CN102496062A - Personnel information fusion method based on Spiking neural network - Google Patents

Abstract

The invention discloses a personnel information fusion method based on a Spiking neural network. The personnel information fusion method comprises the following steps of: (1) collecting information of personnel number, and storing data in a real-time database; (2) reading the data from the real-time database, preprocessing the data, mapping the data into input vectors, and storing the input vectors in a relational database in groups; (3) designing a structural model of the Spiking neural network according to the input type of a sensor; (4) reading data from the relational database, multiplying the weighted value of trust degree by data, inputting the result into the Spiking neural network, and adopting frequency coding for a sensing neural cell; (5) adopting consistency coding for a hidden neural cell; (6) adopting a Hebbian learning arithmetic, updating the sensing neural cell, updating a connecting weighted value among the sensing neural cell, the hidden neural cell and an output neural cell of the Spiking neural network model, and improving system behaviors; and (7) carrying out potentiality activation calculation on input signals of the output neural cell, and determining the number of output pulses of the output neural cell, i.e. the number of personnel in a region.

Description

A kind of personal information fusion method based on the Spiking neural network

Technical field

The invention belongs to areas of information technology, relate in particular to the personal information fusion method based on the Spiking neural network of personal information collection in the building, Multi-source Information Fusion, the design of Spiking neural network and modeling.

Background technology

It is research direction emerging in the building energy saving field that personnel detect with analyzing, and is also concerning building safety simultaneously, and the building operation energy consumption is played great influence.A large amount of constructions of intelligent building, developing rapidly of building automatic level and increasing substantially of economy and living standard are constantly promoting the demand of building to comfort level and operation energy consumption control, have promoted the fast development of personnel's etection theory with technology.Analyze according to the relevant expert, develop and apply advanced personal information detection technique energetically, raising personnel accuracy of detection can be building energy conservation and building safety provides foundation, has very important realistic meaning.

At present, occurred the expensive personnel's measuring instrument that has the simple analysis function on the market, yet for a building, adopted the instrument of this costliness to be covered with whole building, its cost obviously is difficult to let the people accept.Detect for personnel in the building; A feasible solution is exactly to utilize advanced technology of Internet of things, wireless sensor network technology; Adopt detecting sensor cheap, low-power consumption to form the redundant multisensor network of low-power consumption,, improve personnel's accuracy of detection through multisource information fusion technology; Reduce and detect cost, reduce the wiring significantly and the installation of building are transformed.

Aspect multi-sensor information fusion, domestic and international research personnel have proposed multiple fusion method, like D-S evidential reasoning, method of weighted mean, Kalman filtering method, many Bayes' assessments, fuzzy logic theory, neural network, expert system etc.Yet these methods have but run into difficulty when personnel detect information in handling building.At first, the existing wireless network of personnel's detection system in the building has wired ethernet network again, and network type is various; Secondly, handled sensor type is various, can be made up of CO2, temperature, humidity, infrared, RDIF, the first-class polytype sensing data of shooting; Once more, existing discrete message of buildings inner sensor information and continuous information, some still is the information of random probability distribution.Therefore, the personnel in the building detect the heterogeneous network Multi-source Information Fusion and have brought new challenge to Theory of Information Fusion and technical research.

Numerous scholars have turned to bionics to sight, and the ability of the biological treatment information in the nature is very superb.Therefore, as the Multi-source Information Fusion instrument, personnel's detection problem proposes new method in the building in order to solve the Spiking neural network in the present invention.The Spiking neural network is as third generation neural network; Have preceding two generation neural network some advantages; But it is simultaneously different with traditional first generation perceptron network again with second generation Sigmoid neural network; Because adopt the pulse firing signal as neuronic input/output signal, the Spiking neural network can closer be described actual biological nervous system, and has stronger computing power.It not only can realize the fusion of discrete message and continuous information, can also handle the information problem of random probability distribution.Therefore, the Spiking neural network is fit to solve a heterogeneous network Multi-source Information Fusion difficult problem very much.Therefore, the present invention not only can be building energy conservation and building safety provides foundation, and can promote Spiking neural net model establishing and Design and Development, has very important realistic meaning.

