CN102625446B - Method for positioning object tags in Internet of Things - Google Patents

Abstract

The invention provides a method for positioning object tags in the Internet of Things. By the steps of setting relevant variables of the Internet of Things and initializing the Internet of Things, then updating a measured value of the first layer of an SVM (support vector machine), performing accurate position estimation on object tags and obtaining the positions of the object tags by using a fuzzy neural network, the method solves the problems caused during positioning of the object tags in the Internet of Things. The method achieves the effect of accurately positioning the object tags with high efficiency while meeting the positioning requirements for object tags in the Internet of Things.

Description

A kind of object tag location method in Internet of Things

Technical field

The present invention relates to wireless communication technology field, particularly relate to Internet of Things, Sensor Network and wireless transmission.

Background technology

The Internet of Things that sensor technology, distributed information processing, wireless communication technology, network technology and embedded computing technique are combined together, can by internodal concertedness to external world environment carry out real-time monitoring and detection, gather voluntarily environmental information and carry out corresponding parallel computation processing, deal with data can be transferred to user side, realize environment and environment, the information interaction between environment and people.

Internet of Things based on Sensor Network technical foundation, its positioning requirements to object is higher.Two class modes of existing Internet of Things object tag location are the localization method based on range finding and the localization method that need not find range.More common ranging technology has RSSI, TOA, TDOA and AOA etc.RSSI technology has low-power consumption, feature but often occur ± 50% error of range finding result cheaply, the applied environment of the accurate distance result of can not satisfying the demand.The key of TOA ranging technology is how to guarantee that node asks that precise time is synchronous, and this technology is had to certain limitation for Internet of Things.TDOA technology is utilized ultrasonic technology, this ranging technology need to be considered each factor and hyperacoustic transmission range of ultrasonic wave impact, as shown in Figure 1, and the selected ultrasonic propagation of Internet of Things distance is shorter for its positioning principle, the topology of networks of having selected TDOA technical limitations.The support of other hardware of AOA Technology Need, from considering that economically this ranging technology is not suitable for large scale network.In order to reduce the deficiency of these ranging technologies, location algorithm based on range finding adopts various means to reduce the impact of range error on location, as repeatedly measured, circulation location refinement, and these means all can produce a large amount of energy consumptions, so the localization method based on range finding has certain assurance and need research in the cost of location in precision.The localization method that need not find range is a kind of rely on other complicated foundation facilities or extensive centralized calculating realization, its positioning precision is not high, but energy-conservation aspect is obviously better than to the locate mode based on range finding, when position error be less than sensor node radio communication radius 40% time, position error is little on the impact of the correlation techniques such as routing performance and target tracking accuracy, therefore the localization method of coarse positioning precision also can meet the primary demand of Internet of Things, and the Zigbee using in engineering and UWB location model are respectively as shown in Figures 2 and 3.

Therefore, for realizing the object label of Internet of Things, accurately locate, need the efficient localization method of design.

Summary of the invention

Technical problem to be solved by this invention is: solve the object tag location problem of Internet of Things, in the object tag location that meets Internet of Things, require, realize its efficiently and accurately location.

The present invention, for a kind of object tag location method providing in Internet of Things is provided, is characterized in that:

A, Internet of things system correlated variables is set, and it is carried out to initialization, then the measured value of SVM ground floor is upgraded;

B, object label is carried out to exact position estimation;

C, the position of using fuzzy neural network accurately to obtain object label;

In described steps A, make z _tfor observational variable, it is radio frequency signal (RSSI) signal that all Wireless RF identifiers (RFID) receiver obtains, y _tfor hidden variable, it is the rough object label estimated position in the block layer of use SVMs (SVM) grader, x _tfor state variable, it is the accurate object label estimated position at fine layer.The ground floor of SVM is for passing through observational variable z _trough estimate label position y _t, the second layer is for passing through hidden variable z _taccurately estimate label position x _t, as shown in Figure 4, its overview flow chart as shown in Figure 5 for the structure of SVM.

In described steps A, the measured value of SVM ground floor is upgraded.Svm classifier device is used for selecting object label model, and probability description is passed through in the estimated position of object label.Make the estimated position of object label follow Gaussian Profile wherein μ t is average, for variance.Adopt strategy upgrade object label position, wherein δ=| y ' _t-y ' _t-1|, D is the artificial threshold value arranging, y ' _tfor the object label estimated position at moment t, y ' _t-1for the object label estimated position at a upper moment t-1.

