CN106779091A - A kind of periodic vibration signal localization method based on transfinite learning machine and arrival distance - Google Patents

Abstract

The invention discloses a kind of periodic vibration signal localization method based on transfinite learning machine and arrival distance.The present invention comprises the following steps：Step 1, based on known accurate fundamental frequency and reach distance, obtain training range prediction model.Step 2, the unknown fundamental frequency collected at 3 and the node of the above is obtained in the same time period respectively with the periodic vibration signal for reaching distance；Step 3, the periodic vibration signal at any node, carry out accurate fundamental frequency f_iExtraction, and the extraction of FBED characteristic vectors is carried out based on the accurate fundamental frequency for obtaining；Step 4, to periodic vibration signal at any node, extraction obtains FBED characteristic vector Ws, and distance estimations are carried out to characteristic vector W using the ELM forecast models for training, and obtains corresponding range estimation d_i；Step 5, the estimated coordinates for calculating vibration source.The present invention realizes under single node high-precision distance estimations and with the training being exceedingly fast and the speed of real-time estimation, reduces the cost of sensor network laying.

Description

A kind of periodic vibration signal localization method based on transfinite learning machine and arrival distance

Technical field

The invention belongs to field of signal processing, it is related to a kind of periodic vibration signal based on transfinite learning machine and arrival distance Localization method.

Background technology

Traditional high precision nonlinear regression estimation algorithm such as SVMs (SVM), error back propagation (BP) nerve Network and now popular deep learning algorithm, have that model training and real-time estimation speed are relatively slow, be not suitable for live application The application problem of environment.

Traditional vibration source positioning is positioned based on the deflection for reaching time delay based on what multisensor array node was realized Scheme, but it is applied to there are problems that following 2 when earth's surface periodic vibration reaches distance detection：

1. the calculating that traditional targeting scheme reaches time delay is necessarily dependent upon more accurate velocity of wave, but vibration wave in different Jie Velocity of wave has certain difference when being propagated in quality table, and the array positioning mode to depending on accurate delay inequality can be caused greatly Error, influences precision；

2. multisensor array node is higher to sensor and support circuit requirement, and cost is larger, is pushed away without large area Wide condition；

3. single multisensor array node be only capable of realizing vibration source arrival direction angle estimation, it is necessary to 2 and the above it is many Sensor array node realizes that direction intersects the plane coordinates that just can determine that vibration source, therefore considerably increases the laying of sensor Cost, cost performance is relatively low.

The present invention proposes that one kind is shaken for the earth's surface cycle based on learning machine intelligent algorithm and the arrival Distance positioning method of transfiniting The method of source coordinate setting.The method is entered row distance and is estimated based on the learning machine (Extreme Learning Machine, ELM) that transfinites Meter can realize under single node high-precision distance estimations and with the training being exceedingly fast and the speed of real-time estimation.In addition it is based on arriving Single-sensor node only needs 1 vibrating sensor to realize reaching the estimation of distance up in the targeting scheme of distance, and non-traditional Need multiple vibrating sensors to realize the estimation of arrival direction in targeting scheme in array node, reduce sensor network laying Cost, reduce this method popularization difficulty.

The content of the invention

The purpose of the present invention is directed to the deficiencies in the prior art, there is provided a kind of week based on transfinite learning machine and arrival distance Phase vibration signal localization method.

The technical solution adopted for the present invention to solve the technical problems comprises the following steps：

Step 1, acquisition training range prediction model

1.1st, it is L for any one piece of data length based on known accurate fundamental frequency and arrival distance_fKnown fundamental frequency and arrive Up to the periodic vibration signal of distance, FBED is carried out based on known accurate fundamental frequency (FrequencyBandEnergyDistribution) extraction of characteristic vector；

1.2nd, periodic vibration signal and corresponding standard FBED characteristic vectors storehouse is built under collection different distance；

1.3rd, learning machine (ELM) Algorithm for Training range prediction of transfiniting is used based on the FBED characteristic vectors for demarcating arrival distance Model.