Summary of the invention

The objective of the invention is to propose a kind of personal information fusion method based on the Spiking neural network to personnel's Multi-source Information Fusion and Spiking neural net model establishing problem.It detects low-power consumption, redundant sensor network through the structure personnel, carries out Multi-source Information Fusion, proposes the analytical approach that effective personnel detect data; Design precision height, personnel's detection system that cost is low; Not only can guarantee to build the subregion precision, reduce operation energy consumption, resources conseravtion and environment; And can realize personalized control according to personnel's number, for building energy and safety management and optimal control provide foundation.

For achieving the above object, the present invention adopts following technical scheme:

A kind of personal information fusion method based on the Spiking neural network the steps include:

1) utilize CO2 concentration sensor, infrared sensor, personnel RFID, camera and the relative humidity sensing etc. that possibly exist in the space zone to carry out the collection of personnel's information of number, data deposit real-time data base in;

2) reading of data from real-time data base is carried out pre-service, removes unreasonable data, is mapped as input vector, and relational database divides into groups to deposit in;

3), be designed to comprise the three-layer network structure of input layer, hidden layer and output layer, the corresponding sensing neuron of one of them sensor according to sensor input type design Spiking neural network structure model; CO ₂The sensing neuron output signal that concentration sensor, infrared sensor and relative humidity sensor are corresponding gets into implicit neuron, and the corresponding sensing neuron signal of personnel RFID and camera directly gets into output neuron; The output umber of pulse of the corresponding output neuron of space area people number;

4) reading of data from relational database, data multiply by degree of belief weighted value input Spiking neural network, and the sensing neuron adopts frequency coding;

5) implicit neuron adopts the consistance coding, and promptly implicit neuron carries out the logical computing with the pulse of input, and only when two pulses occurred simultaneously, implicit neuron was just in pulse of corresponding time point output;

6) adopt the Hebbian learning algorithm, upgrade the connection weights between the sensing neuron of Spiking neural network model, implicit neuron and the output neuron, improve system action;

7) the output neuron input signal is activated potential and calculate, confirm the output umber of pulse of output neuron, personnel's number that promptly should the zone.

In the said step 4), said frequency coding, represent with following formula:

$I_i=\mathrm{round}\left(\frac{r_i-s_i}{u_i}\right)$

In the formula, I _iBe i the neuronic output umber of pulse of sensing, r _iBe sensor full range, s _iBe sensor input, u _iBe the unit pulse conversion scale, function round is a bracket function.

The implicit neuron of said Spiking neural network adopts the consistance coding, and promptly implicit neuron carries out the logical computing with the pulse of input, and only when two pulses occurred simultaneously, implicit neuron was just in pulse of corresponding time point output.

In the said step 6), the Hebbian learning algorithm is following:

$w_{\mathrm{ij}}\left(t\right)=d\×w_{\mathrm{ij}}(t-1)+r\×\underset{t_j^{\mathrm{output}},t_i^{\mathrm{input}}}{\mathrm{\Σ}}W\left(\mathrm{\Δ}{\mathrm{\τ}}_{\mathrm{ij}}\right)$

In the formula, w _IjBe the weights that are connected between sensing neuron, implicit neuron and output neuron, $W\left(\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}\right)=\left\{\begin{array}{c}[c_1(1-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\widetilde{\mathrm{\&tau;}}}_1})+c_2(1-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\widetilde{\mathrm{\&tau;}}}_2})]\&CenterDot;\mathrm{Exp}\left(\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{\mathrm{\&tau;}}\right),\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}\&GreaterEqual;0\\ {\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="HDA0000117711460000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1\mathrm{Exp}(-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\mathrm{\&tau;}}_1})+c_2\mathrm{Exp}(-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\mathrm{\&tau;}}_2}),\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}<0\end{array}\right.,$ c ₁, c ₂, τ, τ ₁And τ ₂Be constant,

D is an attenuation rate, and r is a learning rate, Δ τ _IjIt is i input pulse and j mistiming of exporting between the pulse

The formula of computing activation potential is following in the said step 7):

$u_i\left(t\right)=\underset{j}{\mathrm{\&Sigma;}}\underset{t_j^{\left(l\right)}}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}^g\mathrm{\&epsiv;}(t-t_j^{\left(l\right)})+\underset{k}{\mathrm{\&Sigma;}}\underset{t_k^{\left(f\right)}}{\mathrm{\&Sigma;}}{w}_{\mathrm{ik}}^y\mathrm{\&epsiv;}(t-t_k^{\left(l\right)})$

In the formula,

Be the weights that are connected between sensing neuron and the output neuron, For being connected weights, u between implicit neuron and the output neuron _i(t) be the input pulse activation potential of i output neuron,

Time when expression sensor neuron j exports l pulse,

Time when the implicit neuron k of expression exports l pulse, the ε function does

τ wherein ₃Be time constant; Output neuron calculates through input signal being activated potential, confirms the output umber of pulse, personnel's number that promptly should the zone.