In described steps A, for improving the reliability of current object label estimated position, adopt most voting methods.Its three continuous label estimated positions of use y " _t-2, y " _t-1, y " _tin majority as current estimated position, i.e. y _t=Gof{y " _t-2, y _t-1, y " _t, G be three continuous label estimated positions y " _t-2, y " _t-1, y " _tin majority.

In described step B, first the measured value that carries out the SVM second layer upgrades, the average of being sampled by estimation module multidigit Gaussian Profile obtaining from SVM ground floor, the weights of each sampling upgrade according to the distance between average and sample position, at m weights of sampling of moment t, are then upgrade and probability of motion distribution rule service time obtain roughly the estimated position △ of object label.

In described step C, design structure of fuzzy neural network is for obtaining the optimum weights of initial estimated location △.Structure of fuzzy neural network comprises three parts: input layer, output layer and hidden layer, it is many single-input single-output system (SISO system)s.The activation primitive of hidden neuron is input neuron is P (n), being predicted as of output neuron n is time sequence number.

In described step C, design fuzzy neural network amending unit structure is for revising predicted weights.It comprises three parts: input layer, output layer and hidden layer, and it is many single-input single-output system (SISO system)s, its structure is as shown in Figure 6.Error after predicted weights are corrected is met $\frac{\mathrm{\&alpha;}}{N}[\underset{X_{m,n}\&Element;A}{\mathrm{\&Sigma;}}\mathrm{\&mu;}{R}_1(X_{i,j},X_{m,n},\mathrm{\&alpha;})-\underset{X_{m,n}\&Element;A}{\mathrm{\&Sigma;}}\mathrm{\&mu;}{R}_2(X_{i,j},X_{m,n},\mathrm{\&alpha;})]<\mathrm{\&gamma;},$ Wherein η is for threshold value is manually set, and span is [0.9,0.99], $\mathrm{\&mu;}{R}_s(X_{i,j},X_{m,n},\mathrm{\&alpha;})=\mathrm{MAX}\{(1-\frac{|X_{i,j},X_{m,n}-\mathrm{\&alpha;}|}{2\mathrm{\&alpha;}}),0\}S=1,$

x _i,jfor a certain period is trained the average of optimum weights data, X _m,nfor with the immediate value of optimum weights size, N be and the quantity of the big or small immediate value of optimum weights that α ∈ [0,9] is scalar factor.

In described step C, use training mechanism to adjust input value and the output valve of Fuzzy Neural Network System, its sub-step is:

A. produce input layer;

B. produce hidden layer;

C. produce an output neuron target;

D. making optimum weights is zero, and initialization cycle number;

E. chromosome population is carried out to initialization;

F. be each link assignment weights;

E. obtain the output of input neuron and the product of weights;

G. obtain average and the deviation of each hidden neuron;

H. obtain the output of all vanishing targets;

I. be each the link assignment weights between hidden neuron and output neuron;

J. fix average and the deviation of each output neuron;

K. obtain the output of all output neuron targets;

L. obtain accumulated error and optimum weights, if former optimum weights are less than current optimum weights, store current optimum weights, if former optimum weights are greater than current optimum weights, the optimum weights before storing;

M. usage counter is counted;

N. use roulette wheel mechanism to select two parents;

O. chromosome is used to intersection, variation and copy step, produce new weights, and give each link by new weights;

If p. number of cycles is greater than the counting of counter, return value sub-step f, and repeat above-mentioned sub-step, when predicted weights are obtained, use detecting unit and decision device functional module to detect weights and whether meet the demands, if meet by genetic algorithm, used, if do not meet use fuzzy neural network functional module to revise predicted weights, wherein η is for threshold value is manually set, and its span is [0.8,0.9].

In described step C, use the optimum weight function of genetic algorithm adjustment neural network parameter and chromosome to assess chromosome individuality.Use interference prediction system evaluation method of measurement,

Optimum weights=w ₁* susceptibility+w ₂* specificity, wherein the subset i of predicted normal training data, [the subset i of the normal training data that 1-is predicted], the subset i of predicted improper training data,

[the subset i of the improper training data that 1-is predicted], p and q are respectively the quantity that training data that chromosome uses is concentrated normal and improper data, w ₁and w ₂for according to the weights of each regular allocation.

