CN109302309B

CN109302309B - Passive sensing indoor positioning method based on deep learning

Info

Publication number: CN109302309B
Application number: CN201810977963.5A
Authority: CN
Inventors: 张涛; 谢志军; 何静; 陈海明
Original assignee: Ningbo University
Current assignee: Ningbo University
Priority date: 2018-08-27
Filing date: 2018-08-27
Publication date: 2021-07-30
Anticipated expiration: 2038-08-27
Also published as: CN109302309A

Abstract

The invention discloses a passive sensing indoor positioning method based on deep learning, which comprises the steps of firstly, acquiring and generating CSI amplitude matrixes of all indoor positions, constructing a deep neural network model, dividing the deep neural network model into four restricted Boltzmann machines, then respectively training the four restricted Boltzmann machines and carrying out iterative optimization updating on parameters obtained by training on the basis of the CSI amplitude matrixes, obtaining fingerprints of all indoor positions as a fingerprint database, and finally realizing the positioning of indoor personnel on the basis of the obtained fingerprint database; the method has the advantages of simple positioning process, high positioning precision and lower cost.

Description

Passive sensing indoor positioning method based on deep learning

Technical Field

The invention relates to a passive sensing positioning method, in particular to a passive sensing indoor positioning method based on deep learning.

Background

With the rapid development of intelligent technology, various intelligent terminal devices enter people's lives. Meanwhile, various intelligent control methods or intelligent monitoring methods matched with the intelligent terminal devices are rapidly developed. The positioning method is an intelligent monitoring method, is mainly used for monitoring people or objects and the like, brings great changes to the life of people, and is widely applied. The existing positioning methods are mainly divided into two categories, namely indoor positioning methods and outdoor positioning methods according to different application areas of the positioning methods. At present, outdoor positioning methods are mature, and indoor positioning methods are generally used for high-precision positioning of indoor personnel, such as people who want to find themselves or lose children quickly in a strange large-scale shopping mall or an office; police seizes a prisoner and hopes to quickly confirm the relationship between a criminal and cognition; the real-time positions of the old and the children at home are remotely monitored, and the like. The application area is limited to a certain indoor space, and the indoor positioning method faces a great challenge due to the complex indoor environment.

With the development of the WIFI technology, channel State information CSI (channel State information) is successfully applied to the field of indoor wireless positioning, and a passive sensing indoor positioning method based on CSI is subject toThere is a wide range of concerns. The existing passive sensing indoor positioning method based on CSI mainly comprises two methods: the first method first divides the room to be monitored evenly into L zones (20)<L<30) Sampling, selecting the center of each region as its position, and recording the positions of the L regions as S₁～S_LConfiguring a router which is not provided with a password and is communicated with a network in a room to be monitored, configuring a computer in the monitoring room, wherein the computer is provided with a wireless network card intel 5300, a ubuntu12.04 system and a CSI Tools, and then a tester stands at a position S in sequence₁～S_LWhere another tester separately obtains the position S by computer₁～S_LThe computer acquires the test data at the current position in real time according to the movement of the personnel when the personnel are positioned indoors, compares the test data at the current position with each reference fingerprint in the indoor position fingerprint library, and takes the position corresponding to the reference fingerprint with the minimum Euclidean distance between the test data and the current test data as the current positioning position; wherein the position S is acquired_lThe specific procedure for the test data at (1, 2, …, L) is: one tester standing at position S_lAnother worker collects N signal data packets through a ping router at a speed of 50 times per second of a computer and stores the N signal data packets in CSI Tools, wherein N is an integer larger than 800 and smaller than 1000, matlab software is used for opening a self-contained matlab folder in the CSI Tools, a read _ bf _ file function in the matlab folder is used for reading a CSI & dat file to obtain N signal data packets, then a get _ scaled _ CSI function is used for respectively opening N signal data packets to obtain N CSI data matrices, each CSI data matrix comprises 30 rows and 3 columns of CSI data, the CSI data are in a complex form, an abs function is used for respectively obtaining CSI amplitude matrices corresponding to the N CSI data matrices, each CSI amplitude matrix comprises 30 rows and 3 columns of CSI amplitude data, an average value of three CSI amplitude data in the same row in each CSI amplitude matrix is respectively obtained, and N CSI average amplitude matrices are obtained, the CSI average amplitude comprises 30 rows and 1 column of average amplitude data, averaging the magnitude moments of each CSIThe square accumulation of 30 CSI average amplitude data in the array is used as the power of each average amplitude matrix to obtain the power corresponding to N CSI average amplitude matrices, the average power corresponding to the N CSI average amplitude matrices is obtained and is used as the position S_lThe test data at (a), i.e. the reference fingerprint. The second method is to construct a position fingerprint database by adopting the same steps as the first method, then to find K position fingerprints with the closest Euclidean distance by utilizing a K-nearest-neighbor method (K represents K positions with the closest characteristics), to mark the positions of the CSI amplitude matrixes of the K position fingerprints respectively, and to find the position with the largest quantity as the current positioning coordinate.

The first CSI-based passive sensing indoor positioning method directly compares the test data at the current position with each reference fingerprint in the indoor position fingerprint database to perform positioning, although the positioning process is simpler, the diversity of the data is also reduced, thereby resulting in low positioning accuracy. The second passive sensing indoor positioning method based on the CSI classifies CSI data by adopting a K-nearest-neighbor method, the calculated amount is small, the precision is improved compared with that of the first passive sensing indoor positioning method based on the CSI, all training values need to be stored, and the cost is greatly increased.

Disclosure of Invention

The invention aims to solve the technical problem of providing a passive sensing indoor positioning method based on deep learning, which has the advantages of simple positioning process, high positioning accuracy and lower cost.

The technical scheme adopted by the invention for solving the technical problems is as follows: a passive sensing indoor positioning method based on deep learning comprises the following steps:

firstly, data acquisition:

step 1: preparing a computer, inserting the commercial wireless network card intel 5300 into a wireless network card expansion slot of a computer host, installing an ubuntu12.04 system and a Linux 802.11n CSI Tool in the computer host, and placing the computer in a monitoring room;

step 2: the room to be positioned is divided into L areas along the transverse direction and the longitudinal direction,the transverse direction is divided into L1 parts along the maximum length direction, the longitudinal direction is divided into L2 parts along the maximum length direction, L1 xL 2, L1 and L2 are integers which are larger than 4 and smaller than 6 respectively, the center of each region is selected as the position of the region, if the region is a non-centrosymmetric region, the intersection point of the transverse sideline and the perpendicular bisector of the longitudinal sideline of the region is taken as the center, and the position of the first region is recorded as S_l1,2, ·, L; selecting any indoor point as an origin, constructing a two-dimensional coordinate system of the indoor area by taking the transverse direction as the x-axis direction and the longitudinal direction as the y-axis direction, and obtaining the first area S through measurement_lCentral point L of_lIs expressed as (x)_l,y_l) A router which is not provided with a password and is communicated with a network is arranged at a certain position in a room to be positioned;

step 3: opening the ubuntu12.04 system in the computer host, connecting the commercial wireless network card intel 5300 arranged on the computer host with the router through a wireless network, and sequentially standing one tester at the position S₁～S_LThe other tester obtains the position S in turn at the computer host₁～S_LThe specific process of the test data is as follows: one tester standing at position S_lWhen the computer is in operation, another tester collects the position S through the ping router at the speed of 50 times per second of the computer host_lTakes N signal data packets as position S_lTest data of (2), position S_lSaving the test data of (a) as a file with a suffix name of dat, and naming the file as csil, wherein N is an integer greater than 800 and less than 1000;

processing data: installing matlab in the ubuntu12.04 system of the computer host, opening a self-contained matlab folder in a Linux 802.11n CSI Tool by using matlab software, and comparing the position S acquired in the step (i)₁～S_LRespectively processing the test data to obtain the position S₁～S_LThe specific process of the amplitude matrix is as follows: reading csil.dat file by adopting read _ bf _ file function in matlab folder, and acquiring position S_lThe number N of signal data packets of (a),then respectively opening the position S by adopting a get _ scaled _ csi function_lN signal data packets of (A) to obtain a position S_lN CSI data matrices, position S_lEach CSI data matrix comprises 30 rows and 3 columns of CSI data, the CSI data are in a complex form, and the position S is respectively calculated by adopting an abs function_lObtaining a position S by using the CSI amplitude matrixes corresponding to the N CSI data matrixes_lAt N CSI magnitude matrices, position S_lEach CSI amplitude matrix respectively comprises 30 rows and 3 columns of CSI amplitude data;

constructing a deep neural network model: the neural network model comprises a visible layer, a first hidden layer, a second hidden layer, a third hidden layer and a fourth hidden layer which are sequentially arranged from top to bottom, wherein the visible layer is provided with K0 neural units, K0 is N30 3 12, and represents a multiplication symbol; said first hidden layer has K1 neural units, 300< K1<500, said second hidden layer has K2 neural units, 200< K2< K1; the third hidden layer has K3 nerve units, 100< K3< K2; the fourth hidden layer has K4 nerve units, 50< K4< K3; all the nerve units in the same layer are not connected with each other, all the nerve units in the two adjacent layers are connected with each other, and each nerve unit has two states: the neural network comprises an activated state and a closed state, wherein the state value of the neural unit in the activated state is 1, and the state value of the neural unit in the closed state is 0; the visible layer and the first hidden layer form a first limited Boltzmann machine, the first hidden layer and the second hidden layer form a second limited Boltzmann machine, the second hidden layer and the third hidden layer form a third limited Boltzmann machine, and the third hidden layer and the fourth hidden layer form a fourth limited Boltzmann machine;

extracting the position S_lThe fingerprint of (2) is prepared by the following specific processes:

step C1, training the first limited Boltzmann machine:

c1-1, constructing a matrix for storing the state quantity of the visible layer, and marking the matrix as v; building a state quantity for storing a first hidden layerA matrix, which is marked as ha; constructing a matrix for storing the offset of the visible layer, and marking the matrix as a; constructing a matrix for storing the offset of the first hidden layer, and recording the matrix as b; constructing a weight matrix storing the connection weight between the first hidden layer and the visible layer, and recording the weight matrix as w¹(ii) a Wherein v ═ v (v)₁,v₂,v₃,...,v_K0)^T，v_k0Represents the state quantity of the K0 th neural unit in the visible layer, the superscript T represents the transpose of the matrix, K0 is 1,2, K0; ha ═ ha (ha)₁,ha₂,ha₃,...,ha_K1)^T，ha_k1Represents the state quantity of the K1 th neural unit in the first hidden layer, K1 being 1,2.., K1; a ═ a₁,a₂,a₃,...,a_K0)^T，a_k0Representing the offset of the k0 th nerve unit in the visible layer, and adopting a random function to a_k0Performing initialization to make a_k0Is a random number between 0 and 0.1; b ═ b₁,b₂,b₃,...,b_K1)^T，b_k1Representing the offset of the k1 th neural unit in the first hidden layer by using a random function pair b_k1Performing initialization to b_k1Is a random number between 0 and 0.1;

representing the connection weight between the k1 th nerve unit in the first hidden layer and the k0 th nerve unit in the visible layer by adopting a random function pair

Carry out initialization to

Is a random number between 0 and 0.1;

c1-2, defining the sequence of each CSI assigned data in the CSI amplitude matrix: locate the CSI amplitude matrix at 1 st column and 1 st rowCSI amplitude data from CSI amplitude data to a 30 th row are used as 1 st CSI amplitude data to a 30 th CSI amplitude data of the matrix, CSI amplitude data from a 2 nd row from the 1 st row CSI amplitude data to a 30 th row are used as 31 st CSI amplitude data to a 60 th CSI amplitude data of the matrix, and CSI amplitude data from a 3 rd row from the 1 st row CSI amplitude data to the 30 th row are used as 61 st CSI amplitude data to 90 CSI amplitude data of the matrix; determining a location S_lThe 12-bit binary number corresponding to the qth CSI amplitude data of the Nth CSI amplitude matrix is as follows: determination of position S_lIf the q-th CSI amplitude data of the N' -th CSI amplitude matrix is an integer or a decimal if the position S_lThe q-th CSI amplitude data of the N' -th CSI amplitude matrix is an integer, the integer is converted into a binary number, 4 0S are supplemented to the right of the obtained binary number to obtain a new binary number, whether the number of the obtained new binary number is equal to 12 or not is judged, and if the number of the obtained new binary number is equal to 12, the new binary number is used as a position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; if the position is more than 12, selecting 12 bits from right to left as the position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; if the number of bits is less than 12, the left side of the new binary number is supplemented with 0 to 12 bits and then the result is used as the position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; if at position S_lIf the q-th CSI amplitude data of the N 'th CSI amplitude matrix is a decimal, converting an integer part and a decimal part of the N' th CSI amplitude matrix into binary numbers respectively, and if the number of bits of the binary numbers obtained by converting the integer part is less than 8, complementing 0 at a high bit of the binary numbers to change the binary numbers into 8 bits; if the number of the binary digits obtained by converting the integer part is more than 8, retaining 8 digits from right to left, and deleting other digits to obtain 8-digit binary digits; if the binary digit obtained by converting the integer part is equal to 8, keeping the binary digit unchanged; if the binary digit number obtained by fractional part conversion is equal to 4, the binary digit number is kept unchanged; if the digit of the binary number obtained by fractional part conversion is less than 4, 0 is complemented at the low bit of the binary number to change the binary number into a 4-bit binary number; number of binary digits obtained if fractional part is convertedIf the number is more than 4, remaining 4 bits from left to right and deleting other bits to obtain a 4-bit binary number, then using the 8-bit binary number obtained after the integer part is processed as a high 8 bit, using the 4-bit binary number obtained after the decimal part is processed as a low 4 bit and splicing the low 4 bit into a 12-bit binary number, and using the spliced 12-bit binary number as a position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; n' ═ 1,2, ·, N; q ═ 1,2, ·, 90; assigning the value of the p-th bit from left to right of the 12-bit binary number corresponding to the q-th CSI amplitude data to v_{12*3*30*(N′-1)+(q-1)*12+p}P1, 2, 12, thereby completing v₁,v₂,v₃,...,v_K0Initial assignment of (1);