The present invention is directed to personnel's Spiking neural network structure that detected the multisensor network design in the building; Input neuron and implicit neuronic coded system have been confirmed; Design the connection weights and confirmed mode and Spiking network using Hebbian mode of learning, provided the output neuron potential computing formula of calculating personnel's number.At first personnel's information of number of the sensor acquisition such as CO2 concentration sensor, infrared sensor, personnel RFID, camera and relative humidity sensor that possibly exist in the same space zone is carried out pre-service, remove unreasonable data, be mapped as input vector; Then data after the pre-service multiply by the sensing neuron of degree of belief weighted value input Spiking neural network input layer, carry out frequency coding; The corresponding implicit neuron of sensing neuron that designs C O2 concentration sensor, infrared sensor and relative humidity sensor are corresponding carries out the consistance coding; Confirm the connection weights between the sensing neuron of Spiking neural network structure model, implicit neuron and the output neuron through the Hebbian learning method; At last output neuron is activated potential and calculate, confirm the output of output neuron, personnel's number that promptly should the zone.

Beneficial effect of the present invention: the present invention adopts third generation neural network Spiking neural network; The mode of simulation biological treatment information is carried out Multi-source Information Fusion; Thereby obtain that precision is higher, adaptability is stronger, the more excellent personnel's number fusion results of robustness, for Environmental security and building energy conservation provide reliable more foundation.

Description of drawings

Fig. 1 is the Spiking neural network structure model of personal information fusion method;

Fig. 2 is a kind of personal information fusion method process flow diagram based on the Spiking neural network.

Embodiment

Below in conjunction with accompanying drawing and embodiment a kind of personal information fusion method based on the Spiking neural network of the present invention is made detailed description.

The present invention proposes a kind of personal information fusion method based on the Spiking neural network, personnel's Multi-source Information Fusion and statistics in being applied to build.Personnel's information of number to five kinds of sensor acquisition such as the CO2 concentration sensor that possibly exist in the area of space, infrared sensor, personnel RFID, camera and relative humidity sensors is carried out pre-service, removes unreasonable data, is mapped as input vector; According to sensor input type design Spiking neural network structure model, the network structure of design comprises input layer, sensing neuron layer, implicit neuron layer, output neuron layer, output layer, shown in accompanying drawing 1.Wherein, The corresponding sensing neuron of sensor; The sensing neuron output signal that CO2 concentration sensor, infrared sensor and relative humidity sensor are corresponding gets into implicit neuron, and the corresponding input neuron of personnel RFID and camera directly gets into output neuron.To pass through the sensing neuron that pretreated data multiply by degree of belief weighted value input Spiking neural network, sensing neuron SF coding, so among Fig. 1 ₁～so _lBe CO ₂The pretreated data of concentration sensor, l is CO ₂The number of concentration sensor; Sr ₁～sr _mBe the pretreated data of infrared sensor, m is CO ₂The number of concentration sensor; Sh ₁～sh _nBe the pretreated data of humidity sensor, n is CO ₂The number of concentration sensor; Sd ₁～sd _jBe the pretreated data of RFID sensor, j is the number of RFID sensor; Sa ₁～sa _kBe the pretreated data of ccd sensor, k is the number of ccd sensor; p _SumBe the output umber of pulse of output neuron, i.e. personnel's number in this area of space.

