WO2016187746A1 - 提高人工神经网络定位性能的方法和装置 - Google Patents

提高人工神经网络定位性能的方法和装置 Download PDF

Info

Publication number
WO2016187746A1
WO2016187746A1 PCT/CN2015/079553 CN2015079553W WO2016187746A1 WO 2016187746 A1 WO2016187746 A1 WO 2016187746A1 CN 2015079553 W CN2015079553 W CN 2015079553W WO 2016187746 A1 WO2016187746 A1 WO 2016187746A1
Authority
WO
WIPO (PCT)
Prior art keywords
initial
pso algorithm
ann
particle
optimized
Prior art date
Application number
PCT/CN2015/079553
Other languages
English (en)
French (fr)
Inventor
李安俭
崔杰
韩静
李红
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to PCT/CN2015/079553 priority Critical patent/WO2016187746A1/zh
Priority to JP2017560712A priority patent/JP6596516B2/ja
Priority to EP15892831.7A priority patent/EP3282781A4/en
Priority to CN201580069876.4A priority patent/CN107113764B/zh
Publication of WO2016187746A1 publication Critical patent/WO2016187746A1/zh
Priority to US15/818,979 priority patent/US20180089566A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S5/00—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations
    • G01S5/02—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations using radio waves
    • G01S5/0252—Radio frequency fingerprinting
    • G01S5/02521—Radio frequency fingerprinting using a radio-map
    • G01S5/02524—Creating or updating the radio-map
    • G01S5/02525—Gathering the radio frequency fingerprints
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W64/00—Locating users or terminals or network equipment for network management purposes, e.g. mobility management

Definitions

  • Embodiments of the present invention relate to the field of wireless communication technologies, and in particular, to a method and apparatus for improving positioning performance of an artificial neural network.
  • the positioning technology is a technology adopted to determine the geographical location of the terminal, and the geographic location information of the terminal can be directly or indirectly obtained by using the resources of the wireless communication network.
  • positioning technology mainly focuses on outdoor positioning.
  • indoor positioning is used for positioning in emergency situations, public safety, commercial and military applications.
  • satellite-based positioning systems have achieved great success in outdoor environments.
  • due to the occlusion of buildings, these systems are difficult to apply to indoor environments.
  • WLAN Wireless Local Area Networks
  • the distance-independent method is energy efficient and low cost, but it is only suitable for networks with high connectivity, which is less accurate.
  • the distance-based approach is based primarily on the measured signal, which is then applied to the mathematical model of the channel. Since it is very difficult to determine the channel model in multipath propagation and fast fading environments, it is difficult to achieve higher precision positioning by these methods.
  • the fingerprint matching positioning method has very good potential to obtain better positioning accuracy. This method has its own unique tags based on each location, such as RSSI. The tag information corresponding to some locations is pre-stored in a database. When locating, the specific location of the UE is determined by matching the tag information in the database.
  • ANN Artificial Neutral Network
  • the training phase is to train the artificial neural network using the geographical location information and fingerprint information of the known location, and then measure the location to be located in the positioning phase.
  • the obtained fingerprint information is input into the trained artificial neural network, thereby obtaining the geographical location information of the point to be located.
  • an initial artificial neural network is generated first, and then the initial artificial neural network is trained to obtain the trained artificial neural network.
  • ANN has excellent performance in the WSN positioning system.
  • the shortcoming of ANN is the search for global optimal aspects, especially in scenarios with incomplete or non-ideal information. Therefore, in the actual indoor environment, due to the existence of multipath and fast decay factors, the method is difficult to have better performance. That is, when the initial artificial neural network generated immediately is trained in the training phase, the trained artificial neural network may not be the optimal artificial neural network, and the positioning using the trained artificial neural network may affect the positioning performance.
  • the embodiment of the invention provides a method and a device for improving the positioning performance of an artificial neural network, and optimizes the weight and offset of each neuron in the ANN by using the PSO algorithm, thereby improving the performance of using the ANN for positioning.