c1-3. construct a first training matrix ha ', ha' (ha ') that holds the first training values of the state quantities of the k1 th neural unit in the first hidden layer'₁,ha′₂,…,ha′_K1)^T，ha′_k1A first training value representing a state quantity of a k1 th neural unit in the first hidden layer; constructing a matrix pa storing the probability of first activation of the k1 th neural unit in the first hidden layer, pa ═ pa₁,pa₂,pa₃,...,pa_K1)^T，pa_k1Representing the probability of the first activation of the k1 th neural unit in the first hidden layer; determining a first training matrix ha' of a first state training value of a k1 th neural unit in the first hidden layer and a matrix pa of a first activation probability of a k1 th neural unit in the first hidden layer, wherein the specific process is as follows:

c1-3-1, recording the activation probability of the k1 th nerve unit in the first hidden layer as Pr (k1), and calculating by adopting the formula (1) to obtain Pr (k 1):

where exp denotes an exponential function, Σ being the accumulated sign, b_k1The current value of the offset for the k1 th neural unit in the first hidden layer,

for the current value of the connection weight between the k1 th neuron in the first hidden layer and the k0 th neuron in the visible layer, v_k0The current value of the state quantity of the k0 th nerve unit in the visible layer;

c1-3-2, assigning the current value of Pr (k1) to pa_k1Generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k1) with the random number, and if the current value of Pr (k1) is greater than the random number, assigning 1 to ha'_k1Assigning 0 to ha 'if the current value of Pr (k1) is not greater than the random number'_k1；

C1-4. construct a second training matrix v ', v ═ v'₁,v′₂,…,v′_K0)^T，v′_k0Training values which are state quantities of k0 th neural units in the visible layer; determining a training value v 'of a state quantity of a k0 th neural unit in a visible layer'_k0The specific process is as follows:

c1-4-1, recording the activation probability of the k0 th nerve unit in the visible layer as Pr (k0), and calculating by adopting the formula (2) to obtain Pr (k 0):

wherein, a_k0The current value of the offset for the k0 th neural cell in the visible layer,

is the current value of the connection weight, ha 'between the k1 th neural unit in the first hidden layer and the k0 th neural unit in the visible layer'_k1A first training value of state quantity of a k1 th neural unit in the first hidden layer;

c1-4-2, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k0) with the random number, and if the current value of Pr (k0) is greater than the random number, assigning 1 to v'_k0Assigning 0 to v 'if the current value of Pr (k1) is not greater than the random number'_k0；

C1-5, constructing a third training matrix ha ", ha ″ (ha ″") storing second training values of state quantities of the k1 th neural unit in the first hidden layer₁,ha″₂,…,ha″_K1)^T，ha″_k1A second training value representing a state quantity of a k1 th neural unit in the first hidden layer; constructing a matrix pa ', pa ═ of (pa'₁,pa′₂,…,pa′_K1)^T，pa′_k1Representing a second probability of activation of the k1 th neural unit in the first hidden layer; determining a third training matrix ha 'of a second training value of the state quantity of the k1 th neural unit in the first hidden layer and a matrix pa' of a second activation probability of the k1 th neural unit in the first hidden layer, which comprises the following specific processes:

c1-5-1, the activation probability Pr (k1) of the k1 th nerve cell in the first hidden layer is updated by adopting the formula (3):

wherein, b_k1The current value of the offset for the k1 th neural unit in the first hidden layer,

for the current value of the connection weight between the k1 th neuron in the first hidden layer and the k0 th neuron in the visible layer, v^′ _k0Training values which are state quantities of k0 th neural units in the visible layer;

c1-5-2 assigning the current value of Pr (k1) to pa'_k1Generating a random number between 0 and 1 by adopting a random function, comparing the current value of Pr (k1) with the random number, and assigning 1 to ha' if the current value of Pr (k1) is greater than the random number_k1And simultaneously assigns 1 to ha_k1(ii) a If the current value of Pr (k1) is not greater than the random number, then a value of 0 is assigned to ha ″_k1And simultaneously assigns 0 to ha_k1；

C1-6. will

Added 0.01 x (pa) to the current value of_k1*v_k0-pa′_k1*v′_k0) Is assigned to

To pair

The value of the weight matrix w is updated to obtain an updated weight matrix w¹(ii) a A is to_k0Current value of (v) plus 0.01 x (v)_k0-v′_k0) Is assigned to a_k0To a, a_k0Updating the value of the visible layer to obtain the updated offset a of the visible layer; b is to_k1Added 0.01 x (pa) to the current value of_k1-pa′_k1) Is assigned to b_k1To b is paired_k1Is updated to obtain the updated offset b of the first hidden layer, wherein pa_k1、v_k0、pa′_k1And v'_k0All the values of (a) are the current values thereof;

step C2: training the second limited Boltzmann machine:

c2-1, constructing a matrix for storing the state quantity of the second hidden layer, and marking the matrix as hb; constructing a matrix for storing the offset of the second hidden layer, and recording the matrix as c; constructing a weight matrix for storing a connection weight between the first hidden layer and the visible layer, and recording the weight matrix as w²；hb＝(hb₁,hb₂,hb₃,...,hb_K2)^T，hb_k2A state value representing the K2 th neural unit in the second hidden layer, K2 ═ 1,2.., K2; c ═ c₁,c₂,c₃,...,c_K2)^T，c_k2Representing the offset of the k2 th neural unit in the second hidden layer;

indicating first concealmentThe connection weight between the k1 th nerve unit in the layer and the k2 th nerve unit in the second hidden layer adopts a random function pair c respectively_k2And

is initialized to c_k2And

respectively, random values of 0-0.1;

c2-2. construct a fourth training matrix hb ', hb ' ═ hb '₁,hb′₂,…,hb′_K2)^T，hb′_k2A first training value representing a state quantity of a k2 th neural unit in the second hidden layer; constructing a matrix pb storing first activation probabilities for the k2 th neural unit in the second hidden layer (pb ═ pb)₁,pb₂,pb₃,...,pb_K2)^T，pb_k2Representing the probability of the first activation of the k2 th neural unit in the second hidden layer; determining a fourth training matrix hb' of the first training value of the state quantity of the k2 th neural unit in the second hidden layer and a matrix pb of the first activation probability of the k2 th neural unit in the second hidden layer, specifically comprising the following steps:

c2-2-1, recording the activation probability of the k2 th nerve unit in the second hidden layer as Pr (k2), and calculating by adopting a formula (4) to obtain Pr (k 2):

wherein, cx₂The current value of the offset for the k2 th neural unit in the second hidden layer,

is the current value, ha, of the connection weight between the k1 th neuron in the first hidden layer and the k2 th neuron in the second hidden layer_k1Is the current value of the state quantity of the k1 th neural unit in the first hidden layer;

c2-2-2, assigning the current value of Pr (k2) to pb_k2Generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k2) with the random number, and if the current value of Pr (k2) is greater than the random number, assigning 1 to hb'_k2(ii) a If the current value of Pr (k2) is not greater than the random number, then 0 is assigned to hb'_k2(ii) a Completing the assignment of a fourth training matrix hb';

c2-3, constructing a fifth training matrix ha '", ha'" for (ha "") storing third training values of state quantities of the k1 th neural unit in the first hidden layer₁,ha″₂,…,ha″_K1)^T，ha″_k1A third training value representing a state quantity of a k1 th neural unit in the first hidden layer; determining a fifth training matrix ha' ″ storing a third training value of the state quantity of the k1 th neural unit in the first hidden layer by the following specific process:

c2-3-1, the activation probability Pr (k1) of the k1 th nerve cell in the first hidden layer is updated by adopting the formula (5):

is the current value of the connection weight, hb ', between the k1 th neural unit in the first hidden layer and the k2 th neural unit in the second hidden layer'_k2A first training value of state quantity of a k2 th neural unit in a second hidden layer;

c2-3-2, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k1) with the random number, and if the current value of Pr (k1) is greater than the random number, assigning 1 to ha ″'_k1(ii) a If the current value of Pr (k1) is not greater than the random number, then a value of 0 is assigned to ha ″'_k1；

C2-4. constructing the second training value for storing the state quantity of the k2 th neural unit in the second hidden layerThe sixth training matrix hb ", hb ═ hb ″ (hb ″)₁,hb″₂,…,hb″_K2)^T，hb″_k2A second training value representing a state quantity of a k2 th neural unit in the second hidden layer; constructing a matrix pb ', pb ' ═ b (pb '₁,pb′₂,…,pb′_K2)^T，pb′_k2The method comprises the following steps of representing the second activation probability of the k2 th neural unit in the second hidden layer, determining a sixth training matrix hb 'of a second training value for storing the state quantity of the k2 th neural unit in the second hidden layer and a matrix pb' of the second activation probability of the k2 th neural unit in the second hidden layer, wherein the specific process is as follows:

c2-4-1, the activation probability Pr (k2) of the k2 th nerve cell in the second hidden layer is updated by adopting the formula (6):

wherein, c_k2The current value of the offset for the k2 th neural unit in the second hidden layer,

is the current value of the connection weight, ha ″, between the k1 th neural unit in the first hidden layer and the k2 th neural unit in the second hidden layer'_k1A third training value of state quantities of a k1 th neural unit in the first hidden layer;

c2-4-2 assigning the current value of Pr (k2) to pb'_k2Generating a random number between 0 and 1 by adopting a random function, comparing Pr (k2) with the current value with the random number, and if the current value of Pr (k2) is greater than the random number, assigning 1 to hb ″', respectively_k2And hb_k2If the current value of Pr (k2) is not greater than the random number, then 0 is assigned to hb ″', respectively_k2And hb_k2；

C2-5. will

Current value of plus0.01*(pb_k2*ha_k1-pb′_k2*ha″′_k1) Is assigned to

To pair

Updating to obtain updated weight matrix w²B is mixing_k1Current value of (a) plus 0.01 (ha)_k1-ha″′_k1) Is assigned to b_k1To b is paired_k1Updating to obtain the offset b of the updated first hidden layer, and updating c_k2Current value of plus 0.01 (pb)_k2-pb′_k2) To c_k2To c for_k2Updating to obtain the offset c of the updated second hidden layer; wherein pb_k2、ha_k1、pb′_k2And ha'_k1The values of (a) are their current values, respectively;

step C3: training the third limited Boltzmann machine:

c3-1, constructing a matrix for storing the state quantity of the third hidden layer, and marking the matrix as hc; constructing a matrix for storing the offset of the third hidden layer, and recording the matrix as d; constructing a weight matrix for storing the connection weight between the second hidden layer and the third hidden layer, and recording the weight matrix as w³；hc＝(hc₁,hc₂,hc₃,...,hc_K3)^T，hc_k3A state value representing the K3 th neural unit in the third hidden layer, K3 ═ 1,2.., K3; d ═ d (d)₁,d₂,d₃,...,d_K3)^T，d_k3Representing the offset of the k3 th neural unit in the third hidden layer;

representing the connection weight between the k2 th neural unit in the second hidden layer and the k3 th neural unit in the third hidden layer respectivelyMachine function pair d_k3And

initializing to d_k3And

respectively, random values of 0-0.1;

c3-2. construct a seventh training matrix hc ', hc' ═ hc 'of the first training values of the state quantities of the k3 th neural unit in the third hidden layer'₁,hc′₂,…,hc′_K3)^T，hc′_k3A first training value representing a state quantity of a k3 th neural unit in the third hidden layer; constructing a matrix pc storing the probability of first activation of the k3 th neural unit in the third hidden layer (pc ═ c)₁,pc₂,pc₃,...,pc_K3)^T，pc_k3Represents the first activation probability of the k3 th neural unit in the third hidden layer; determining a seventh training matrix hc' storing the first training value of the state quantity of the k3 th neural unit in the third hidden layer and a matrix pc storing the first activation probability of the k3 th neural unit in the third hidden layer, wherein the specific process is as follows:

c3-2-1, recording the activation probability of the k3 th nerve unit in the third hidden layer as Pr (k3), and calculating by adopting a formula (7) to obtain Pr (k 3):

wherein d is_k3The current value of the offset for the k3 th neural unit in the third hidden layer,

for the current value of the connection weight, hb, between the k2 th neuron in the second hidden layer and the k3 th neuron in the third hidden layer_k2Is the current value of the state value of the k2 th neural unit in the second hidden layer;

c3-2-2, assigning the current value of Pr (k3) to pc_k3Obtaining an updated matrix pc, usingThe machine function generates a random number between 0 and 1, compares the current value of Pr (k3) with the random number, and assigns 1 to hc 'if the current value of Pr (k3) is greater than the random number'_k3(ii) a If the current value of Pr (k3) is not greater than the random number, then 0 is assigned to hc'_k3；