Beneficial effect of the present invention is, realizing under the object tag location prerequisite of Internet of Things, improved the accuracy of its positioning precision and the prediction of object label position.

Accompanying drawing explanation

Fig. 1 is TDOA positioning principle schematic diagram;

Fig. 2 is Zigbee location model schematic diagram;

Fig. 3 is UWB location model schematic diagram;

Fig. 4 is SVM structural representation;

Fig. 5 is total process flow operation schematic diagram;

Fig. 6 is fuzzy neural network amending unit structural representation.

Embodiment

The object label that the present invention is directed to Internet of Things is orientation problem accurately, reducing under the prerequisite of using method complexity, provides a kind of object tag location method that solves Internet of Things.

For achieving the above object, technical scheme of the present invention is as follows:

Step 1, arranges Internet of things system correlated variables, and it is carried out to initialization.Make z _tfor observational variable, it is the RSSI signal that all RFID receivers obtain, y _tfor hidden variable, it is the rough label estimated position in the block layer of use SVMs (SVM) grader, x _tfor state variable, it is the accurate label estimated position at fine layer, and ground floor is for passing through observational variable z _trough estimate label position y _t, the second layer is for passing through hidden variable z _taccurately estimate label position x _t.

Step 2, upgrades the measured value of SVM ground floor.Svm classifier device is used for selecting label model, and probability description is passed through in the estimated position of label.Make the estimated position of label follow Gaussian Profile μ wherein _tfor average, for variance.Adopt strategy upgrade label position, wherein δ=| y ' _t-y ' _t-1|, D is the artificial threshold value arranging, y ' _tfor the label estimated position at moment t, y ' _t-1for the label estimated position at a upper moment t-1.

For improving the reliability of current label estimated position, adopt most voting methods, use three continuous label estimated positions y " _t-2, y " _t-1, y " _tin majority as current estimated position, i.e. y _t=Gof{y " _t-2, y " _t-1, y " _t, G be three continuous label estimated positions y " _t-2, y " _t-1, y " _tin majority.

Step 3, carries out exact position estimation to object label.First the measured value that carries out the second layer upgrades, the average of being sampled by estimation module multidigit Gaussian Profile obtaining from ground floor, and the weights of each sampling upgrade according to the distance between average and sample position, at m weights of sampling of moment t, are $w_t^m=\frac{\mathrm{\&alpha;}\left(m\right)p\left(y_t^l\right|x_t^m)}{\underset{s=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{\&alpha;}\left(s\right)p\left(y_t^l\right|x_t^s)};$ Then upgrade service time and probability of motion distribution $p\left(x_{t+1}\right|Y_{1:t})=\underset{m=1}{\overset{M}{\mathrm{\&Sigma;}}}{w}_t^{m}p\left(x_{t+1}\right|x_t^m)$ Obtain roughly the estimated position △ of object label.

Step 4, design structure of fuzzy neural network is for obtaining the optimum weights of initial estimated location △.Structure of fuzzy neural network comprises three parts: input layer, output layer and hidden layer, and it is many single-input single-output system (SISO system)s, its structure is as shown in Figure 1.The activation primitive of hidden neuron is input neuron is P (n), being predicted as of output neuron n is time sequence number.

Step 5, design fuzzy neural network amending unit structure is for revising predicted weights.It comprises three parts: input layer, output layer and hidden layer, and it is many single-input single-output system (SISO system)s, its structure is as shown in Figure 2.Error after predicted weights are corrected is met $\frac{\mathrm{\&alpha;}}{N}[\underset{X_{m,n}\&Element;A}{\mathrm{\&Sigma;}}\mathrm{\&mu;}{R}_1(X_{i,j},X_{m,n},\mathrm{\&alpha;})-\underset{X_{m,n}\&Element;A}{\mathrm{\&Sigma;}}\mathrm{\&mu;}{R}_2(X_{i,j},X_{m,n},\mathrm{\&alpha;})]<\mathrm{\&gamma;},$ Wherein η is for threshold value is manually set, and span is [0.9,0.99], $\mathrm{\&mu;}{R}_s(X_{i,j},X_{m,n},\mathrm{\&alpha;})=\mathrm{MAX}\{(1-\frac{|X_{i,j},X_{m,n}-\mathrm{\&alpha;}|}{2\mathrm{\&alpha;}}),0\}S=1,$ x _i,jfor a certain period is trained the average of optimum weights data, X _m,nfor with the immediate value of optimum weights size, N be and the quantity of the big or small immediate value of optimum weights that α ∈ [0,9] is scalar factor.