  • the first aspect provides a method for improving the positioning performance of an artificial neural network, including:
  • Fingerprint information of the at least one test point is used as an input layer neuron, and geographic location information of the at least one test point is used as an output layer neuron, and an initial input, an output layer, and an initial of each neuron in the hidden layer are randomly selected. Weight and initial offset, establishing an initial ANN;
  • the optimized initial ANN is trained to obtain the trained ANN.
  • the using the PSO optimizes the initial weight and the initial offset of each neuron of the initial ANN to obtain an optimized initial ANN, including:
  • the optimized initial ANN is established using the optimized weight and the optimized offset of each neuron of the initial ANN.
  • the PSO algorithm is used to optimize an initial weight and an initial offset of each neuron of the initial ANN to obtain an optimized initial ANN.
  • the optimized initial ANN is established using the optimized weight and the optimized offset of each neuron of the initial ANN.
  • the calculating the fitness of each of the PSO algorithms includes:
  • Updating the optimal particles in the population of the PSO algorithm includes:
  • the particles with the smallest error are used as the best particles in the population of the PSO algorithm.
  • the ANN is a backward propagating ANN or a generalized regression linear network. .
  • the fingerprint information is RSSI.
  • the second aspect provides an apparatus for improving the positioning performance of an artificial neural network, including:
  • An obtaining module configured to acquire geographic location information and fingerprint information of at least one test point
  • An initial information module configured to use fingerprint information of the at least one test point as an input layer neuron, and use the geographical location information of the at least one test point as an output layer neuron, and randomly select an input layer, an output layer, and a hidden layer Initial weight and initial offset of each neuron, establishing an initial ANN;
  • An optimization module configured to optimize an initial weight and an initial offset of each neuron of the initial ANN using a PSO algorithm to obtain an optimized initial ANN
  • the training module is configured to train the optimized initial ANN to obtain the trained ANN.
  • the optimization module is specifically configured to use an initial weight and an initial offset of each neuron in the initial ANN as a particle of a PSO algorithm.
  • Initial position randomly selecting the initial velocity of each particle, establishing a population of the PSO algorithm; in each generation of the PSO algorithm iterative process, sequentially calculating the fitness of each particle in the PSO algorithm, and updating the PSO algorithm
  • Optimal particles in the population updating the position of each particle in the PSO algorithm population and its velocity until the iterative algebra of the PSO algorithm ends; using the optimal position of each particle in the PSO algorithm population as the initial ANN
  • the optimized weight and the optimized offset of each neuron; the optimized initial ANN is established using the optimized weight and the optimized offset of each neuron of the initial ANN.
  • the optimization module is specifically configured to use an initial weight and an initial offset of each neuron in the initial ANN as a particle of a PSO algorithm.
  • Initial position randomly selecting the initial velocity of each particle, establishing a population of the PSO algorithm; in each generation of the iterative process of the PSO algorithm, sequentially calculating each of the PSO algorithms Adaptability of the child, updating the optimal particles in the population of the PSO algorithm, updating the position of each particle in the PSO algorithm population and its speed until the sum of the fitness of all the particles in the PSO algorithm is less than a preset threshold; Using the optimal position of each particle in the PSO algorithm population as the optimized weight and the optimized offset of each neuron of the initial ANN; using the optimized weight and optimization of each neuron of the initial ANN The post-offset establishes the optimized initial ANN.
  • the optimization module is specifically configured to use the location of each particle in the PSO algorithm as an initial a weight and an offset in the ANN; calculating, as the input layer neuron, fingerprint information of the at least one test point, calculating an output layer neuron of the initial ANN; calculating the output layer neuron and the at least one test point
  • the error of the geographical location information, the error is taken as the fitness of the particle; the particle with the smallest error is used as the optimal particle in the population of the PSO algorithm.
  • the ANN is a backward propagating ANN or a generalized regression linear network. .
  • the fingerprint information is an indication RSSI.
  • the method and device for improving the positioning performance of the artificial neural network when the artificial neural network is used for wireless positioning, when the initial artificial neural network is established by using the random initial weight and the offset, the PSO algorithm is used for the initial artificial
  • the initial weight and initial offset of each layer of the neural network are optimized, and the optimized initial artificial neural network is obtained.