C3-3 constructing an eighth training matrix hb '", hb'" 'of (hb' "') that deposits a third training value for the state quantity of the k2 th neural unit in the second hidden layer'₁,hb″′₂,…,hb″′_K2)^T，hb″′_k2And determining an eighth training matrix hb' ″ for storing the third training value of the state quantity of the k2 th neural unit in the second hidden layer by using the third training value representing the state quantity of the k2 th neural unit in the second hidden layer, which comprises the following specific processes:

c3-3-1, updating the activation probability Pr (k2) of the k2 th neural unit in the second hidden layer by using the formula (8):

is the current value of the connection weight, hc 'between the k2 th neural unit in the second hidden layer and the k3 th neural unit in the third hidden layer'_k3A first training value of state quantity of a k3 th neural unit in a third hidden layer;

c3-3-2, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k2) with the random number, and if the current value of Pr (k2) is greater than the random number, assigning 1 to hb ″'_k2(ii) a If the current value of Pr (k2) is not greater than the random number, then 0 is assigned to hb ″'_k2；

C3-4, constructing a ninth training matrix hc ", hc ═ hc ″ (hc ″") that stores the second training values of the state quantities of the k3 th neural unit in the third hidden layer₁,hc″₂,…,hc″_K3)^T，hc″_k3A second training value representing a state quantity of a k3 th neural unit in the third hidden layer; constructing a matrix pc ', pc ═ c (pc'₁,pc′₂,…,pc′_K3)^T，pc′_k3Represents the second activation probability of the k3 th neural unit in the third hidden layer; the ninth training matrix hc ″ storing the second training value of the state quantity of the k3 th neural unit in the third hidden layer and the matrix pc' of the second activation probability of the k3 th neural unit in the third hidden layer are determined by the following specific processes:

c3-4-1, recording the activation probability of the k3 th nerve unit in the third hidden layer as Pr (k3), and calculating by adopting the formula (9) to obtain Pr (k 3):

is the current value of the connection weight, hb ″, between the k2 th neural unit in the second hidden layer and the k3 th neural unit in the third hidden layer'_k2A third training value which is the state quantity of the k2 th neural unit in the second hidden layer;

c3-4-2 assigning the current value of Pr (k3) to pc'_k3Generating a random number between 0 and 1 by adopting a random function, comparing the current value of Pr (k3) with the random number, and if the current value of Pr (k3) is greater than the random number, assigning 1 to hc ″', respectively_k3And hc and_k3(ii) a If the current value of Pr (k3) is not greater than the random number, then 0 is assigned to hc ″', respectively_k3And hc and_k3；

c3-5. will

Current value of (c) plus 0.01 × (pc)_k3*hb_k2-pc′_k3*hb″′_k2) Is assigned to

To pair

Updating to obtain updated weight matrix w³(ii) a C is to_k2Current value of plus 0.01 (hb)_k2-hb″′_k2) To c_k2To c for_k2Updating to obtain the offset c of the updated second hidden layer; will d_k3Current value of (c) plus 0.01 × (pc)_k3-pc′_k3) Is assigned to d_k3To d is paired_k3Updating to obtain the offset d of the updated third hidden layer, wherein pc_k3、hb_k2、pc′_k3And hb'_k2The values of (a) are their current values, respectively;

step C4: training the fourth limited Boltzmann machine:

c4-1, constructing a matrix for storing the state quantity of the fourth hidden layer, and marking the matrix as hd; constructing a matrix for storing the offset of the fourth hidden layer, and recording the matrix as e; constructing a weight matrix storing the connection weight between the third hidden layer and the fourth hidden layer, and recording the weight matrix as w⁴；hd＝(hd₁,hd₂,hd₃,...,hd_K4)^T，hd_k4A state value representing the K4 th neural unit in the fourth hidden layer, K4 ═ 1,2.., K4; e ═ e (e)₁,e₂,e₃,...,e_K4)^T，e_k4Represents the offset of the k4 th neural unit of the fourth hidden layer,

representing the connection weight between the k3 th neural unit in the third hidden layer and the k4 th neural unit in the fourth hidden layer by adopting a random function_k4And

respectively initializing to be random values of 0-0.1;

c4-2. construct a tenth training matrix hd 'that stores the first training values of the state quantities of the k4 th neural unit in the fourth hidden layer, hd ═ hd'₁,hd′₂,…,hd′_K4)^T，hd′_k4A first training value representing a state quantity of a k4 th neural unit in the fourth hidden layer; constructing a matrix pd storing the probability of the first activation of the k4 th neural unit in the fourth hidden layer (pd ═ pd)₁,pd₂,pd₃,...,pd_K4)^T， pd_k4Represents the probability of the first activation of the k4 th neural unit in the fourth hidden layer; the specific process of determining the tenth training matrix hd' storing the first training value of the state quantity of the k4 th neural unit in the fourth hidden layer and the matrix pd storing the first activation probability of the k4 th neural unit in the fourth hidden layer is as follows:

c4-2-1, recording the activation probability of the k4 th nerve unit in the fourth hidden layer as Pr (k4), and calculating by adopting a formula (10) to obtain Pr (k 4):

wherein e is_k4The current value of the offset for the k4 th neural unit of the fourth hidden layer,

is the current value of the connection weight, hc, between the k3 th neuron in the third hidden layer and the k4 th neuron in the fourth hidden layer_k3The current value of the state quantity of the k3 th neural unit in the third hidden layer;

c4-2-2, assigning the current value of Pr (k4) to pd_k4Generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k4) with the random number, and if Pr (k4) is greater than the random number, assigning 1 to hd'_k4Assigning 0 to hd 'if the current value of Pr (k4) is not greater than the random number'_k4；

C4-3 constructing an eleventh training matrix hc '", hc'" ═ h ″ ', of third training values of state quantities of k3 th neural units in a third hidden layer'₁,hc″′₂,…,hc″′_K3)^T，hc″′_k3And determining an eleventh training matrix hc' for storing the third training value of the state quantity of the k3 th neural unit in the third hidden layer by using the third training value representing the state quantity of the k3 th neural unit in the third hidden layer, wherein the specific process is as follows:

c4-3-1, the activation probability Pr (k3) of the k3 th nerve cell in the third hidden layer is updated by adopting the formula (11):

is the current value of the connection weight, hd ', between the k3 th neural unit in the third hidden layer and the k4 th neural unit in the fourth hidden layer'_k4A first training value of state quantity of a k4 th neural unit in a fourth hidden layer;

c4-3-2, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k3) with the random number, and if the current value of Pr (k3) is greater than the random number, assigning 1 to hc ″'_k3If the current value of Pr (k3) is not greater than the random number, then 0 is assigned to hc'_k3；

C4-4, constructing a twelfth training matrix hd ", hd" ═ hd ″, which stores the second training values of the state quantities of the k4 th neural unit in the fourth hidden layer₁,hd″₂,…,hd″_K4)^T，hd″_k4A second training value representing a state quantity of a k4 th neural unit in the fourth hidden layer; constructing a matrix pd ', pd' ═ of (pd ') of the second activation probability of the k4 th neural unit in the fourth hidden layer'₁,pd′₂,…,pd′_K4)^T， pd′_k4Represents the second activation probability of the k4 th neural unit in the fourth hidden layer; determining a twelfth training matrix hd 'storing a second training value of the state quantity of the k4 th neural unit in the fourth hidden layer and a matrix pd' of a second activation probability of the k4 th neural unit in the fourth hidden layer, which comprises the following specific processes:

c4-4-1, updating the activation probability Pr (k4) of the k4 th neural unit in the fourth hidden layer by adopting the formula (12):

is the current value, hc ″, of the connection weight between the k3 th neural unit in the third hidden layer and the k4 th neural unit in the fourth hidden layer'_k3A third training value which is the state quantity of the k3 th neural unit in the third hidden layer;

c4-4-2 assigning the current value of Pr (k4) to pd'_k4Generating a random number between 0 and 1 by adopting a random function, comparing the current value of Pr (k4) with the random number, and if the current value of Pr (k4) is greater than the random number, assigning 1 to hd ″', respectively_k4And hc and_k4if the current value of Pr (k4) is not greater than the random number, then 0 is assigned to hd ″', respectively_k4And hc and_k4；

c4-5. will

Current value of plus 0.01 x (pd)_k4*hc_k3-pd′_k4*hc″′_k3) Is assigned to

To pair

Updating to obtain updated weight matrix w⁴(ii) a Will d_k3Current value of (d) plus 0.01 (hc)_k3-hc″′_k3) Is assigned to d_k3To d is paired_k3Updating to obtain the updated offset d of the third hidden layer; e is to be₄₃Current value of plus 0.01 x (pd)_k4-pd′_k4) To e_k4To e is aligned with_k4Updating to obtain the offset e of the updated fourth hidden layer; pdk therein₄、hc_k3、pd′_k4And hc'_k3The values of (a) are their current values, respectively;

step C5: setting an iteration variable Y, and initializing the iteration variable Y to enable the initial value of the iteration variable Y to be 1;

and step C6, carrying out the Y-th iteration updating on the activation probability of each nerve unit in the visible layer, the first hidden layer, the second hidden layer, the third hidden layer and the fourth hidden layer, wherein the specific process is as follows:

c6-1, updating the activation probability Pr (k1) of the k1 th nerve cell in the first hidden layer by adopting the formula (13):

c6-2, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k1) with the random number, and if the current value of Pr (k1) is larger than the random number, assigning 1 to ha_k1Pair ha_k1Updating is carried out; if the current value of Pr (k1) is not greater than the random number, then a value of 0 is assigned to ha_k1Pair ha_k1Updating is carried out;

c6-3, updating the activation probability Pr (k2) of the k2 neural unit in the second hidden layer by adopting the formula (14):

c6-4, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k2) with the random number, and if the current value of Pr (k2) is larger than the random number, assigning 1 to hd_k2To hb_k2Updating is carried out; if the current value of Pr (k2) is not greater than the random number, then a value of 0 is assigned to hb_k2To hb_k2Updating is carried out;

c6-5, updating the activation probability Pr (k3) of the k3 neural unit in the third hidden layer by adopting the formula (15):

for the current value of the connection weight, hb, between the k2 th neuron in the second hidden layer and the k3 th neuron in the third hidden layer_k2Is the current value of the state quantity of the k2 th nerve unit in the second hidden layer;

c6-6, generating a random number between 0 and 1 by using a random function, and comparing the current value of Pr (k3) with the random numberThe random numbers are compared and if the current value of Pr (k3) is greater than the random number, then 1 is assigned to hc_k3To hc_k3Updating is carried out; if the current value of Pr (k3) is not greater than the random number, then a value of 0 is assigned to hc_k3To hc_k3Updating is carried out;

c6-7, updating the activation probability Pr (k4) of the k4 neural cell in the fourth hidden layer by adopting the formula (16):

wherein e is_k4The current value of the offset for the k4 th neural unit in the fourth hidden layer,

c6-8, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k4) with the random number, and if the current value of Pr (k4) is larger than the random number, assigning 1 to hd_k4To hd_k4Updating is carried out; if the current value of Pr (k4) is not greater than the random number, then a value of 0 is assigned to hd_k4To hd_k4Updating is carried out;

c6-9, updating the activation probability Pr (k3) of the k3 th neural cell in the third hidden layer again by adopting the formula (17):

the k3 th nerve cell in the third hidden layer and the fourth hidden layerCurrent value of connection weight between k4 neural units, hd_k4Is the current value of the state quantity of the k4 th neural unit in the fourth hidden layer;

c6-10, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k3) with the random number, and if the current value of Pr (k3) is larger than the random number, assigning 1 to hc_k3To hc_k3Updating is carried out; if the current value of Pr (k3) is not greater than the random number, then a value of 0 is assigned to hc_k3To hc_k3Updating is carried out;

c6-11, updating the activation probability Pr (k2) of the k2 neural cell in the second hidden layer again by adopting the formula (18):

is the current value of the connection weight, hc, between the k2 th neuron in the second hidden layer and the k3 th neuron in the third hidden layer_k3The current value of the state quantity of the k3 th neural unit in the third hidden layer;

c6-12, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k2) with the random number, and if the current value of Pr (k2) is larger than the random number, assigning 1 to hb_k2To hb_k2Updating is carried out; if the current value of Pr (k2) is not greater than the random number, then a value of 0 is assigned to hb_k2To hb_k2Updating is carried out;

c6-13, updating the activation probability Pr (k1) of the k1 th neural cell in the first hidden layer again by adopting the formula (19):

for the current value of the connection weight, hb, between the k1 th neuron in the first hidden layer and the k2 th neuron in the second hidden layer_k2Is the current value of the state quantity of the k2 th nerve unit in the second hidden layer;

c6-14, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k1) with the random number, and if the current value of Pr (k1) is larger than the random number, assigning 1 to ha_k1Pair ha_k1Updating is carried out; if the current value of Pr (k1) is not greater than the random number, then a value of 0 is assigned to ha_k1Pair ha_k1Updating is carried out;

c6-15, updating the activation probability Pr (k0) of the k0 th nerve cell in the visible layer by adopting the formula (20):

is the current value of the connection weight, ha, between the k1 th neuron in the first hidden layer and the k0 th neuron in the visible layer_k1Is the current value of the state quantity of the k1 th neural unit in the first hidden layer;

c6-16: generating a random number between 0 and 1 by adopting a random function, comparing the current value of Pr (k0) with the random number, and if the current value of Pr (k0) is greater than the random number, assigning 1 to v'_k0To v'_k0Updating is carried out; if the current value of Pr (k0) is not greater than the random number, then 0 is assigned to v'_k0To v'_k0Updating is carried out;

step C7: creating N reconstructed amplitude matrixes for storing reconstructed CSI amplitude data, wherein each reconstructed amplitude matrix isThe amplitude value matrix can store 30 rows and 3 columns of reconstructed CSI amplitude value data, the sequencing definition of the 30 rows and 3 columns of reconstructed CSI amplitude value data which are stored in each reconstructed amplitude value matrix is the same as that of the CS amplitude value matrix, and v't-v'_11+tAccording to v'_tv′_t+1v′_t+2v′_t+3v′_t+4…v′_t+ ₉v′_t+10v′_t+11The order combination of the 12 bit binary numbers is 12 bit binary numbers, t is 12N +1, N is 0, 1,2, …, K0/12-1, the 1 st bit to the 8 th bit binary numbers from left to right of the 12 bit binary numbers are used as corresponding decimal numbers obtained by calculating the integral part of the reconstructed CSI amplitude data, the 9 th bit to the 12 th bit are used as corresponding decimal numbers obtained by calculating the decimal part of the reconstructed CSI amplitude data, the decimal numbers corresponding to the integral part are used as integers, the decimal numbers corresponding to the decimal part are used as the decimal part combination of the decimal part as the decimal number corresponding to the 12 bit binary numbers, and the decimal numbers are used as the (t-1)/12) +1) -90 (N ' -1) th reconstructed CSI amplitude data of the N ' reconstructed amplitude matrix to be stored in the N ' reconstructed amplitude matrix,