Step 2, the node P for obtaining in the same time period 3 and the above respectively_i(i=1,2,3 ...) place collects Unknown fundamental frequency and the periodic vibration signal S for reaching distance_i(n), n=1,2 ..., L_f, data length is all L_f, sample frequency be F_s；

Step 3, for any node P_iThe periodic vibration signal S at (i=1,2,3 ...) place_iN (), carries out accurate fundamental frequency f_i Extraction, and based on the accurate fundamental frequency f for obtaining_iCarry out the extraction of FBED characteristic vectors；

3.1st, based on conventional fundamental frequency extraction algorithm, such as autocorrelation sequence method, average amplitude difference method or Cepstrum Method carry out base Frequency is extracted and obtains fundamental frequency estimation valueAnd match immediate accurate fundamental frequency f_i；

3.2nd, based on accurate fundamental frequency from periodic vibration signal S_iExtracted on (n) and obtain FBED characteristic vector Ws.

Step 4, to any node P_iLocate periodic vibration signal S_iN (), extraction obtains FBED characteristic vector Ws, using training ELM forecast models distance estimations are carried out to characteristic vector W, obtain corresponding range estimation d_i；

Step 5, the estimated coordinates for calculating vibration source, it is specific as follows：Based on periodic vibration signal S at any 3 nodes_i N (), the range estimation respectively obtained by ELM forecast models, the positioning mode for being then based on reaching distance is calculated this and shakes The estimated coordinates in dynamic source；

Vibration source estimated coordinates are calculated as follows：

Set any 3 nodes and be respectively P₁(x₁,y₁)、P₂(x₂,y₂)、P₃(x₃,y₃), and the estimated coordinates of vibration source are (x, y), then the formula based on the positioning mode for reaching distance is as follows：

Wherein

Wherein, d₁、d₂、d₃Node P is referred to respectively₁(x₁,y₁)、P₂(x₂,y₂)、P₃(x₃,y₃) range estimation that obtains of place, And γ₁、γ₂Intermediate variable during to calculate；

FBED characteristic vector pickup processes in described step 1.1 and step 3.2 are identicals, and its processing procedure is base In certain fundamental frequency f_fd, extract N_bDimension FBED characteristic vectorsSpecific formula for calculation group is as follows：

1) dimension N_bDetermine formula：

2) i-th frequency band bound scope [f_L(i),f_R(i)] computing formula：

Additionally due to frequency band range may have with half spectral limit of PSD (f) conflict, so to f_R(N_b) have one amendment Operation：

f_R(N_b)=min [F_s/2,f_R(N_b)]；

3) characteristic vector before normalizingComputing formula：

PSD (f) is power spectral density (Power Spectral Density) sequence of this segment signal herein, and it is used as one Conventional Digital Signal Processing frequency-domain analysis object is planted, with many general computational methods, here is omitted；

4) FBED characteristic vectorsComputing formula it is as follows：

Described FBED characteristic vectors are existing relatively stable FBED (FrequencyBandEnergyDistributio N) characteristic vector, it would however also be possible to employ other cover the distance feature up to range information, has no effect on the validity of subsequent processes.

Due to using existing general fundamental frequency extraction algorithm and fundamental frequency matching process in described step 3.1, so Specific scheme can be according to demand selected herein, have no effect on the validity of subsequent processes.

The present invention has the beneficial effect that：

Carrying out distance estimations can be real for learning machine (Extreme Learning Machine, ELM) based on transfiniting for the method High-precision distance estimations and with the training being exceedingly fast and the speed of real-time estimation under existing single node.

In addition based on single-sensor node in the targeting scheme for reaching distance only need 1 vibrating sensor realize reaching away from From estimation, and need multiple vibrating sensors to realize the estimation of arrival direction in non-traditional targeting scheme in array node, drop The cost that low sensor network is laid, that is, reduce the difficulty of this method popularization.

Brief description of the drawings

Fig. 1 reaches distance estimations and localization method schematic flow sheet

Specific embodiment

The present invention is elaborated with reference to the accompanying drawings and detailed description.

As shown in figure 1, a kind of periodic vibration signal localization method based on transfinite learning machine and arrival distance, specifically includes Following steps：

Step 1, acquisition training range prediction model

1.1st, it is L for any one piece of data length based on known accurate fundamental frequency and arrival distance_fKnown fundamental frequency and arrive Up to the periodic vibration signal of distance, FBED (FrequencyBandEnergyDistributi are carried out based on known accurate fundamental frequency On) the extraction of characteristic vector；

1.2nd, periodic vibration signal and corresponding standard FBED characteristic vectors storehouse is built under collection different distance；

1.3rd, learning machine (ELM) Algorithm for Training range prediction of transfiniting is used based on the FBED characteristic vectors for demarcating arrival distance Model.