  • the optimized artificial neural network is trained to obtain the trained artificial neural network.
  • the initial weight and initiality are obtained by using the PSO algorithm.
  • the optimization of the offset can be used to obtain the globally optimal initial weight and the initial offset. Therefore, the positioning performance of the artificial neural network provided by the embodiment of the present invention can be improved.
  • 1 is a schematic structural diagram of an artificial neural network algorithm
  • Embodiment 1 is a flowchart of Embodiment 1 of a method for improving positioning performance of an artificial neural network according to an embodiment of the present invention
  • FIG. 3 is a flowchart of Embodiment 2 of a method for improving positioning performance of an artificial neural network according to an embodiment of the present disclosure
  • Embodiment 4 is a flowchart of Embodiment 3 of a method for improving positioning performance of an artificial neural network according to an embodiment of the present invention
  • FIG. 5 is a flowchart of Embodiment 4 of a method for improving positioning performance of an artificial neural network according to an embodiment of the present invention
  • FIG. 6 is a schematic structural diagram of Embodiment 1 of an apparatus for improving positioning performance of an artificial neural network according to an embodiment of the present disclosure
  • FIG. 7 is a schematic structural diagram of Embodiment 2 of an apparatus for improving positioning performance of an artificial neural network according to an embodiment of the present invention.
  • ANN is a well-known pattern matching algorithm applied to multi-layer and multi-join artificial neural networks.
  • Artificial neurons simulate the neurons of a living body by using an activation function.
  • Each neuron has an activation function that is responsible for mapping the input of the neuron to its output.
  • the structure of a neural network depends on how the artificial neurons of different layers are connected to each other.
  • Each neuron has its own weight and offset. The weights and offsets can be adjusted during the training phase. This process of learning is called supervised learning and is used to find an optimal input-to-output mapping function.
  • the artificial neural network calculates an output below a set error threshold
  • the artificial neural network ends its training phase.
  • the trained network can then be used to solve pattern recognition problems.
  • Figure 1 is a schematic diagram of the structure of an artificial neural network algorithm. As shown in Figure 1, the entire artificial neural network is divided into an input layer, a hidden layer, and an output layer. Wherein, each node of the input layer, the hidden layer, and the output layer (also known as neurons) are related to each other, each node represents a specific function output, and the connection between each two nodes represents a weighted value, ie weight, of the signal through the connection.
  • each node of the input layer, the hidden layer, and the output layer also known as neurons
  • each node represents a specific function output
  • the connection between each two nodes represents a weighted value, ie weight, of the signal through the connection.
  • the purpose of training the artificial neural network is to find the functional relationship and the weight and offset of each layer node, so as to obtain the nonlinear functional relationship between the input layer node and the output layer node.
  • the fingerprint information needs to be measured at some measurement points first in the training phase of the ANN.
  • the fingerprint information as an RSSI as an example
  • the RSSI of each received wireless access point (AP) is measured at a measurement point of known geographical location information (for example, latitude and longitude information).
  • a plurality of RSSIs of one measurement point are used as input neurons of the artificial neural network, and the position information of the measurement points is used as an output neuron, and the artificial neural network is trained using the RSSI and position information of the plurality of measurement points until the artificial neural network The output is below the set error threshold. This completes the training process of the artificial neural network.
  • the RSSI of each AP received by the to-be-positioned point is input into the trained artificial neural network, and the position information with the positioning point is obtained, thereby completing the positioning.
  • the process of training the network in the process of training the network, it is first necessary to randomly select the initial weight and initial offset of each node (neuron), and then start the artificial nerve with the initial weight and the initial offset.
  • the network is trained.
  • the randomly selected initial weight and initial offset may have a large difference from the optimal weight and offset.
  • the weight and offset may be converged to a local optimal value. .
  • the use of locally optimal weights and offsets for wireless positioning will have an impact on positioning performance.
  • Embodiment 1 is a flowchart of Embodiment 1 of a method for improving positioning performance of an artificial neural network according to an embodiment of the present invention. As shown in FIG. 2, the method in this embodiment includes:
  • Step S201 Acquire geographical location information and fingerprint information of at least one test point.