INT is an integer function, and N reconstructed amplitude matrixes respectively storing 30 rows and 3 columns of reconstructed CSI amplitude data are obtained;

step C8: calculating the Euclidean distance between the Nth 'reconstructed amplitude matrix and the Nth' CSI amplitude matrix, accumulating the Euclidean distance between the 1 st reconstructed amplitude matrix and the 1 st CSI amplitude matrix to the Euclidean distance between the Nth reconstructed amplitude matrix and the Nth CSI amplitude matrix, taking the obtained sum as an output error, and recording the output error as phi;

step C9: offset amounts for visible layer, first hidden layer, second hidden layer, third hidden layer and fourth hidden layer and w¹，w²，w³And w⁴Respectively updating, and the specific process is as follows:

c9-1. construction of the k0 th of the visible layerMatrix δ a, δ a ═ of residual terms of the neural units (δ a)₁,δa₂,…,δa_K0)，δa_k0Residual terms for the k0 th neural unit in the visible layer; calculating a residual error term delta a of the k0 th nerve unit in the visible layer by adopting the formula (21)_k0：

δa_k0＝-(v′_k0-v_k0)*v′_k0*(1-v′_k0) (21)

Wherein, v'_k0The current value of the training value, v, of the state quantity of the k0 th neural unit in the visible layer_k0The current value of the state quantity of the k0 th nerve unit in the visible layer;

c9-2. construct a matrix δ b that holds the residual terms of the k1 th neural unit in the first hidden layer (δ b ═ b)₁,δb₂,…,δb_K1)，δb_k1Residual terms for the k1 th neural unit in the first hidden layer; calculating a residual error term delta b of the k1 th nerve unit in the first hidden layer by adopting a formula (22)_k1：

Wherein,

is the current value of the connection weight, deltaa, between the k1 th neural unit in the visible layer and the first hidden layer_k0Is the current value of the residual term for the k0 th neural unit in the visible layer, ha_k1Is the current value of the state quantity of the k1 th neural unit in the first hidden layer;

c9-3. construct a matrix δ C that holds the residual terms of the k2 th neural unit in the second hidden layer (δ C ═ δ C)₁,δc₂,…,δc_K2)，δc_k2Residual terms for the k2 th neural unit in the second hidden layer; calculating a residual error term deltac of the k2 th neural unit in the second hidden layer by adopting a formula (23)_k2：

Wherein,

is the current value of the connection weight between the k1 th neuron in the first hidden layer and the k2 th neuron in the second hidden layer, δ b_k1Is the current value of the residual term of the k1 th neural unit in the first hidden layer, hb_k2Is the current value of the state quantity of the k2 th nerve unit in the second hidden layer;

c9-4. construct a matrix δ d storing the residual terms of the k3 th neural unit in the third hidden layer (δ d ═ d)₁,δd₂,…,δd_K3)，δd_k3Residual terms for the k3 th neural unit in the third hidden layer; calculating a residual error term delta d of the k3 th nerve unit in the third hidden layer by adopting a formula (24)_k3：

Wherein,

is the current value of the connection weight, deltac, between the k2 th neuron in the second hidden layer and the k3 th neuron in the third hidden layer_k2Is the current value, hc, of the residual term of the k2 th neural unit in the second hidden layer_k3The current value of the state quantity of the k3 th neural unit in the third hidden layer;

c9-5. construct a matrix δ e storing the residual terms of the k4 th neural unit in the fourth hidden layer (δ e ═ e₁,δe₂,…,δeK₄)，δe_k4Calculating a residual term delta e of the k4 neural unit in the fourth hidden layer by adopting a formula (25) for the residual term of the k4 neural unit in the fourth hidden layer_k4：

Wherein,

is the current value of the connection weight, δ d, between the k3 th neuron in the third hidden layer and the k4 th neuron in the fourth hidden layer_k3Is the current value of the residual term, hd, for the k3 th neural unit in the third hidden layer_k4Is the current value of the state quantity of the k4 th neural unit in the fourth hidden layer;

c9-6. will

Is added to the current value of (d) by 0.5 x δ d_k3hd_k4Is assigned to

To pair

Updating is carried out; will be provided with

Is added to the current value of 0.5 × δ c_k2hc_k3Is assigned to

To pair

Updating is carried out; will be provided with

Is added by 0.5 × δ b_k1hb_k2Is assigned to

To pair

Is updated to

Is added to the current value of (d) by 0.5 x δ a_k1ha_k1Is assigned to

To pair

Updating is carried out; e is to be_k4Is added with 0.5 × δ e_k4Is assigned to e_k4(ii) a To e_k4Updating is carried out; will d_k3Is added to the current value of (d) by 0.5 x δ d_k3Is assigned to d_k3(ii) a To d_k3Updating is carried out; c is to_k2Is added to the current value of 0.5 × δ c_k2Is assigned to c_k2To c for_k2Updating is carried out; b is to_k1Is added by 0.5 × δ b_k1Is assigned to b_k1To b is paired_k1Updating is carried out; a is to_k0Is added to the current value of (d) by 0.5 x δ a_k0Is assigned to a_k0To a, a_k0Performing an update, wherein δ d_k3、hd_k4、δc_k2、hc_k3、δb_k1、hb_k2、δa_k1、ha_k1And δ e_k4The values of (a) are their current values, respectively;

step C10: judging whether the output error phi is less than 5 and Y is equal to 10000, if at least one of the two conditions is satisfied, determining w¹、w²、w³、w⁴The current values of a, b, c, d and e are taken as S_lIf neither of the two conditions is true, the value of Y is updated by adding 1 to the current value of Y, and the process returns to the step C6 to the step C9 to carry out the next iteration until at least one of the two conditions is true to obtain S_lThe fingerprint of (2);

fifthly, positioning indoor personnel, and the specific process is as follows:

fifthly-1, acquiring N signal data packets of the current position of the indoor person to be positioned in real time through a computer in the monitoring room, and saving the acquired N signal data packets of the current position of the indoor person as data to be detected as a file named CSI.dat;

fifthly-2, obtaining N CSI amplitude matrixes of the current position according to the N signal data packets of the current position by the same method, and calculating to obtain the mean, the variance sigma and the standard deviation std of the N CSI amplitude matrixes of the current position;

-3 setting a coefficient of variation, and recording the coefficient of variation as λ, and calculating the coefficient of variation λ by using a formula (26):

fifthly, determining a 12-bit binary number corresponding to the mth CSI amplitude data of the Nth' th CSI amplitude matrix in the N CSI amplitude matrices at the current position, wherein the specific process is as follows: judging whether the m-th CSI amplitude data of the N 'th CSI amplitude matrix in the N CSI amplitude matrices at the current position is an integer or a decimal, if the m-th CSI amplitude data of the N' th CSI amplitude matrix in the N CSI amplitude matrices at the current position is an integer, converting the m-th CSI amplitude data into a binary number, supplementing 4 0 s to the right of the obtained binary number to obtain a new binary number, then judging whether the number of bits of the obtained new binary number is equal to 12, if the m-th CSI amplitude data is equal to 12, taking the new binary number as a 12-bit binary number corresponding to the m-th CSI amplitude data of the N 'th CSI amplitude matrix in the N CSI amplitude matrices at the current position, if the m-th CSI amplitude data is greater than 12, selecting 12 bits from right to left as 12-bit binary numbers corresponding to the m-th CSI amplitude data of the N' th CSI amplitude matrix in the N CSI amplitude matrices at the current position, and if the m-th CSI amplitude data is less than 12, after 0 is supplemented to the left of the new binary number until the number of bits is 12 bits, the new binary number is used as a 12-bit binary number corresponding to the mth CSI amplitude data of the Nth' CSI amplitude matrix in the N CSI amplitude matrices at the current position; if the m-th CSI amplitude data of the N' th CSI amplitude matrix in the N CSI amplitude matrices at the current position is a decimal, respectively converting an integer part and a decimal part of the M-th CSI amplitude data into binary numbers, if the number of bits of the binary numbers obtained by converting the integer part is less than 8, complementing 0 at the high bits of the binary numbers to change the binary numbers into 8-bit binary numbers, and if the number of bits of the binary numbers obtained by converting the integer part is more than 8, deleting other bits of the 8-bit binary numbers from right to left to obtain 8-bit binary numbersA binary number, wherein if the number of bits of the binary number obtained by converting the integer part is equal to 8, the binary number is kept unchanged, if the number of bits of the binary number obtained by converting the decimal part is equal to 4, the binary number is kept unchanged, if the number of bits of the binary number obtained by converting the decimal part is less than 4, 0 is supplemented to the lower bits of the binary number to change the binary number into a 4-bit binary number, if the number of bits of the binary number obtained by converting the decimal part is greater than 4, 4 bits of the binary number are left from left to right and deleted to obtain a 4-bit binary number, the 8-bit binary number obtained by processing the integer part is used as the upper 8 bits, the 4-bit binary number obtained by processing the decimal part is used as the lower 4 bits to be spliced into a 12-bit binary number, and the spliced 12-bit binary number is used as a 12-bit binary number corresponding to the mth CSI amplitude data of the Nth' CSI amplitude matrix in the N CSI amplitude matrices at the current position; n' ″ 1,2., N; 1,2, ·, 90; assigning the value of the x-th bit counted from left to right to v by the twelve-bit binary value of the m-th amplitude data_{12*3*30*(N″′-1)+(m-1)*12+x}To v is to v₁,v₂,v₃,...,v_K0Updating, wherein x is 1,2, 12;

fifthly, 5 will be at the position S_lThe probability of obtaining v is denoted as Pr (v)_l) Determining Pr (v)_l) The specific process comprises the following steps:

5-1, adding S_lFingerprint w of¹、w²、w³、w⁴And a, b, C, d and e are used as current values, output errors are recalculated according to the methods of the step C6, the step C7 and the step C8, and the output errors calculated at the moment are recorded as phi¹；

5-2, calculating by formula (27) to obtain Pr (v)_l)：

Fifthly, recording the probability of the indoor position as

Let Pr (Lo) be uniformly distributed in the indoor position probability_l) 1/L, locating the indoor person to be positioned at the position S_lIs denoted as Pr (S)_l) Calculating by adopting a formula (28) to obtain Pr (S)_l)：

7 recording the coordinate of the current position of the indoor person to be positioned as (x)_l′,y_l′) Calculating (x) using equations (29) and (30)_l′,y_l′)：

(x) calculated by (8)_l′,y_l′) Namely the positioning coordinates of the indoor person to be positioned.

Compared with the prior art, the invention has the advantages that a deep neural network model with a visible layer and four hidden layers is constructed by utilizing a deep learning technology and is used for extracting fingerprints of each position in a room to be measured to establish a fingerprint database, when the fingerprint of each position is extracted, the deep neural network model is divided into four limited Boltzmann machines, the four limited Boltzmann machines are respectively trained to obtain a weight matrix and a matrix of offset, then the obtained weight matrix and the matrix of offset are subjected to iterative updating optimization by utilizing residual error items until an optimization termination condition is met, so that the weight matrix and the matrix of offset in the deep neural network model with higher precision are obtained to improve the precision, the four weight matrices in the 4 trained limited Boltzmann machines are used, and the matrices of offset of the visible layer and the four hidden layers are stored as the fingerprint of the current position, the degree of the space occupied by the data is far smaller than the degree of the space occupied by the acquired data, the cost is low, so that a database is established for positioning indoor personnel after the fingerprints at all positions are extracted, the fingerprints in the fingerprint database can be directly used when the indoor personnel are positioned, and the positioning process is simple; by using the position S_lThe output error is recalculated, and the output error is calculated based on the recalculated output errorTo obtain at position S_lProbability of obtaining v Pr (v)_l) Then according to the calculated probability Pr (v)_l) Obtaining the position S of the indoor person to be positioned_lThe probability of the positioning coordinate of the indoor person to be positioned is obtained finally, and the result positioning precision is high.