Accompanying drawing 2 is based on the process flow diagram of the personal information fusion method of Spiking neural network.The various data of sensor acquisition at first get into real-time data base, from real-time data base, read then, carry out pre-service, and relational database divides into groups to deposit in.Carry out reading by sensor type, handle respectively, the sensing neuron output signal that CO2 concentration sensor, infrared sensor and relative humidity sensor are corresponding gets into implicit neuron.Output neuron receives implicit neuron and the neuronic pulse of sensing, activates potential and calculates, and confirms the personnel's data in this construction area through the output umber of pulse then.The Spiking neural network utilizes the input pulse of output neuron and output pulse to carry out Hebbian study, upgrades the connection weights between implicit neuron and the output neuron, improves system action.

The sensing neuron of Spiking neural network adopts frequency coding among the present invention, representes with following formula:

$I_i=\mathrm{round}\left(\frac{r_i-s_i}{u_i}\right)$

In the formula, I _iBe i the neuronic output umber of pulse of sensing, r _iBe sensor full range, s _iBe sensor input, u _iBe the unit pulse conversion scale, function round is a bracket function.

The sensing neuron output signal that CO2 concentration sensor, infrared sensor and relative humidity sensor are corresponding gets into implicit neuron, and implicit neuron adopts the consistance coding.Hidden neuron carries out the logical computing with the pulse of input, and only when two pulses occurred simultaneously, hidden neuron was just in pulse of corresponding time point output.

The formula of the output neuron computing activation potential of Spiking neural network is following:

$u_i\left(t\right)=\underset{j}{\mathrm{\&Sigma;}}\underset{t_j^{\left(l\right)}}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}^g\mathrm{\&epsiv;}(t-t_j^{\left(l\right)})+\underset{k}{\mathrm{\&Sigma;}}\underset{t_k^{\left(f\right)}}{\mathrm{\&Sigma;}}{w}_{\mathrm{ik}}^y\mathrm{\&epsiv;}(t-t_k^{\left(l\right)})$

In the formula,

Be the weights that are connected between sensing neuron and the output neuron,

For being connected weights, u between implicit neuron and the output neuron _i(t) be the input pulse activation potential of i output neuron,

Time when expression sensor neuron j exports l pulse,

Time when the implicit neuron k of expression exports l pulse, the ε function does

τ wherein ₃Be time constant; Output neuron calculates through input signal being activated potential, confirms the output umber of pulse, personnel's number that promptly should the zone.

Spiking neural network among the present invention adopts the Hebbian learning algorithm, and it is following to embody formula:

$w_{\mathrm{ij}}\left(t\right)=d\&times;w_{\mathrm{ij}}(t-1)+r\&times;\underset{t_j^{\mathrm{output}},t_i^{\mathrm{input}}}{\mathrm{\&Sigma;}}W\left(\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{ij}}\right)$

In the formula, w _IjBe the weights that are connected between sensing neuron, implicit neuron and output neuron, $W\left(\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}\right)=\left\{\begin{array}{c}[c_1(1-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\widetilde{\mathrm{\&tau;}}}_1})+c_2(1-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\widetilde{\mathrm{\&tau;}}}_2})]\&CenterDot;\mathrm{Exp}\left(\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{\mathrm{\&tau;}}\right),\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}\&GreaterEqual;0\\ {\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="HDA0000117711460000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1\mathrm{Exp}(-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\mathrm{\&tau;}}_1})+c_2\mathrm{Exp}(-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\mathrm{\&tau;}}_2}),\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}<0\end{array}\right.,$ c ₁, c ₂, τ, τ ₁And τ ₂Be constant,

D is an attenuation rate, and r is a learning rate, Δ τ _IjIt is i input pulse and j mistiming of exporting between the pulse

Although principle of the present invention is combined embodiment to show and describes; But it will be understood to those of skill in the art that under the situation that does not depart from principle of the present invention and essence; Change these embodiments; Like every layer of neuron number of Spiking neural network model, the change of input layer coding form etc., its scope also falls in claim of the present invention and the equivalent institute restricted portion thereof.