  • the positioning method provided by this embodiment is based on the ANN algorithm shown in FIG. 1.
  • the positioning method provided in this embodiment includes two stages of a training phase and a positioning phase.
  • the training phase trains the artificial neural network based on the known position information and fingerprint information of some measurement points, and obtains the weight and offset of each layer of the artificial neural network.
  • the measured fingerprint information is input to the trained artificial neural network to obtain the location information of the location to be located.
  • the training phase of positioning is first performed.
  • geographical location information and fingerprint information of at least one test point need to be acquired, and the number of test points is increased.
  • the artificial neural network obtained by training has better performance.
  • the fingerprint information of each test point may be any measurement quantity of the specific mark that can be characterized by the test point, such as RSSI and the like. Since in a wireless network, each test point may receive signals from multiple APs, the fingerprint information of each test point may be composed of multiple measurements.
  • the geographical location information of the test point can be determined by any known positioning method, for example, using a positioning system such as a Global Positioning System (GPS).
  • GPS Global Positioning System
  • the geographical location information of the test point is generally a latitude and longitude value.
  • Step S202 using the fingerprint information of the at least one test point as the input layer neuron, using the geographical location information of the at least one test point as the output layer neuron, and randomly selecting the initial weight of each of the input layer, the output layer, and the hidden layer.
  • the initial artificial neural network is established with the value and initial offset.
  • an initial artificial neural network needs to be established.
  • the fingerprint information of the at least one test point is used as the input layer neuron
  • the geographical location information of the at least one test point is respectively used as the output layer neuron.
  • the initial weight and initial offset of each neuron in the input layer, the output layer and the hidden layer are randomly selected to establish an initial artificial neural network.
  • the number of input layer neurons in the initial artificial neural network differs according to the number of measured quantities measured at each test point.
  • the number of input layer neurons in the initial artificial neural network is two, that is, the geographical position of the test point. Longitude and dimension values.
  • the initial artificial neural network establishment method of this step is the same as the existing method of applying ANN to the wireless positioning technology.
  • Step S203 using Particle Swarm Optimization (PSO) algorithm to optimize the initial weight and initial offset of each neuron of the initial artificial neural network, and obtain an optimized initial artificial neural network.
  • PSO Particle Swarm Optimization
  • the problem of optimal weight and offset may not be obtained.
  • the PSO algorithm is applied to the artificial nerve.
  • the PSO algorithm is used to optimize the initial weight and initial offset of the randomly selected initial artificial neural network, so as to obtain the optimized initial artificial neural network.
  • the PSO algorithm belongs to the particle swarm theory and was discovered when simulating a simplified social model. It simulates the characteristics of birds and fish stocks. In a particle swarm optimization, there are a group of candidate schemes that become individuals or particles. At the same time, these individuals or particles evolve through cooperation or competition.
  • Each particle has a fitness value determined by the objective function, and each particle knows the best (pbest) and current position it has found so far.
  • Each particle also knows the best position (gbest) found by all particles in the entire particle population so far.
  • Each particle follows the optimal particle in the entire particle swarm to search in space. After multiple iterations, it finally finds the optimal solution in the whole space (the best position in the whole space).
  • the PSO algorithm has advantages in finding the global optimal solution. Therefore, this embodiment applies the PSO algorithm to the artificial neural network, and uses the PSO algorithm to find the optimal initial weight and the optimal initial offset of the initial artificial neural network. Get the optimal initial artificial neural network.
  • the initial weight and initial offset of the initial artificial neural network are used as particles of the PSO algorithm. Through the competition and cooperation between particles, these particles can find their best position in the search space.
  • the weights and offsets of the corresponding iterations of each particle are substituted into the artificial neural network, the fingerprint information of each test point is used as an input, and then each test point is calculated and processed by the artificial neural network. The error between the resulting output and the actual geographic location information of the test point.
  • the error minimum weight and offset are the best positions (gbest) for this round of PSO algorithm iteration.
  • Combining the above errors of each test point is the total error of the iterative process of the PSO algorithm.