Detailed Description

The present invention will be described in further detail with reference to examples.

Example (b): a passive sensing indoor positioning method based on deep learning comprises the following steps:

firstly, data acquisition:

step 2: dividing an indoor to be positioned into L areas along the transverse direction and the longitudinal direction, wherein the transverse direction is divided into L1 parts along the maximum length direction, the longitudinal direction is divided into L2 parts along the maximum length direction, so that L areas are obtained, L is L1 multiplied by L2, L1 and L2 are integers which are larger than 4 and smaller than 6 respectively, the center of each area is selected as the position of the area, if the area is a non-centrosymmetric area, the intersection point of the transverse sideline and the perpendicular bisector of the longitudinal sideline of the area is taken as the center, and the position of the first area is recorded as S_l1,2, ·, L; selecting any indoor point as an origin, constructing a two-dimensional coordinate system of the indoor area by taking the transverse direction as the x-axis direction and the longitudinal direction as the y-axis direction, and obtaining the first area S through measurement_lCentral point L of_lIs expressed as (x)_l,y_l) A router which is not provided with a password and is communicated with a network is arranged at a certain position in a room to be positioned;

step 3: opening the ubuntu12.04 system in the computer host, connecting the commercial wireless network card intel 5300 arranged on the computer host with the router through a wireless network, and sequentially standing one tester at the position S₁～S_LTo another test personThe personnel sequentially obtain the positions S at the computer host₁～S_LThe specific process of the test data is as follows: one tester standing at position S_lWhen the computer is in operation, another tester collects the position S through the ping router at the speed of 50 times per second of the computer host_lTakes N signal data packets as position S_lTest data of (2), position S_lSaving the test data of (a) as a file with a suffix name of dat, and naming the file as csil, wherein N is an integer greater than 800 and less than 1000;

processing data: installing matlab in the ubuntu12.04 system of the computer host, opening a self-contained matlab folder in a Linux 802.11n CSI Tool by using matlab software, and comparing the position S acquired in the step (i)₁～S_LRespectively processing the test data to obtain the position S₁～S_lThe specific process of the amplitude matrix is as follows: reading csil.dat file by adopting read _ bf _ file function in matlab folder, and acquiring position S_lRespectively opening the position S by adopting a get _ scaled _ csi function_lN signal data packets of (A) to obtain a position S_lN CSI data matrices, position S_lEach CSI data matrix comprises 30 rows and 3 columns of CSI data, the CSI data are in a complex form, and the position S is respectively calculated by adopting an abs function_lObtaining a position S by using the CSI amplitude matrixes corresponding to the N CSI data matrixes_lAt N CSI magnitude matrices, position S_lEach CSI amplitude matrix respectively comprises 30 rows and 3 columns of CSI amplitude data;

constructing a deep neural network model: the neural network model comprises a visible layer, a first hidden layer, a second hidden layer, a third hidden layer and a fourth hidden layer which are sequentially arranged from top to bottom, wherein the visible layer is provided with K0 neural units, and K0 is N30 3 12 which represents a multiplication symbol; the first hidden layer has K1 neural units, 300< K1<500, the second hidden has K2 neural units, 200< K2< K1; the third hidden layer has K3 neural units, 100< K3< K2; the fourth hidden layer has K4 neural units, 50< K4< K3; all nerve units on the same layer are not connected with each other, all nerve units on two adjacent layers are connected with each other, and each nerve unit has two states: the neural network comprises an activated state and a closed state, wherein the state value of the neural unit in the activated state is 1, and the state value of the neural unit in the closed state is 0; the visible layer and the first hidden layer form a first limited Boltzmann machine, the first hidden layer and the second hidden layer form a second limited Boltzmann machine, the second hidden layer and the third hidden layer form a third limited Boltzmann machine, and the third hidden layer and the fourth hidden layer form a fourth limited Boltzmann machine;

step C1, training the first limited Boltzmann machine:

c1-1, constructing a matrix for storing the state quantity of the visible layer, and marking the matrix as v; constructing a matrix for storing the state quantity of the first hidden layer, and recording the matrix as ha; constructing a matrix for storing the offset of the visible layer, and marking the matrix as a; constructing a matrix for storing the offset of the first hidden layer, and recording the matrix as b; constructing a weight matrix storing the connection weight between the first hidden layer and the visible layer, and recording the weight matrix as w¹(ii) a Wherein v ═ v (v)₁,v₂,v₃,...,v_K0)^T，v_k0Represents the state quantity of the K0 th neural unit in the visible layer, the superscript T represents the transpose of the matrix, K0 is 1,2, K0; ha ═ ha (ha)₁,ha₂,ha₃,...,ha_K1)^T，ha_k1Represents the state quantity of the K1 th neural unit in the first hidden layer, K1 being 1,2.., K1; a ═ a₁,a₂,a₃,...,a_K0)^T，a_k0Representing the offset of the k0 th nerve unit in the visible layer, and adopting a random function to a_k0Performing initialization to make a_k0Is a random number between 0 and 0.1; b ═ b₁,b₂,b₃,...,b_K1)^T，b_k1Representing the offset of the k1 th neural unit in the first hidden layer by using a random function pair b_k1Performing initialization to b_k1Is a random number between 0 and 0.1;

Carry out initialization to

Is a random number between 0 and 0.1;

c1-2, defining the sequence of each CSI assigned data in the CSI amplitude matrix: taking CSI amplitude data from 1 st column and 1 st row to 30 th row in a CSI amplitude matrix as 1 st CSI amplitude data to 30 th CSI amplitude data of the matrix, taking 1 st column and 1 st row and 30 th row CSI amplitude data of the matrix as 31 st CSI amplitude data to 60 th CSI amplitude data of the matrix, and taking 3 rd column and 1 st row and 30 th row CSI amplitude data as 61 st CSI amplitude data to 90 th CSI amplitude data of the matrix; determining a location S_lThe 12-bit binary number corresponding to the qth CSI amplitude data of the Nth CSI amplitude matrix is as follows: determination of position S_lIf the q-th CSI amplitude data of the N' -th CSI amplitude matrix is an integer or a decimal if the position S_lThe q-th CSI amplitude data of the N' -th CSI amplitude matrix is an integer, the integer is converted into a binary number, 4 0S are supplemented to the right of the obtained binary number to obtain a new binary number, whether the number of the obtained new binary number is equal to 12 or not is judged, and if the number of the obtained new binary number is equal to 12, the new binary number is used as a position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; if the position is more than 12, selecting 12 bits from right to left as the position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; if less than 12, the left side of the new binary number is complemented with 0 to 12 bits and then used as the positionS_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; if at position S_lIf the q-th CSI amplitude data of the N 'th CSI amplitude matrix is a decimal, converting an integer part and a decimal part of the N' th CSI amplitude matrix into binary numbers respectively, and if the number of bits of the binary numbers obtained by converting the integer part is less than 8, complementing 0 at a high bit of the binary numbers to change the binary numbers into 8 bits; if the number of the binary digits obtained by converting the integer part is more than 8, retaining 8 digits from right to left, and deleting other digits to obtain 8-digit binary digits; if the binary digit obtained by converting the integer part is equal to 8, keeping the binary digit unchanged; if the binary digit number obtained by fractional part conversion is equal to 4, the binary digit number is kept unchanged; if the digit of the binary number obtained by fractional part conversion is less than 4, 0 is complemented at the low bit of the binary number to change the binary number into a 4-bit binary number; if the digit of the binary number obtained by the fractional part conversion is more than 4, the other 4 bits are left from left to right and deleted to obtain a 4-bit binary number, then 8-bit binary number obtained by processing the integer part is used as the high 8 bits, the 4-bit binary number obtained by processing the fractional part is used as the low 4 bits and spliced into a 12-bit binary number, and the spliced 12-bit binary number is used as a position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; n' ═ 1,2, ·, N; q ═ 1,2, ·, 90; assigning the value of the p-th bit from left to right of the 12-bit binary number corresponding to the q-th CSI amplitude data to v_{12*3*30*(N′-1)+(q-1)*12+p}P1, 2, 12, thereby completing v₁,v₂,v₃,...,v_K0Initial assignment of (1);

c1-3. construct a first training matrix ha ', ha' (ha ') that holds the first training values of the state quantities of the k1 th neural unit in the first hidden layer'₁,ha′₂,…,ha′_K1)^T，ha′_k1A first training value representing a state quantity of a k1 th neural unit in the first hidden layer; constructing a matrix pa storing the probability of first activation of the k1 th neural unit in the first hidden layer, pa ═ pa₁,pa₂,pa₃,...,pa_K1)^T，pa_k1Indicating the k1 th god in the first hidden layerA first activation probability by cell; determining a first training matrix ha' of a first state training value of a k1 th neural unit in the first hidden layer and a matrix pa of a first activation probability of a k1 th neural unit in the first hidden layer, wherein the specific process is as follows:

is the current value, v ', of the connection weight between the k1 th neural unit in the first hidden layer and the k0 th neural unit in the visible layer'_k0Training values which are state quantities of k0 th neural units in the visible layer;

C1-6. will

Added 0.01 x (pa) to the current value of_k1*vk₀-pa′_k1*v′_k0) Is assigned to

To pair

The value of the weight matrix w is updated to obtain an updated weight matrix w¹(ii) a A is to_k0Current value of plus 0.01 (vk)₀-v′_k0) Is assigned to a_k0To a, a_k0Updating the value of the visible layer to obtain the updated offset a of the visible layer; b is to_k1Added 0.01 x (pa) to the current value of_k1-pa′_k1) Is assigned to b_k1To b is paired_k1Is updated to obtain the updated offset b of the first hidden layer, wherein pa_k1、vk₀、pa′_k1And v'_k0All the values of (a) are the current values thereof;

step C2: training a second limited Boltzmann machine:

c2-1, constructing a matrix for storing the state quantity of the second hidden layer, and marking the matrix as hb; build up and store a second hidden layerThe matrix of the offset of (1), which is denoted as c; constructing a weight matrix for storing a connection weight between the first hidden layer and the visible layer, and recording the weight matrix as w²；hb＝(hb₁,hb₂,hb₃,...,hb_K2)^T，hb_k2A state value representing the K2 th neural unit in the second hidden layer, K2 ═ 1,2.., K2; c ═ c₁,c₂,c₃,...,c_K2)^T，c_k2Representing the offset of the k2 th neural unit in the second hidden layer;

representing the connection weight between the k1 th nerve unit in the first hidden layer and the k2 th nerve unit in the second hidden layer, and respectively adopting random functions to pair c_k2And

is initialized to c_k2And

respectively, random values of 0-0.1;

c2-2. construct a fourth training matrix hb ', hb ' ═ hb '₁,hb′₂,…,hb′_K2)^T，hb′_k2A first training value representing a state quantity of a k2 th neural unit in the second hidden layer; constructing a matrix pb storing first activation probabilities for the k2 th neural unit in the second hidden layer (pb ═ pb)₁,pb₂,pb₃,...,pb_K2)^T，pb_k2Representing the probability of the first activation of the k2 th neural unit in the second hidden layer; determining a fourth training matrix hb' of the first training values of the state quantities of the k2 th neural unit in the second hidden layer and a matrix pb of the first activation probabilities of the k2 th neural unit in the second hidden layer, in particularThe process is as follows:

c2-3. constructing a fifth training matrix ha '", ha'" ═ ha '"of (ha'" ', of the third training values of the state quantities of the k1 th neural unit in the first hidden layer'₁,ha″′₂,…,ha″′_K1)^T，ha″′_k1A third training value representing a state quantity of a k1 th neural unit in the first hidden layer; determining a fifth training matrix ha' ″ storing a third training value of the state quantity of the k1 th neural unit in the first hidden layer by the following specific process:

C2-4, constructing a sixth training matrix hb ", hb" ═ hb ″ (hb ″) storing second training values of state quantities of the k2 th neural unit in the second hidden layer₁,hb″₂,…,hb″_K2)^T，hb″_k2A second training value representing a state quantity of a k2 th neural unit in the second hidden layer; constructing a matrix pb ', pb ' ═ b (pb '₁,pb′₂,…,pb′_K2)^T，pb′_k2The method comprises the following steps of representing the second activation probability of the k2 th neural unit in the second hidden layer, determining a sixth training matrix hb 'of a second training value for storing the state quantity of the k2 th neural unit in the second hidden layer and a matrix pb' of the second activation probability of the k2 th neural unit in the second hidden layer, wherein the specific process is as follows:

is the k1 th god in the first hidden layerVia the current value of the connection weight between the cell and the k2 th neural cell in the second hidden layer, ha'_k1A third training value of state quantities of a k1 th neural unit in the first hidden layer;