Step 6, is used training method to adjust input value and the output valve of Fuzzy Neural Network System.Its sub-step is:

A. produce input layer;

B. produce hidden layer;

C. produce an output neuron target;

D. making optimum weights is zero, and initialization cycle number;

E. chromosome population is carried out to initialization;

F. be each link assignment weights;

E. obtain the output of input neuron and the product of weights;

G. obtain average and the deviation of each hidden neuron;

H. obtain the output of all vanishing targets;

I. be each the link assignment weights between hidden neuron and output neuron;

J. fix average and the deviation of each output neuron;

K. obtain the output of all output neuron targets;

L. obtain accumulated error and optimum weights, if former optimum weights are less than current optimum weights, store current optimum weights, if former optimum weights are greater than current optimum weights, the optimum weights before storing;

M. usage counter is counted;

N. use roulette wheel mechanism to select two parents;

O. to steps such as chromosome are used intersection, makes a variation and copies, produce new weights, and give each link by new weights;

If p. number of cycles is greater than the counting of counter, return value sub-step f, and repeat above-mentioned sub-step, when predicted weights are obtained, use detecting unit and decision device functional module to detect weights and whether meet the demands, if meet wherein η is for threshold value is manually set, and span is [0.8,0.9], by genetic algorithm, used, if do not meet use fuzzy neural network functional module to revise predicted weights.

Step 7, is used genetic algorithm to adjust neural network parameter.

Step 8, is used the optimum weight function of chromosome to assess individuality.Use interference prediction system evaluation method of measurement,

Optimum weights=w ₁* susceptibility+w ₂* specificity, wherein the subset i of predicted normal training data, [the subset i of the normal training data that 1-is predicted], the subset i of predicted improper training data, [the subset i of the improper training data that 1-is predicted], p and q are respectively the quantity that training data that chromosome uses is concentrated normal and improper data, w ₁and w ₂for according to the weights of each regular allocation.

Claims (7)

1. the object tag location method in Internet of Things, solves the object tag location problem of Internet of Things, in the object tag location that meets Internet of Things, requires, and realizes its efficiently and accurately location, comprises the steps:

A, Internet of things system correlated variables is set, and it is carried out to initialization, then the measured value of SVM ground floor is upgraded, be specially and make z _tfor observational variable, it is radio frequency signal (RSSI) signal that all Wireless RF identifiers (RFID) receiver obtains, y _tfor hidden variable, it is the rough object label estimated position in the block layer of use SVMs (SVM) grader, x _tfor state variable, it is the accurate object label estimated position at fine layer, and the ground floor of SVM is for passing through observational variable z _trough estimate label position y _t, the second layer is for passing through hidden variable y _taccurately estimate label position x _t;

B, object label is carried out to exact position estimation, be specially the measured value renewal of first carrying out the SVM second layer, the average of being sampled by estimation module multidigit Gaussian Profile obtaining from SVM ground floor, the weights of each sampling upgrade according to the distance between average and sample position, at m weights of sampling of moment t, are then upgrade and probability of motion distribution rule service time obtain roughly the estimated position △ of object label;

C, the position of using fuzzy neural network accurately to obtain object label, be specially design structure of fuzzy neural network for obtaining the optimum weights of initial estimated location △, structure of fuzzy neural network comprises three parts: input layer, output layer and hidden layer, it is many single-input single-output system (SISO system)s, and the activation primitive of hidden neuron is input neuron is P (n), being predicted as of output neuron n is time sequence number.

2. according to the method for claim 1, for described steps A, it is characterized in that: the measured value to SVM ground floor upgrades, svm classifier device is used for selecting object label model, and the estimated position of object label, by probability description, makes the estimated position of object label follow Gaussian Profile μ wherein _tfor average,

Variance, adopts strategy upgrade object label position, wherein δ=| y ' _t-y ' _t-1|, D is the artificial threshold value arranging, y ' _tfor the object label estimated position at moment t, y ' _t-1for the object label estimated position at a upper moment t-1.