  • the goal of optimizing the initial weight and the initial offset using the PSO algorithm is to make the total error less than a preset threshold. If the calculated total error is not less than the preset threshold, the next calculation of the PSO algorithm is performed, otherwise The optimized initial weight and initial offset will be obtained, resulting in an optimized initial artificial neural network. Or the goal of optimizing the initial weight and the initial offset using the PSO algorithm is that when the number of iterations of the PSO algorithm reaches a preset number of times, when the number of iterations reaches a preset number of times, the optimization process ends, thereby obtaining an optimized initial artificial neural network. .
  • Step S204 training the optimized initial artificial neural network to obtain a trained artificial neural network.
  • the optimized initial artificial neural network will be trained using the ANN algorithm.
  • the ANN algorithm used for training the optimized initial artificial neural network can be any kind of ANN algorithm, such as Back-Propagation-ANN (BP-ANN) generalized regression linear network (General Regression) Neural Network, GRNN), etc.
  • BP-ANN Back-Propagation-ANN
  • General Regression General Regression
  • GRNN General Regression Neural Network
  • each neuron learns from each other and finally obtains the artificial neural network after training. Thereby completing the training process of the artificial neural network.
  • the trained artificial neural network After completing the training process of the artificial neural network, the trained artificial neural network can be used for positioning.
  • the trained artificial neural network is obtained, it is equivalent to obtaining a mapping function from the fingerprint information of the location point to the geographical location information.
  • the fingerprint information of the device to be located is obtained first, and the fingerprint information needs to be the same measurement amount as the fingerprint information acquired in the training phase in step S201.
  • After obtaining the fingerprint information of the point to be located of the device it is used as the input layer neuron of the trained artificial neural network, and the input layer neurons processed by the trained artificial neural network are the pending of the device. Location information for the location. Thereby, the positioning process for the positioning point is completed.
  • the initial artificial neural network when the artificial neural network is used for wireless positioning, when the initial artificial neural network is established by using the random initial weight and the offset, the initial artificial neural network is used by using the PSO algorithm.
  • the initial weight and initial offset of each layer are optimized, and the optimized initial artificial neural network is obtained.
  • the optimized artificial neural network is trained to obtain the trained artificial neural network.
  • the initial weight and initial offset are used due to the PSO algorithm.
  • the optimization of the quantity can obtain the global optimal initial weight and the initial offset. Therefore, the positioning performance of the artificial neural network generated by the positioning method provided by the embodiment is high.
  • FIG. 3 is a flowchart of Embodiment 2 of a method for improving positioning performance of an artificial neural network according to an embodiment of the present invention. Specifically, the embodiment is a specific execution process of step S203 in the embodiment shown in FIG. 2 . As shown in FIG. 3, the method in this embodiment includes:
  • Step S301 using the initial weight and the initial offset of each neuron in the initial artificial neural network as the initial position of the particle of the PSO algorithm, randomly selecting the initial velocity of each particle, and establishing a population of the PSO algorithm.
  • the initial weight and initial offset of each neuron in the initial artificial neural network are taken as the initial positions of the particles in the PSO algorithm.
  • the initial velocity of each particle is randomly selected to complete the establishment of the PSO algorithm population.
  • the goal of the end of the PSO algorithm is that the iterative algebra ends, that is, the iterative algebra reaches the predetermined algebra. In each iteration, you can cycle through the steps in turn S302-S305:
  • Step S302 calculating the fitness of each particle in the PSO algorithm.
  • Step S303 updating the optimal particles in the population of the PSO algorithm.
  • the optimal position (gbest) in the population of the PSO algorithm is updated.
  • Step S304 updating each particle position and its velocity in the PSO algorithm population.
  • step S305 it is determined whether the iteration algebra of the PSO algorithm ends.
  • step S302 is repeatedly performed, otherwise step S306 is performed.
  • Step S306 the optimal position of each particle in the PSO algorithm population is used as the optimized weight and the optimized offset of each neuron of the initial artificial neural network.
  • the optimal position (pbest) of each particle in the PSO algorithm is used as the optimized weight and the optimized offset of each neuron of the initial artificial neural network corresponding to the particle.
  • Step S307 using the optimized weight of each neuron of the initial artificial neural network and the optimized offset to establish an optimized initial artificial neural network.