Step 2, the node P for obtaining in the same time period 3 and the above respectively_i(i=1,2,3 ...) place collects Unknown fundamental frequency and the periodic vibration signal S for reaching distance_i(n), n=1,2 ..., L_f, data length is all L_f, sample frequency be F_s；

Step 3, for any node P_iThe periodic vibration signal S at (i=1,2,3 ...) place_iN (), carries out accurate fundamental frequency f_i Extraction, and based on the accurate fundamental frequency f for obtaining_iCarry out the extraction of FBED characteristic vectors；

3.1st, based on conventional fundamental frequency extraction algorithm, such as autocorrelation sequence method, average amplitude difference method or Cepstrum Method carry out base Frequency is extracted and obtains fundamental frequency estimation valueAnd match immediate accurate fundamental frequency f_i；

3.2nd, based on accurate fundamental frequency from periodic vibration signal S_iExtracted on (n) and obtain FBED characteristic vector Ws.

Step 4, to any node P_iLocate periodic vibration signal S_iN (), extraction obtains FBED characteristic vector Ws, using training ELM forecast models distance estimations are carried out to characteristic vector W, obtain corresponding range estimation d_i；

Step 5, the estimated coordinates for calculating vibration source, it is specific as follows：Based on periodic vibration signal S at any 3 nodes_i N (), the range estimation respectively obtained by ELM forecast models, the positioning mode for being then based on reaching distance is calculated this and shakes The estimated coordinates in dynamic source；

Vibration source estimated coordinates are calculated as follows：

Set any 3 nodes and be respectively P₁(x₁,y₁)、P₂(x₂,y₂)、P₃(x₃,y₃), and the estimated coordinates of vibration source are (x, y), then the formula based on the positioning mode for reaching distance is as follows：

Wherein

Wherein, d₁、d₂、d₃Node P is referred to respectively₁(x₁,y₁)、P₂(x₂,y₂)、P₃(x₃,y₃) range estimation that obtains of place, And γ₁、γ₂Intermediate variable during to calculate；

FBED characteristic vector pickup processes in described step 1.1 and step 3.2 are identicals, and its processing procedure is base In certain fundamental frequency f_fd, extract N_bDimension FBED characteristic vectorsSpecific formula for calculation group is as follows：

1) dimension N_bDetermine formula：

2) i-th frequency band bound scope [f_L(i),f_R(i)] computing formula：

Additionally due to frequency band range may have with half spectral limit of PSD (f) conflict, so to f_R(N_b) have one amendment Operation：

f_R(N_b)=min [F_s/2,f_R(N_b)]；

3) characteristic vector before normalizingComputing formula：

PSD (f) is power spectral density (Power Spectral Density) sequence of this segment signal herein, and it is used as one Conventional Digital Signal Processing frequency-domain analysis object is planted, with many general computational methods, here is omitted；

4) FBED characteristic vectorsComputing formula：

Described FBED characteristic vectors are existing relatively stable FBED (FrequencyBandEnergyDistributio N) characteristic vector, it would however also be possible to employ other cover the distance feature up to range information, has no effect on the validity of subsequent processes.

In step 3.1, conventional fundamental frequency extraction algorithm generally needs the data length Lf obtained in step 2 to comprise at least 2 and the actual fundamental frequency cycles of the above, to ensure to extract to obtain fundamental frequency information therein.

It is understood that the essence of FBED distance features vector is periodic vibration signal power spectrum in step 1.1 and 3.2 The compression expression form of degree.It is that energy is concentrated on during (a) is based on periodic vibration signal as 2 basic foundations of distance feature Near fundamental frequency integral multiple；(b) with propagation distance (arrival distance) change, the decay of the different frequency composition of periodic vibration signal Degree is different.Each fundamental frequency integral multiple is used in formula in note of the present invention 1 for frequency band range center, cover width are base The frequency band range of frequency size.In fact can according to demand select each frequency bands Band wide to reach different anti-broadband noise and bases The effect of frequency evaluated error.

Positioning in step 5 can be done into one to each vibration source estimated coordinates that all 3 combination of nodes are obtained as needed Step is integrated, to reach more preferably locating effect.The present invention positioning side of displaying based on arrival distance by taking 3 minimum nodes as an example Method, so no longer being repeated with regard to this.Additionally, 3 vibrating sensors need not point-blank in 3 node locatings, and preferably exist Earth's surface Triangle-Profile at an acute angle simultaneously ensures rational spacing, smaller and to the vibration source of all directions to ensure distance estimations error Locating effect is approached.