3. according to the method for claim 1, for described steps A, it is characterized in that: for improving the reliability of current object label estimated position, adopt most voting methods, its three continuous label estimated positions of use y " _t-2, y " _t-1, y " _tin majority as current estimated position, i.e. y _t=Gof{y " _t-2, y " _t-1, y " _t, G be three continuous label estimated positions y " _t-2, y " _t-1, y " _tin majority.

4. according to the method for claim 1, for described step C, it is characterized in that: design structure of fuzzy neural network is for obtaining the optimum weights of initial estimated location △, structure of fuzzy neural network comprises three parts: input layer, output layer and hidden layer, it is many single-input single-output system (SISO system)s, and the activation primitive of hidden neuron is input neuron is P (n), being predicted as of output neuron n is time sequence number.

5. according to the method for claim 1, for described step C, it is characterized in that: design fuzzy neural network amending unit structure is for revising predicted weights, it comprises three parts: input layer, output layer and hidden layer, it is many single-input single-output system (SISO system)s, and the error after predicted weights are corrected is met $\frac{\mathrm{\&alpha;}}{N}[\underset{X_{m,n}\&Element;A}{\mathrm{\&Sigma;}}\mathrm{\&mu;}{R}_1(X_{i,j},X_{m,n},\mathrm{\&alpha;})-\underset{X_{m,n}\&Element;A}{\mathrm{\&Sigma;}}\mathrm{\&mu;}{R}_2(X_{i,j},X_{m,n},\mathrm{\&alpha;})]<\mathrm{\&gamma;},$ Wherein γ is for threshold value is manually set, and its span is [0.9,0.99], $\mathrm{\&mu;}{R}_s(X_{i,j},X_{m,n},\mathrm{\&alpha;})=\mathrm{MAX}\{(1-\frac{|X_{i,j},X_{m,n}-\mathrm{\&alpha;}|}{2\mathrm{\&alpha;}}),0\}S=1,$

x _i,jfor a certain period is trained the average of optimum weights data, X _m,nfor with the immediate value of optimum weights size, N be and the quantity of the big or small immediate value of optimum weights that α ∈ [0,9] is scalar factor.

6. according to the method for claim 1, for described step C, it is characterized in that: use training mechanism to adjust input value and the output valve of Fuzzy Neural Network System, its sub-step is:

A. produce input layer;

B. produce hidden layer;

C. produce an output neuron target;

D. making optimum weights is zero, and initialization cycle number;

E. chromosome population is carried out to initialization;

F. be each link assignment weights;

E. obtain the output of input neuron and the product of weights;

G. obtain average and the deviation of each hidden neuron;

H. obtain the output of all vanishing targets;

I. be each the link assignment weights between hidden neuron and output neuron;

J. fix average and the deviation of each output neuron;

K. obtain the output of all output neuron targets;

L. obtain accumulated error and optimum weights, if former optimum weights are less than current optimum weights, store current optimum weights, if former optimum weights are greater than current optimum weights, the optimum weights before storing;

M. usage counter is counted;

N. use roulette wheel mechanism to select two parents;

O. chromosome is used to intersection, variation and copy step, produce new weights, and give each link by new weights;

If p. number of cycles is greater than the counting of counter, return value sub-step f, and repeat above-mentioned sub-step, when predicted weights are obtained, use detecting unit and decision device functional module to detect weights and whether meet the demands, if meet by genetic algorithm, used, if do not meet use fuzzy neural network functional module to revise predicted weights, wherein η is for threshold value is manually set, and its span is [0.8,0.9].

7. according to the method for claim 1, for described step C, it is characterized in that: use the optimum weight function of genetic algorithm adjustment neural network parameter and chromosome to assess chromosome individuality, use interference prediction system evaluation method of measurement,

Optimum weights=w ₁* susceptibility+w ₂* specificity, wherein the subset i of predicted normal training data, [the subset i of the normal training data that 1-is predicted], the subset i of predicted improper training data, [the subset i of the improper training data that 1-is predicted], p and q are respectively the quantity that training data that chromosome uses is concentrated normal and improper data, w ₁and w ₂for according to the weights of each regular allocation.