  • FIG. 4 is a flowchart of Embodiment 3 of a method for improving positioning performance of an artificial neural network according to an embodiment of the present invention. Specifically, this embodiment is another specific execution process of step S203 in the embodiment shown in FIG. 2 . As shown in FIG. 4, the method in this embodiment includes:
  • Step S401 using the initial weight and initial offset of each neuron in the initial artificial neural network as the initial position of the particle of the PSO algorithm, randomly selecting the initial velocity of each particle, and establishing a population of the PSO algorithm.
  • the initial weight and initial offset of each neuron in the initial artificial neural network are taken as The initial position of each particle in the PSO algorithm.
  • the initial velocity of each particle is randomly selected to complete the establishment of the PSO algorithm population.
  • steps S402-S405 may be sequentially performed in a loop:
  • Step S402 calculating the fitness of each particle in the PSO algorithm.
  • Step S403 updating the optimal particles in the population of the PSO algorithm.
  • the optimal position (gbest) in the population of the PSO algorithm is updated.
  • Step S404 updating the position of each particle in the PSO algorithm population and its velocity.
  • Step S405 determining whether the sum of the fitnesss of all the particles of the PSO algorithm is less than a preset threshold.
  • the fitness of each particle in the PSO algorithm indicates its error with the optimal position, and the smaller the fitness, the smaller the error. Therefore, when the fitness of each particle of the PSO algorithm is small, each particle is closer to the optimal position.
  • the sum of the fitness of all the particles in the PSO algorithm is less than the preset threshold, the average distance between the particles optimized by the PSO algorithm and the optimal position will be closer, so that the optimized artificial neural network generated by the optimized particles is used.
  • the optimized weights and optimized offsets for each neuron will also be preferred. Therefore, after each iteration of the PSO algorithm, the sum of the fitnesss of all the particles is calculated.
  • step S402 If the sum of the fitnesss of all the particles of the PSO algorithm is not less than the preset threshold, step S402 is repeatedly performed, otherwise step S406 is performed.
  • the preset threshold can be preset in the system according to experience. The smaller the preset threshold, the better the result of the PSO algorithm optimization, but the longer the optimization time will be.
  • Step S406 the optimal position of each particle in the PSO algorithm population is used as the optimized weight and the optimized offset of each neuron of the initial artificial neural network.
  • the optimal position (pbest) of each particle in the PSO algorithm is used as the initial artificial neural network corresponding to the particle.
  • the optimized weight of the element and the optimized offset is used as the initial artificial neural network corresponding to the particle.
  • Step S407 an optimized initial artificial neural network is established by using the optimized weight of each neuron of the initial artificial neural network and the optimized offset.
  • FIG. 5 is a flowchart of Embodiment 4 of a method for improving the positioning performance of an artificial neural network according to an embodiment of the present invention.
  • the embodiment is the step S302 and the step S303 (or the implementation shown in FIG. 4) in the embodiment shown in FIG. A specific execution process of step S402 and step S403) in the example.
  • the method in this embodiment includes:
  • step S501 the position of each particle in the PSO algorithm is used as the weight and offset in the initial artificial neural network.
  • Step S502 Calculate the output layer neurons of the initial artificial neural network by using the fingerprint information of the at least one test point as the input layer neuron.
  • Step S503 calculating an error of the geographic location information of the output layer neuron and the at least one test point, and using the error as the fitness of the particle.
  • step S504 the particle with the smallest error is used as the optimal particle in the population of the PSO algorithm.
  • this embodiment is a specific method for calculating the fitness of each particle in the PSO algorithm and selecting the optimal particle in the PSO algorithm. That is to say, the fingerprint information and geographical location information of the test point are used as the optimization target of the PSO algorithm, and the particle is optimized by the PSO algorithm.
  • FIG. 6 is a schematic structural diagram of Embodiment 1 of an apparatus for improving positioning performance of an artificial neural network according to an embodiment of the present disclosure. As shown in FIG. 6, the method in this embodiment includes:
  • the obtaining module 61 is configured to acquire geographic location information and fingerprint information of at least one test point.