C2-5. will

Current value of plus 0.01 (pb)_k2*ha_k1-pb′_k2*ha″′_k1) Is assigned to

To pair

step C3: training a third restricted boltzmann machine:

c3-1, constructing a matrix for storing the state quantity of the third hidden layer, and marking the matrix as hc; constructing a matrix for storing the offset of the third hidden layer, and recording the matrix as d; building a connection weight between the second hidden layer and the third hidden layerThe weight matrix of (2) is recorded as w³；hc＝(hc₁,hc₂,hc₃,...,hc_K3)^T，hc_k3A state value representing the K3 th neural unit in the third hidden layer, K3 ═ 1,2.., K3; d ═ d (d)₁,d₂,d₃,...,dK₃)^T，d_k3Representing the offset of the k3 th neural unit in the third hidden layer;

representing the connection weight between the k2 th nerve unit in the second hidden layer and the k3 th nerve unit in the third hidden layer, and respectively adopting random functions to pair d_k3And

initializing to d_k3And

respectively, random values of 0-0.1;

c3-2. construct a seventh training matrix hc ', hc' ═ hc 'of the first training values of the state quantities of the k3 th neural unit in the third hidden layer'₁,hc′₂,…,hc′_K3)^T，hc′_k3A first training value representing a state quantity of a k3 th neural unit in the third hidden layer; constructing a matrix pc storing the probability of first activation of the k3 th neural unit in the third hidden layer (pc ═ c)₁,pc₂,pc₃,...,pcK₃)^T，pc_k3Represents the first activation probability of the k3 th neural unit in the third hidden layer; determining a seventh training matrix hc' storing the first training value of the state quantity of the k3 th neural unit in the third hidden layer and a matrix pc storing the first activation probability of the k3 th neural unit in the third hidden layer, wherein the specific process is as follows:

c3-2-1. excitation of the k3 th neural unit in the third hidden layerThe activity probability is recorded as Pr (k3), and Pr (k3) is calculated by adopting formula (7):

c3-2-2, assigning the current value of Pr (k3) to pc_k3Obtaining an updated matrix pc, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k3) with the random number, and if the current value of Pr (k3) is greater than the random number, assigning 1 to hc'_k3(ii) a If the current value of Pr (k3) is not greater than the random number, then 0 is assigned to hc'_k3；

c3-5. will

To pair

step C4: training a fourth restricted boltzmann machine:

respectively initializing to be random values of 0-0.1;

c4-2-1, recording the activation probability of the k4 nerve unit in the fourth hidden layer as Pr (k4), and calculating by adopting a formula (10)

Pr(k4)：

C4-3. constructing an eleventh training matrix hc "storing third training values for the state quantities of the k3 th neural unit in the third hidden layer^′， hc″′＝(hc″′₁,hc″′₂,…,hc″′_K3)^T，hc″′_k3And determining an eleventh training matrix hc' for storing the third training value of the state quantity of the k3 th neural unit in the third hidden layer by using the third training value representing the state quantity of the k3 th neural unit in the third hidden layer, wherein the specific process is as follows:

is the current value of the connection weight, hd ', between the k3 th neural unit in the third hidden layer and the k4 th neural unit in the fourth hidden layer'_k4For the k4 th nerve sheet in the fourth hidden layerA first training value of a state quantity of the element;

c4-4-2 assigning the current value of Pr (k4) to pd'_k4Generating a 0E to E by random function1, comparing the current value of Pr (k4) with the random number, and if the current value of Pr (k4) is greater than the random number, assigning 1 to hd ″, respectively_k4And hc and_k4if the current value of Pr (k4) is not greater than the random number, then 0 is assigned to hd ″', respectively_k4And hc and_k4；

c4-5. will

Current value of plus 0.01 x (pd)_k4*hc_k3-pd′_k4*hc″′_k3) Is assigned to

To pair

Updating to obtain updated weight matrix w⁴(ii) a Will d_k3Current value of (d) plus 0.01 (hc)_k3-hc″′_k3) Is assigned to d_k3To d is paired_k3Updating to obtain the updated offset d of the third hidden layer; e is to be₄₃Current value of plus 0.01 x (pd)_k4-pd′_k4) To e_k4To e is aligned with_k4Updating to obtain the offset e of the updated fourth hidden layer; wherein pd is_k4、hc_k3、pd′_k4And hc'_k3The values of (a) are their current values, respectively;

c6-4, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k2) with the random number, and if the current value of Pr (k2) is larger than the random number, assigning 1 to hb_k2To hb_k2Updating is carried out; if the current value of Pr (k2) is not greater than the random number, then a value of 0 is assigned to hb_k2To hb_k2Updating is carried out;

c6-6, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k3) with the random number, and if the current value of Pr (k3) is larger than the random number, assigning 1 to hc_k3To hc_k3Updating is carried out; if the current value of Pr (k3) is not greater than the random number, then a value of 0 is assigned to hc_k3To hc_k3Updating is carried out;

c6-8, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k4) with the random number, and if the current value of Pr (k4) is larger than the random number, assigning 1 to the random numberValue given to hd_k4To hd_k4Updating is carried out; if the current value of Pr (k4) is not greater than the random number, then a value of 0 is assigned to hd_k4To hd_k4Updating is carried out;

as the current value of the connection weight, hd, between the k3 th neuron in the third hidden layer and the k4 th neuron in the fourth hidden layer_k4Is the current value of the state quantity of the k4 th neural unit in the fourth hidden layer;

is the current value of the connection weight, hc, between the k2 th neuron in the second hidden layer and the k3 th neuron in the third hidden layer_k3Is as followsThe current value of the state quantity of the k3 th neural unit in the three hidden layers;

step C7: creating N reconstructed amplitude matrixes for storing the reconstructed CSI amplitude data, wherein each reconstructed amplitude matrix can store 30 rows and 3 columns of reconstructed CSI amplitude data, the sequencing definition of the 30 rows and 3 columns of reconstructed CSI amplitude data stored in each reconstructed amplitude matrix is the same as that of the CS amplitude matrix, and v't-v'_11+tAccording to v'_tv′_t+1v′_t+2v′_t+3v′_t+4…v′_t+ ₉v′_t+10v′_t+11The order combination of the 12 bit binary numbers is 12 bit binary numbers, t is 12N +1, N is 0, 1,2, …, K0/12-1, the 1 st bit to the 8 th bit binary numbers from left to right of the 12 bit binary numbers are used as corresponding decimal numbers obtained by calculating the integral part of the reconstructed CSI amplitude data, the 9 th bit to the 12 th bit are used as corresponding decimal numbers obtained by calculating the decimal part of the reconstructed CSI amplitude data, the decimal numbers corresponding to the integral part are used as integers, the decimal numbers corresponding to the decimal part are used as the decimal part combination of the decimal part as the decimal number corresponding to the 12 bit binary numbers, and the decimal numbers are used as the (t-1)/12) +1) -90 (N ' -1) th reconstructed CSI amplitude data of the N ' reconstructed amplitude matrix to be stored in the N ' reconstructed amplitude matrix,

c9-1. construct a matrix δ a storing the residual terms of the k0 th neural unit in the visible layer (δ a ═ d₁,δa₂,…,δaK₀)，δa_k0Residual terms for the k0 th neural unit in the visible layer; calculating a residual error term delta a of the k0 th nerve unit in the visible layer by adopting the formula (21)_k0：

δa_k0＝-(v′_k0-v_k0)*v′_k0*(1-v′_k0) (21)

Wherein,

Wherein,

c9-4. construct a matrix δ d storing the residual terms of the k3 th neural unit in the third hidden layer (δ d ═ d)₁,δd₂,…,δdK₃)，δd_k3Residual terms for the k3 th neural unit in the third hidden layer; calculating a residual error term delta d of the k3 th nerve unit in the third hidden layer by adopting a formula (24)_k3：

Wherein,

is the current value of the connection weight, deltac, between the k2 th neuron in the second hidden layer and the k3 th neuron in the third hidden layer_k2Is a second hidden layerCurrent value of residual term, hc, of the k2 th neural unit in (A)_k3The current value of the state quantity of the k3 th neural unit in the third hidden layer;

c9-5. construct a matrix δ e storing the residual terms of the k4 th neural unit in the fourth hidden layer (δ e ═ e₁,δe₂,…,δe_K4)，δe_k4Calculating a residual term delta e of the k4 neural unit in the fourth hidden layer by adopting a formula (25) for the residual term of the k4 neural unit in the fourth hidden layer_k4：

Wherein,

c9-6. will

Is added to the current value of (d) by 0.5 x δ d_k3hd_k4Is assigned to

To pair

Updating is carried out; will be provided with

Is added to the current value of 0.5 × δ c_k2hc_k3Is assigned to

To pair

Updating is carried out; will be provided with

Is added by 0.5 × δ b_k1hb_k2Is assigned to

To pair

Is updated to

Is added to the current value of (d) by 0.5 x δ a_k1ha_k1Is assigned to

To pair

step C10: judging whether the output error phi is less than 5 and Y is equal to 10000, if at least one of the two conditions is satisfied, determining w¹、w²、w³、w⁴The current values of a, b, c, d and e are taken as S_lThe fingerprint of (a) the fingerprint of (b),if neither condition is satisfied, updating the value of Y by adding 1 to the current value of Y, returning to step C6-step C9 for the next iteration until at least one of the two conditions is satisfied to obtain S_lThe fingerprint of (2);

fifthly, positioning indoor personnel, and the specific process is as follows:

fifthly, determining a 12-bit binary number corresponding to the mth CSI amplitude data of the Nth' th CSI amplitude matrix in the N CSI amplitude matrices at the current position, wherein the specific process is as follows: judging whether the m-th CSI amplitude data of the N 'th CSI amplitude matrix in the N CSI amplitude matrices at the current position is an integer or a decimal, if the m-th CSI amplitude data of the N' th CSI amplitude matrix in the N CSI amplitude matrices at the current position is an integer, converting the m-th CSI amplitude data into a binary number, supplementing 4 0 s to the right of the obtained binary number to obtain a new binary number, then judging whether the number of bits of the obtained new binary number is equal to 12, if the m-th CSI amplitude data is equal to 12, taking the new binary number as a 12-bit binary number corresponding to the m-th CSI amplitude data of the N 'th CSI amplitude matrix in the N CSI amplitude matrices at the current position, if the m-th CSI amplitude data is greater than 12, selecting 12 bits from right to left as 12-bit binary numbers corresponding to the m-th CSI amplitude data of the N' th CSI amplitude matrix in the N CSI amplitude matrices at the current position, and if the m-th CSI amplitude data is less than 12, thenAfter 0 is supplemented to the left of the new binary number to the digit of 12 bits, the new binary number is used as a 12-bit binary number corresponding to the mth CSI amplitude data of the Nth' CSI amplitude matrix in the N CSI amplitude matrices at the current position; if the mth CSI amplitude data of the Nth' th CSI amplitude matrix in the N CSI amplitude matrices at the current position is a decimal, respectively converting an integer part and a decimal part of the current position into binary numbers, if the number of bits of the binary numbers obtained by converting the integer part is less than 8, complementing 0 at the high order of the binary numbers into 8-bit binary numbers, if the number of bits of the binary numbers obtained by converting the integer part is more than 8, deleting other bits of the 8-bit binary numbers from the right to the left to obtain 8-bit binary numbers, if the number of bits of the binary numbers obtained by converting the integer part is equal to 8, keeping unchanged, if the number of bits of the binary numbers obtained by converting the decimal part is equal to 4, complementing 0 at the low order of the binary numbers obtained by converting the decimal part into 4-bit binary numbers, if the number of bits obtained by converting the decimal part is more than 4, then, 4 bits of other binary numbers are left from left to right, 4 bits of other binary numbers are obtained by deletion, the 8-bit binary number obtained after the integer part is processed is used as a high 8 bit, the 4-bit binary number obtained after the decimal part is processed is used as a low 4 bit, the low 4 bit is spliced into a 12-bit binary number, and the 12-bit binary number obtained by splicing is used as a 12-bit binary number corresponding to the mth CSI amplitude data of the Nth' CSI amplitude matrix in the N CSI amplitude matrices at the current position; n' ″ 1,2., N; 1,2, ·, 90; assigning the value of the x-th bit counted from left to right to v by the twelve-bit binary value of the m-th amplitude data_{12*3*30*(N″′-1)+(m-1)*12+x}To v is to v₁,v₂,v₃,...,v_K0Updating, wherein x is 1,2, 12;

5-1, adding S_lFingerprint w of¹、w²、w³、w⁴And a, b, C, d and e are used as current values, the output errors are recalculated according to the methods of the step C6, the step C7 and the step C8, and the output errors calculated at the moment are recorded as the current valuesφ¹；