Claims (2)

1. it is a kind of learning machine and to reach the periodic vibration signal localization method of distance based on transfiniting, it is characterised in that including following step Suddenly：

Step 1, acquisition training range prediction model

1.1st, it is L for any one piece of data length based on known accurate fundamental frequency and arrival distance_fKnown fundamental frequency and reach away from From periodic vibration signal, the extraction of FBED characteristic vectors is carried out based on known accurate fundamental frequency；

1.2nd, periodic vibration signal and corresponding standard FBED characteristic vectors storehouse is built under collection different distance；

1.3rd, the learning machine Algorithm for Training range prediction model that transfinites is used based on the FBED characteristic vectors for demarcating arrival distance；

Step 2, the unknown fundamental frequency that is collected at 3 and the node of the above is obtained in the same time period respectively and distance is reached Periodic vibration signal S_i(n), n=1,2 ..., L_f, data length is all L_f, sample frequency be F_s；

Step 3, for any node P_iThe periodic vibration signal S at place_iN (), carries out accurate fundamental frequency f_iExtraction, wherein i=1,2, 3..., and based on the accurate fundamental frequency f for obtaining_iCarry out the extraction of FBED characteristic vectors；

3.1st, based on conventional fundamental frequency extraction algorithm, such as autocorrelation sequence method, average amplitude difference method or Cepstrum Method carry out fundamental frequency and carry Take acquisition fundamental frequency estimation valueAnd match immediate accurate fundamental frequency f_i；

3.2nd, based on accurate fundamental frequency from periodic vibration signal S_iExtracted on (n) and obtain FBED characteristic vector Ws；

Step 4, to any node P_iLocate periodic vibration signal S_iN (), extraction obtains FBED characteristic vector Ws, using what is trained ELM forecast models carry out distance estimations to characteristic vector W, obtain corresponding range estimation d_i；

Step 5, the estimated coordinates for calculating vibration source, it is specific as follows：Based on periodic vibration signal S at any 3 nodes_iN (), leads to The range estimation that ELM forecast models are respectively obtained is crossed, the positioning mode for being then based on reaching distance is calculated the vibration source Estimated coordinates；

Vibration source estimated coordinates are calculated as follows：

Set any 3 nodes and be respectively P₁(x₁,y₁)、P₂(x₂,y₂)、P₃(x₃,y₃), and vibration source estimated coordinates for (x, Y), then the formula based on the positioning mode for reaching distance is as follows：

$x=\frac{{\mathrm{\γ}}_1(y_2-y_1)-{\mathrm{\γ}}_2(y_3-y_2)}{(x_3-x_2)(y_2-y_1)-(x_2-x_1)(y_3-y_2)},$

$y=\frac{{\mathrm{\γ}}_1(x_2-x_1)-{\mathrm{\γ}}_2(x_3-x_2)}{(x_2-x_1)(y_3-y_2)-(x_3-x_2)(y_2-y_1)},$

Wherein

${\mathrm{\γ}}_2=\frac{1}{2}\[(x_2^2+y_2^2)-(x_1^2+y_1^2)+d_1^2-d_2^2\];$

Wherein, d₁、d₂、d₃Node P is referred to respectively₁(x₁,y₁)、P₂(x₂,y₂)、P₃(x₃,y₃) range estimation that obtains of place, and γ₁、γ₂Intermediate variable during to calculate.

2. it is according to claim 1 it is a kind of based on transfinite learning machine and reach distance periodic vibration signal localization method, It is characterized in that：

FBED characteristic vector pickup processes in described step 1.1 and step 3.2 are identicals, and its processing procedure is based on certain Fundamental frequency f_fd, extract N_bDimension FBED characteristic vectorsSpecific formula for calculation group is as follows：

1) dimension N_bDetermine formula：

2) i-th frequency band bound scope [f_L(i),f_R(i)] computing formula：

I=1,2 ..., N_b.

Additionally due to frequency band range may have with half spectral limit of PSD (f) conflict, so to f_R(N_b) there is an operation for amendment：

f_R(N_b)=min [F_s/2,f_R(N_b)]；

3) characteristic vector before normalizingComputing formula：

$v_i=\underset{f=f_L\left(i\right)}{\overset{f_R\left(i\right)}{\Σ}}PSD\left(f\right),$

4) FBED characteristic vectorsComputing formula：

$W=\frac{V}{\left|\right|V|{|}_1}.$