Claims (5)

1. the personal information fusion method based on the Spiking neural network is characterized in that,

1) utilize CO2 concentration sensor, infrared sensor, personnel RFID, camera and the relative humidity sensing etc. that possibly exist in the space zone to carry out the collection of personnel's information of number, data deposit real-time data base in;

2) reading of data from real-time data base is carried out pre-service, removes unreasonable data, is mapped as input vector, and relational database divides into groups to deposit in;

3), be designed to comprise the three-layer network structure of input layer, hidden layer and output layer, the corresponding sensing neuron of one of them sensor according to sensor input type design Spiking neural network structure model; CO ₂The sensing neuron output signal that concentration sensor, infrared sensor and relative humidity sensor are corresponding gets into implicit neuron, and the corresponding sensing neuron signal of personnel RFID and camera directly gets into output neuron; The output umber of pulse of the corresponding output neuron of space area people number;

4) reading of data from relational database, data multiply by degree of belief weighted value input Spiking neural network, and the sensing neuron adopts frequency coding;

5) implicit neuron adopts the consistance coding, and promptly implicit neuron carries out the logical computing with the pulse of input, and only when two pulses occurred simultaneously, implicit neuron was just in pulse of corresponding time point output;

6) adopt the Hebbian learning algorithm, upgrade the connection weights between the sensing neuron of Spiking neural network model, implicit neuron and the output neuron, improve system action;

7) the output neuron input signal is activated potential and calculate, confirm the output umber of pulse of output neuron, personnel's number that promptly should the zone.

2. the personal information fusion method based on the Spiking neural network as claimed in claim 1 is characterized in that, in the said step 4), and said frequency coding, represent with following formula:

$I_i=\mathrm{round}\left(\frac{r_i-s_i}{u_i}\right)$

, in the formula, I _iBe i the neuronic output umber of pulse of sensing, r _iBe sensor full range, s _iBe sensor input, u _iBe the unit pulse conversion scale, function round is a bracket function.

3. the personal information fusion method based on the Spiking neural network as claimed in claim 1; It is characterized in that; The implicit neuron of said Spiking neural network adopts the consistance coding; Promptly implicit neuron carries out the logical computing with the pulse of input, and only when two pulses occurred simultaneously, implicit neuron was just in pulse of corresponding time point output.

4. the personal information fusion method based on the Spiking neural network as claimed in claim 1 is characterized in that, in the said step 6), the Hebbian learning algorithm is following: algorithm is following:

$w_{\mathrm{ij}}\left(t\right)=d\&times;w_{\mathrm{ij}}(t-1)+r\&times;\underset{t_j^{\mathrm{output}},t_i^{\mathrm{input}}}{\mathrm{\&Sigma;}}W\left(\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{ij}}\right)$

In the formula, w _IjBe the weights that are connected between sensing neuron, implicit neuron and output neuron, $W\left(\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}\right)=\left\{\begin{array}{c}[c_1(1-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\widetilde{\mathrm{\&tau;}}}_1})+c_2(1-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\widetilde{\mathrm{\&tau;}}}_2})]\&CenterDot;\mathrm{Exp}\left(\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{\mathrm{\&tau;}}\right),\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}\&GreaterEqual;0\\ c_1\mathrm{Exp}(-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\mathrm{\&tau;}}_1})+c_2\mathrm{Exp}(-\frac{\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}}{{\mathrm{\&tau;}}_2}),\mathrm{\&Delta;}{\mathrm{\&tau;}}_{\mathrm{Ij}}<0\end{array}\right.,$ c ₁, c ₂, τ, τ ₁And τ ₂Be constant,

D is an attenuation rate, and r is a learning rate, Δ τ _IjIt is i input pulse and j mistiming of exporting between the pulse

5. the personal information fusion method based on the Spiking neural network as claimed in claim 1 is characterized in that the formula of computing activation potential is following in the said step 7):

$u_i\left(t\right)=\underset{j}{\mathrm{\&Sigma;}}\underset{t_j^{\left(l\right)}}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}^g\mathrm{\&epsiv;}(t-t_j^{\left(l\right)})+\underset{k}{\mathrm{\&Sigma;}}\underset{t_k^{\left(f\right)}}{\mathrm{\&Sigma;}}{w}_{\mathrm{ik}}^y\mathrm{\&epsiv;}(t-t_k^{\left(l\right)})$

In the formula,

Be the weights that are connected between sensing neuron and the output neuron, For being connected weights, u between implicit neuron and the output neuron _i(t) be the input pulse activation potential of i output neuron,

Time when expression sensor neuron j exports l pulse,

Time when the implicit neuron k of expression exports l pulse, the ε function does

τ wherein ₃Be time constant; Output neuron calculates through input signal being activated potential, confirms the output umber of pulse, personnel's number that promptly should the zone.

Images