5-2, calculating by formula (27) to obtain Pr (v)_l)：

Fifthly, recording the probability of the indoor position as

Claims

1. A passive sensing indoor positioning method based on deep learning is characterized by comprising the following steps:

firstly, data acquisition:

step 1: preparing a computer, inserting the commercial wireless network card intel 5300 into a wireless network card expansion slot of a computer host, installing an ubuntul2.04 system and a Linux 802.11n CSI Tool in the computer host, and placing the computer in a monitoring room;

step 2: dividing the room to be positioned into L areas along the transverse direction and the longitudinal direction, wherein the transverse direction is divided into L1 parts along the maximum length direction of the transverse direction, and the longitudinal direction is divided into the maximum length direction of the longitudinal directionDividing the length direction of the region into L2 parts evenly, thereby obtaining L regions, wherein L is L1 xL 2, L1 and L2 are integers which are more than 4 and less than 6 respectively, the center of each region is selected as the position of the region, if the region is a non-centrosymmetric region, the intersection point of the perpendicular bisector of the transverse sideline and the longitudinal sideline of the region is taken as the center, and the position of the first region is recorded as S_l1,2, ·, L; selecting any indoor point as an origin, constructing a two-dimensional coordinate system of the indoor area by taking the transverse direction as the x-axis direction and the longitudinal direction as the y-axis direction, and obtaining the first area S through measurement_lCentral point L of_lIs expressed as (x)_l，y_l) A router which is not provided with a password and is communicated with a network is arranged at a certain position in a room to be positioned;

step 3: opening the ubuntul2.04 system in the computer host, connecting the commercial wireless network card intel 5300 arranged on the computer host with the router through a wireless network, and sequentially standing a tester at the position S₁～S_LThe other tester obtains the position S in turn at the computer host₁～S_LThe specific process of the test data is as follows: one tester standing at position S_lWhen the computer is in operation, another tester collects the position S through the ping router at the speed of 50 times per second of the computer host_lTakes N signal data packets as position S_lTest data of (2), position S_lSaving the test data of (a) as a file with a suffix name of dat, and naming the file as csil, wherein N is an integer greater than 800 and less than 1000;

processing data: installing matlab in the ubuntu12.04 system of the computer host, opening a self-contained matlab folder in a Linux 802.11n CSI Tool by using matlab software, and comparing the position S acquired in the step (i)₁～S_LRespectively processing the test data to obtain the position S₁～S_LThe specific process of the amplitude matrix is as follows: reading csil.dat file by adopting read _ bf _ file function in matlab folder, and acquiring position S_lRespectively opening the position S by adopting a get _ scaled _ csi function_lN signal data packets ofObtaining a position S_lN CSI data matrices, position S_lEach CSI data matrix comprises 30 rows and 3 columns of CSI data, the CSI data are in a complex form, and the position S is respectively calculated by adopting an abs function_lObtaining a position S by using the CSI amplitude matrixes corresponding to the N CSI data matrixes_lAt N CSI magnitude matrices, position S_lEach CSI amplitude matrix respectively comprises 30 rows and 3 columns of CSI amplitude data;

constructing a deep neural network model: the neural network model comprises a visible layer, a first hidden layer, a second hidden layer, a third hidden layer and a fourth hidden layer which are sequentially arranged from top to bottom, wherein the visible layer is provided with K0 neural units, K0 is N30 3 12, and represents a multiplication symbol; the first hidden layer has K1 nerve cells, 300< K1<500, the second hidden layer has K2 nerve cells, 200< K2< K1; the third hidden layer is provided with K3 nerve units, 100< K3< K2; the fourth hidden layer is provided with K4 nerve units, 50< K4< K3; all the nerve units in the same layer are not connected with each other, all the nerve units in the two adjacent layers are connected with each other, and each nerve unit has two states: the neural network comprises an activated state and a closed state, wherein the state value of the neural unit in the activated state is 1, and the state value of the neural unit in the closed state is 0; the visible layer and the first hidden layer form a first limited Boltzmann machine, the first hidden layer and the second hidden layer form a second limited Boltzmann machine, the second hidden layer and the third hidden layer form a third limited Boltzmann machine, and the third hidden layer and the fourth hidden layer form a fourth limited Boltzmann machine;

step C1: training the first limited Boltzmann machine:

c1-1, constructing a matrix for storing the state quantity of the visible layer, and marking the matrix as v; constructing a matrix for storing the state quantity of the first hidden layer, and recording the matrix as ha; constructing a matrix for storing the offset of the visible layer, and marking the matrix as a; construction ofStoring a matrix of the offset of the first hidden layer, and marking the matrix as b; constructing a weight matrix storing the connection weight between the first hidden layer and the visible layer, and recording the weight matrix as w¹(ii) a Wherein v ═ v (v)₁，v₂，v₃，...，v_K0)^T，v_k0Represents the state quantity of the K0 th nerve unit in the visible layer, the superscript T represents the transpose of the matrix, K0 is 1,2 …, K0; ha ═ ha (ha)₁，ha₂，ha₃，...，ha_K1)^T，ha_k1Represents the state quantity of the K1 th nerve unit in the first hidden layer, wherein K1 is 1,2, … and K1; a ═ a₁，a₂，a₃，...，a_K0)^T，a_k0Representing the offset of the k0 th nerve unit in the visible layer, and adopting a random function to a_k0Performing initialization to make a_k0Is a random number between 0 and 0.1; b ═ b₁，b₂，b₃，...，b_K1)^T，b_k1Representing the offset of the k1 th neural unit in the first hidden layer by using a random function pair b_k1Performing initialization to b_k1Is a random number between 0 and 0.1;

Carry out initialization to

Is a random number between 0 and 0.1;

c1-2, defining the sequence of each CSI assigned data in the CSI amplitude matrix: sending the CSI amplitude data positioned in the 1 st column and the 1 st row in the CSI amplitude matrix to the 1 stThe CSI amplitude data of 30 rows are used as the 1 st CSI amplitude data to the 30 th CSI amplitude data of the matrix, the 1 st row CSI amplitude data to the 30 th row CSI amplitude data of the 2 nd column are used as the 31 st CSI amplitude data to the 60 th CSI amplitude data of the matrix, and the 1 st row CSI amplitude data to the 30 th row CSI amplitude data of the 3 rd column are used as the 61 st CSI amplitude data to the 90 th CSI amplitude data of the matrix; determining a location S_lThe 12-bit binary number corresponding to the qth CSI amplitude data of the Nth CSI amplitude matrix is as follows: determination of position S_lIf the q-th CSI amplitude data of the N' -th CSI amplitude matrix is an integer or a decimal if the position S_lThe q-th CSI amplitude data of the N' -th CSI amplitude matrix is an integer, the integer is converted into a binary number, 4 0S are supplemented to the right of the obtained binary number to obtain a new binary number, whether the number of the obtained new binary number is equal to 12 or not is judged, and if the number of the obtained new binary number is equal to 12, the new binary number is used as a position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; if the position is more than 12, selecting 12 bits from right to left as the position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; if the number of bits is less than 12, the left side of the new binary number is supplemented with 0 to 12 bits and then the result is used as the position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; if at position S_lIf the q-th CSI amplitude data of the N 'th CSI amplitude matrix is a decimal, converting an integer part and a decimal part of the N' th CSI amplitude matrix into binary numbers respectively, and if the number of bits of the binary numbers obtained by converting the integer part is less than 8, complementing 0 at a high bit of the binary numbers to change the binary numbers into 8 bits; if the number of the binary digits obtained by converting the integer part is more than 8, retaining 8 digits from right to left, and deleting other digits to obtain 8-digit binary digits; if the binary digit obtained by converting the integer part is equal to 8, keeping the binary digit unchanged; if the binary digit number obtained by fractional part conversion is equal to 4, the binary digit number is kept unchanged; if the digit of the binary number obtained by fractional part conversion is less than 4, 0 is complemented at the low bit of the binary number to change the binary number into a 4-bit binary number; if the number of binary digits obtained by fractional part conversion is more than 4, then going from left to rightDeleting other 4 bits to obtain a 4-bit binary number, splicing the 8-bit binary number obtained after processing the integer part as a high 8 bit and the 4-bit binary number obtained after processing the decimal part as a low 4 bit to obtain a 12-bit binary number, and splicing the 12-bit binary number obtained as a position S_lA 12-bit binary number corresponding to the q-th CSI amplitude data of the N' -th CSI amplitude matrix; n ═ 1,2, …, N; q is 1,2, …, 90; assigning the value of the p-th bit from left to right of the 12-bit binary number corresponding to the q-th CSI amplitude data to v_{12*3*30*(N′-1)+(q-1)*12+p}P is 1,2, …, 12, thereby completing v₁，v₂，v₃，...，v_K0Initial assignment of (1);

c1-3. construct a first training matrix ha ', ha' (ha ') that holds the first training values of the state quantities of the k1 th neural unit in the first hidden layer'₁，ha′₂，…，ha′_K1)T，ha′_k1A first training value representing a state quantity of a k1 th neural unit in the first hidden layer; constructing a matrix pa storing the probability of first activation of the k1 th neural unit in the first hidden layer, pa ═ pa₁，pa₂，pa₃，...，pa_K1)^T，pa_k1Representing the probability of the first activation of the k1 th neural unit in the first hidden layer; determining a first training matrix ha' of a first state training value of a k1 th neural unit in the first hidden layer and a matrix pa of a first activation probability of a k1 th neural unit in the first hidden layer, wherein the specific process is as follows:

c1-3-2 assigning the current value of Pr (k1) to Pa_k1Generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k1) with the random number, and if the current value of Pr (k1) is greater than the random number, assigning 1 to ha'_k1Assigning 0 to ha 'if the current value of Pr (k1) is not greater than the random number'_k1；

C1-4. construct a second training matrix v ', v ═ v'₁，v′₂，…，v′_K0)^T，v′_k0Training values which are state quantities of k0 th neural units in the visible layer; determining a training value v 'of a state quantity of a k0 th neural unit in a visible layer'_k0The specific process is as follows:

c1-4-2, generating a random number between 0 and 1 by using a random function, comparing the current value of Pr (k0) with the random number, and if the current value of Pr (k0) is greater than the random number, assigning 1 to the random numberValue to v'_k0Assigning 0 to v 'if the current value of Pr (k1) is not greater than the random number'_k0；

C1-5, constructing a third training matrix ha', ha ″ (ha ″ ") storing second training values of state quantities of the k1 th neural unit in the first hidden layer₁，ha″₂，…，ha″_k1)T，ha″_k1A second training value representing a state quantity of a k1 th neural unit in the first hidden layer; constructing a matrix pa ', pa ═ of (pa'₁，pa′₂，…，pa′_K1)^T，pa′_k1Representing a second probability of activation of the k1 th neural unit in the first hidden layer; determining a third training matrix ha ″ of the second training value of the state quantity of the k1 th neural unit in the first hidden layer and a matrix pa' of the second activation probability of the k1 th neural unit in the first hidden layer by the following specific processes:

C1-6. will

To pair

step C2: training the second limited Boltzmann machine:

c2-1, constructing a matrix for storing the state quantity of the second hidden layer, and marking the matrix as hb; constructing a matrix for storing the offset of the second hidden layer, and recording the matrix as c; constructing a weight matrix for storing a connection weight between the first hidden layer and the visible layer, and recording the weight matrix as w²；hb＝(hb₁，hb₂，hb₃，...，hb_K2)^T，hb_k2Represents the state value of the K2 th nerve unit in the second hidden layer, K2 is 1,2, …, K2; c ═ c₁，c₂，c₃，..，c_K2)^T，c_k2Representing the offset of the k2 th neural unit in the second hidden layer;

is initialized to c_k2And

respectively, random values of 0-0.1;

c2-2. construct a fourth training matrix hb ', hb ' ═ hb '₁，hb′₂，…，hb′_K2)^T，hb′_k2A first training value representing a state quantity of a k2 th neural unit in the second hidden layer; constructing a matrix pb storing first activation probabilities for the k2 th neural unit in the second hidden layer (pb ═ pb)₁，pb₂，pb₃，...，pb_K2)^T，pb_k2Representing the probability of the first activation of the k2 th neural unit in the second hidden layer; determining a fourth training matrix hb' of the first training value of the state quantity of the k2 th neural unit in the second hidden layer and a matrix pb of the first activation probability of the k2 th neural unit in the second hidden layer, specifically comprising the following steps:

wherein，c_k2The current value of the offset for the k2 th neural unit in the second hidden layer,

c2-3. constructing a fifth training matrix ha '", ha'" ═ ha '"of (ha'" ', of the third training values of the state quantities of the k1 th neural unit in the first hidden layer'₁，ha″′₂，…，ha″′_K1)T，ha″′_k1A third training value representing a state quantity of a k1 th neural unit in the first hidden layer; determining a fifth training matrix ha' ″ storing a third training value of the state quantity of the k1 th neural unit in the first hidden layer by the following specific process:

for the current value of the connection weight between the k1 th neuron in the first hidden layer and the k2 th neuron in the second hidden layer,hb′_k2a first training value of state quantity of a k2 th neural unit in a second hidden layer;

C2-4, constructing a sixth training matrix hb', hb ″ (hb ″ ") storing second training values of state quantities of the k2 th neural unit in the second hidden layer₁，hb″₂，…，hb″_K2)T，hb″_k2A second training value representing a state quantity of a k2 th neural unit in the second hidden layer; constructing a matrix pb ', pb ' ═ b (pb '₁，pb′₂，…，pb′_K2)^T，pb′_k2The method comprises the following steps of representing the second activation probability of the k2 th neural unit in the second hidden layer, determining a sixth training matrix hb 'of a second training value for storing the state quantity of the k2 th neural unit in the second hidden layer and a matrix pb' of the second activation probability of the k2 th neural unit in the second hidden layer, wherein the specific process is as follows:

is the current value of the connection weight, ha ″, between the k1 th neural unit in the first hidden layer and the k2 th neural unit in the second hidden layer'_k1A third state quantity of the k1 th neural unit in the first hidden layerA secondary training value;

C2-5. will

Current value of plus 0.01 (pb)_k2*ha_k1-pb′_k2*ha″′_k1) Is assigned to

To pair

step C3: training the third limited Boltzmann machine:

c3-1, constructing a matrix for storing the state quantity of the third hidden layer, and marking the matrix as hc; constructing a matrix for storing the offset of the third hidden layer, and recording the matrix as d; constructing a weight matrix for storing the connection weight between the second hidden layer and the third hidden layer, and recording the weight matrix as w³；hc＝(hc₁，hc₂，hc₃，...，hc_K3)^T，hc_k3Represents the state value of the K3 th neural unit in the third hidden layer, K3 is 1,2, …, K3; d ═ d (d)₁，d₂，d₃，...，d_K3)^T，d_k3Representing the offset of the k3 th neural unit in the third hidden layer;

initializing to d_k3And

respectively, random values of 0-0.1;

c3-2. construct a seventh training matrix hc ', hc' ═ hc 'of the first training values of the state quantities of the k3 th neural unit in the third hidden layer'₁，hc′₂，…，hc′_K3)^T，hc′_k3A first training value representing a state quantity of a k3 th neural unit in the third hidden layer; constructing a matrix pc storing the probability of first activation of the k3 th neural unit in the third hidden layer (pc ═ c)₁，pc₂，pc₃，...，pc_K3)^T，Pc_k3Represents the first activation probability of the k3 th neural unit in the third hidden layer; determining a seventh training matrix hc' storing the first training value of the state quantity of the k3 th neural unit in the third hidden layer and a matrix pc storing the first activation probability of the k3 th neural unit in the third hidden layer, wherein the specific process is as follows:

C3-3, constructing an eighth training matrix hb '″, hb' ″ ═ h b '″, h b ″, which deposits a third training value of the state quantity of the k2 th neural unit in the second hidden layer'₁，hb〞′₂，…，hb″′_k2)^T，hb″′_k2And determining an eighth training matrix hb' ″ for storing the third training value of the state quantity of the k2 th neural unit in the second hidden layer by using the third training value representing the state quantity of the k2 th neural unit in the second hidden layer, which comprises the following specific processes:

C3-4, constructing a ninth training matrix hc ", hc ═ hc ″ (hc ″") that stores the second training values of the state quantities of the k3 th neural unit in the third hidden layer₁，hc″₂，…，hc″_K3)^T，hc″_k3A second training value representing a state quantity of a k3 th neural unit in the third hidden layer; constructing a matrix pc ', pc ═ c (pc'₁，pc′₂，…，pc′_K3)^T，pc′_k3Represents the second activation probability of the k3 th neural unit in the third hidden layer; the ninth training matrix hc 'for storing the second training value of the state quantity of the k3 th neural unit in the third hidden layer and the matrix pc' for the second activation probability of the k3 th neural unit in the third hidden layer are determined by the following specific processes:

c3-5. will

To pair

step C4: training the fourth limited Boltzmann machine:

c4-1, constructing a matrix for storing the state quantity of the fourth hidden layer, and applying the matrixIs marked as hd; constructing a matrix for storing the offset of the fourth hidden layer, and recording the matrix as e; constructing a weight matrix storing the connection weight between the third hidden layer and the fourth hidden layer, and recording the weight matrix as w⁴；hd＝(hd₁，hd₂，hd₃，...，hd_K4)^T，hd_k4A state value representing the K4 th neural unit in the fourth hidden layer, K4 ═ 1,2.., K4; e ═ e (e)₁，e₂，e₃，...，e_K4)^T，e_k4Represents the offset of the k4 th neural unit of the fourth hidden layer,

respectively initializing to be random values of 0-0.1;

c4-2. construct a tenth training matrix hd 'that stores the first training values of the state quantities of the k4 th neural unit in the fourth hidden layer, hd ═ hd'₁，hd′₂，…，hd′_k4)^T，hd′_k4A first training value representing a state quantity of a k4 th neural unit in the fourth hidden layer; constructing a matrix pd storing the probability of the first activation of the k4 th neural unit in the fourth hidden layer (pd ═ pd)₁，pd₂，pd₃，...，pd_K4)^T，pd_k4Represents the probability of the first activation of the k4 th neural unit in the fourth hidden layer; the specific process of determining the tenth training matrix hd' storing the first training value of the state quantity of the k4 th neural unit in the fourth hidden layer and the matrix pd storing the first activation probability of the k4 th neural unit in the fourth hidden layer is as follows:

C4-3 constructing an eleventh training matrix hc '", hc'" ═ h ″ ', of third training values of state quantities of k3 th neural units in a third hidden layer'₁，hc″′₂，…，hc″′_k3)^T，hc″′_k3And determining an eleventh training matrix hc' for storing the third training value of the state quantity of the k3 th neural unit in the third hidden layer by using the third training value representing the state quantity of the k3 th neural unit in the third hidden layer, wherein the specific process is as follows:

C4-4, constructing a twelfth training matrix hd ″ storing second training values of state quantities of the k4 th neural unit in the fourth hidden layer, hd ″ (hd ″)₁，hd″₂，…，hd″_K4)^T，hd″_k4A second training value representing a state quantity of a k4 th neural unit in the fourth hidden layer; constructing a matrix pd ', pd' ═ of (pd ') of the second activation probability of the k4 th neural unit in the fourth hidden layer'₁，pd′₂，…，pd′_K4)^T，pd′_k4Represents the second activation probability of the k4 th neural unit in the fourth hidden layer; determining a twelfth training matrix hd ' ' storing the second training value of the state quantity of the k4 th neural unit in the fourth hidden layer and a matrix pd ' of the second activation probability of the k4 th neural unit in the fourth hidden layer, which comprises the following specific processes:

c4-5. will

Current value of plus 0.01 x (pd)_k4*hc_k3-pd′_k4*hc″′_k3) Is assigned to

To pair

step C6: carrying out the Y-th iteration updating on the activation probability of each nerve unit in the visible layer, the first hidden layer, the second hidden layer, the third hidden layer and the fourth hidden layer, wherein the specific process is as follows:

between the k1 th nerve cell in the first hidden layer and the k2 th nerve cell in the second hidden layerCurrent value of the connection weight of ha_k1Is the current value of the state quantity of the k1 th neural unit in the first hidden layer;

c6-14, generating a random number between 0 and 1 by using a random function, and carrying out the current value of Pr (k1) and the random numberIn contrast, if the current value of Pr (k1) is greater than the random number, then a value of 1 is assigned to ha_k1Pair ha_k1Updating is carried out; if the current value of Pr (k1) is not greater than the random number, then a value of 0 is assigned to ha_k1Pair ha_k1Updating is carried out;

step C7: creating N reconstructed amplitude matrixes for storing reconstructed CSI amplitude data, wherein each reconstructed amplitude matrix can store 30 rows and 3 columns of reconstructed CSI amplitude data, the sequencing definition of the 30 rows and 3 columns of reconstructed CSI amplitude data stored in each reconstructed amplitude matrix is the same as that of the CS amplitude matrix, and v'_t～v′_11+tAccording to v'_tv′_t+1v′_t+2v′_t+3v′_t+4…v′_t+9v′_t+10v′_t+11The 12-bit binary number is combined into a 12-bit binary number, t is 12n +1, n is 0, 1,2, … and K0/12-1, and the 1 st bit to the 8 th bit from left to right of the 12-bit binary number is used as a reconstruction binary numberCorresponding decimal numbers obtained by calculating the integral part of the CSI amplitude data, the 9 th to the 12 th digits are taken as corresponding decimal numbers obtained by calculating the fraction part of the reconstructed CSI amplitude data, the decimal numbers corresponding to the integral part are taken as integers, the decimal numbers corresponding to the fraction part are taken as the fraction part to be combined as the decimal number corresponding to the 12-digit binary number, the decimal number is taken as the (t-1)/12) +1) -90 (N' -1) th reconstructed CSI amplitude data of the Nth reconstructed amplitude matrix and is stored in the Nth reconstructed amplitude matrix,

c9-1. construct a matrix δ a storing the residual terms of the k0 th neural unit in the visible layer (δ a ═ d₁，δa₂，…，δa_K0)，δa_k0Residual terms for the k0 th neural unit in the visible layer; calculating a residual error term delta a of the k0 th nerve unit in the visible layer by adopting the formula (21)_k0：

δa_k0＝-(v′_k0-v_k0)*v′_k0*(1-v′_k0) (21)

Wherein, v'_k0Training of state quantities for the k0 th neural unit in the visible layerCurrent value of exercise, v_k0The current value of the state quantity of the k0 th nerve unit in the visible layer;

c9-2. construct a matrix δ b that holds the residual terms of the k1 th neural unit in the first hidden layer (δ b ═ b)₁，δb₂，…，δb_K1)，δb_k1Residual terms for the k1 th neural unit in the first hidden layer; calculating a residual error term delta b of the k1 th nerve unit in the first hidden layer by adopting a formula (22)_k1：

Wherein,

c9-3. construct a matrix δ C that holds the residual terms of the k2 th neural unit in the second hidden layer (δ C ═ δ C)₁，δc₂，…，δc_K2)，δc_k2Residual terms for the k2 th neural unit in the second hidden layer; calculating a residual error term deltac of the k2 th neural unit in the second hidden layer by adopting a formula (23)_k2：

Wherein,

c9-4. construct a matrix δ d storing the residual terms of the k3 th neural unit in the third hidden layer (δ d ═ d)₁，δd₂，…，δd_K3)，δd_k3Residual terms for the k3 th neural unit in the third hidden layer; calculating a residual error term delta d of the k3 th nerve unit in the third hidden layer by adopting a formula (24)_k3：

Wherein,

c9-5. construct a matrix δ e storing the residual terms of the k4 th neural unit in the fourth hidden layer (δ e ═ e₁，δe₂，…，δe_K4)，δe_k4Calculating a residual term delta e of the k4 neural unit in the fourth hidden layer by adopting a formula (25) for the residual term of the k4 neural unit in the fourth hidden layer_k4：

Wherein,

is the current value of the connection weight, δ d, between the k3 th neuron in the third hidden layer and the k4 th neuron in the fourth hidden layer_k3Is the residual error of the k3 th nerve unit in the third hidden layerCurrent value of the term, hd_k4Is the current value of the state quantity of the k4 th neural unit in the fourth hidden layer;

c9-6. will

Is added to the current value of (d) by 0.5 x δ d_k3hd_k4Is assigned to

To pair

Updating is carried out; will be provided with

Is added to the current value of 0.5 × δ c_k2hc_k3Is assigned to

To pair

Updating is carried out; will be provided with

Is added by 0.5 × δ b_k1hb_k2Is assigned to

To pair

Is updated to

Is added to the current value of (d) by 0.5 x δ a_k1ha_k1Is assigned to

To pair

fifthly, positioning indoor personnel, and the specific process is as follows:

fifthly, determining a 12-bit binary number corresponding to the mth CSI amplitude data of the Nth' th CSI amplitude matrix in the N CSI amplitude matrices at the current position, wherein the specific process is as follows: judging whether the m-th CSI amplitude data of the N 'th CSI amplitude matrix in the N CSI amplitude matrices at the current position is an integer or a decimal, if the m-th CSI amplitude data of the N' th CSI amplitude matrix in the N CSI amplitude matrices at the current position is an integer, converting the m-th CSI amplitude data into a binary number, supplementing 4 0 s to the right of the obtained binary number to obtain a new binary number, then judging whether the number of bits of the obtained new binary number is equal to 12, if the m-th CSI amplitude data is equal to 12, taking the new binary number as a 12-bit binary number corresponding to the m-th CSI amplitude data of the N 'th CSI amplitude matrix in the N CSI amplitude matrices at the current position, if the m-th CSI amplitude data is greater than 12, selecting 12 bits from right to left as 12-bit binary numbers corresponding to the m-th CSI amplitude data of the N' th CSI amplitude matrix in the N CSI amplitude matrices at the current position, and if the m-th CSI amplitude data is less than 12, after 0 is supplemented to the left of the new binary number until the number of bits is 12 bits, the new binary number is used as a 12-bit binary number corresponding to the mth CSI amplitude data of the Nth' CSI amplitude matrix in the N CSI amplitude matrices at the current position; if the m-th CSI amplitude data of the N' th CSI amplitude matrix in the N CSI amplitude matrices at the current position is a decimal, respectively converting an integer part and a decimal part of the M-th CSI amplitude data into binary numbers, if the number of bits of the binary numbers obtained by converting the integer part is less than 8, complementing 0 at the high bits of the binary numbers to change the binary numbers into 8-bit binary numbers, if the number of bits of the binary numbers obtained by converting the integer part is more than 8, remaining other bits of the 8-bit binary numbers from right to left to delete the 8-bit binary numbers to obtain the 8-bit binary numbers, if the number of bits of the binary numbers obtained by converting the integer part is equal to 8, keeping the number of bits obtained by converting the decimal part unchanged, if the number of bits obtained by converting the decimal part is equal to 4, keeping the number of bits obtained by converting the decimal part unchanged, and if the number of bits obtained by converting the decimal part is less than 4If the number of bits of the binary number obtained by converting the decimal part is greater than 4, remaining 4 bits from left to right and deleting the other bits to obtain 4-bit binary numbers, using the 8-bit binary number obtained by processing the integer part as a high 8 bit, splicing the 4-bit binary number obtained by processing the decimal part as a low 4 bit to obtain a 12-bit binary number, and using the spliced 12-bit binary number as a 12-bit binary number corresponding to the mth CSI amplitude data of the Nth' CSI amplitude matrix in the N CSI amplitude matrices at the current position; n ″' -1, 2, …, N; m is 1,2, …, 90; assigning the value of the x-th bit counted from left to right to v by the twelve-bit binary value of the m-th amplitude data_{12*3*30*(N″′-1)+(m-1)*12+x}To v is to v₁，v₂，v₃，...，v_K0Updating, wherein x is 1,2, …, 12;

5-2, calculating by formula (27) to obtain Pr (v)_l)：

Fifthly, recording the probability of the indoor position as

7 recording the coordinate of the current position of the indoor person to be positioned as (x)_l′，y_l′) Calculating (x) using equations (29) and (30)_l′，y_l′)：

(x) calculated by (8)_l′，y_l′) Namely the positioning coordinates of the indoor person to be positioned.