  • the initial information module 62 is configured to use the fingerprint information of the at least one test point as the input layer neuron, the geographic location information of the at least one test point as the output layer neuron, and randomly select each of the input layer, the output layer, and the hidden layer.
  • the initial weight of the element and the initial offset establish the initial ANN.
  • the optimization module 63 is configured to optimize the initial weight and the initial offset of each neuron of the initial ANN using the PSO algorithm to obtain an optimized initial ANN.
  • the training module 64 is configured to train the optimized initial ANN to obtain the trained ANN.
  • the positioning device provided by this embodiment is used to implement the technical solution of the method embodiment shown in FIG. 2, and the implementation principle and technical effects are similar, and details are not described herein again.
  • the optimization module 63 is specifically configured to use the initial weight and the initial offset of each neuron in the initial ANN as the initial position of the particle of the PSO algorithm. Randomly selecting the initial velocity of each particle to establish a population of the PSO algorithm; in each generation of the PSO algorithm iterative process, calculating the fitness of each particle in the PSO algorithm and updating the population of the PSO algorithm Optimal particles, updating each particle position and its velocity in the PSO algorithm population until the iterative algebra of the PSO algorithm ends; using the optimal position of each particle in the PSO algorithm population as the initial ANN for each nerve The optimized weight of the element and the optimized offset; the optimized initial ANN is established using the optimized weight and the optimized offset of each neuron of the initial ANN.
  • the optimization module 63 is specifically configured to use the initial weight and the initial offset of each neuron in the initial ANN as the initial position of the particle of the PSO algorithm, and randomly select each particle.
  • Initial velocity establishing a population of the PSO algorithm; in each generation of the PSO algorithm iterative process, sequentially calculating the fitness of each particle in the PSO algorithm, and updating the optimal particle in the population of the PSO algorithm Updating the position of each particle in the PSO algorithm population and its velocity until the sum of the fitnesss of all the particles in the PSO algorithm is less than a preset threshold; using the optimal position of each particle in the PSO algorithm population as the initial
  • the optimization module 63 is specifically configured to use each particle in the PSO algorithm as a weight and an offset in the initial ANN; and the fingerprint information of the at least one test point.
  • As an input layer neuron calculating an output layer neuron of the initial ANN; calculating an error of the geographic location information of the output layer neuron and the at least one test point, using the error as the fitness of the particle; The smallest particle is the best particle in the population of the PSO algorithm.
  • the artificial neural network is BP-ANN or GRNN.
  • the fingerprint information is RSSI.
  • the positioning apparatus provided by the embodiment of the present invention may be disposed in a base station in a wireless network, or may be an additional network device independent of the base station.
  • FIG. 7 is a schematic structural diagram of Embodiment 2 of an apparatus for improving positioning performance of an artificial neural network according to an embodiment of the present invention.
  • the positioning apparatus of this embodiment includes: a receiver 71 and a processor 72.
  • the positioning device may further include a memory 73.
  • the receiver 71, the processor 72 and the memory 73 can be connected through a system bus or other means.
  • the system bus connection is taken as an example in FIG. 7; the system bus can be an Industrial Standard Architecture (ISA). Bus, Peripheral Component Interconnect (PCI) bus or Extended Industrial Standard Architecture (EISA) bus.
  • ISA Industrial Standard Architecture
  • PCI Peripheral Component Interconnect
  • EISA Extended Industrial Standard Architecture
  • the system bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only one line is shown in Figure 7, but it does not mean that there is only one bus or one type of bus.
  • the receiver 71 is configured to acquire geographic location information and fingerprint information of at least one test point.
  • the processor 72 is configured to use fingerprint information of the at least one test point as an input layer neuron, and use the geographical location information of the at least one test point as an output layer neuron, and randomly select an input layer, an output layer, and a hidden layer.
  • Initial weight and initial offset of each neuron an initial ANN is established; an initial weight and an initial offset of each neuron of the initial ANN are optimized using a PSO algorithm to obtain an optimized initial ANN; The optimized initial ANN is trained to obtain the trained ANN.
  • the memory 73 is configured to store the data received by the receiver 71 and provide the data to the processor 72 for processing.
  • the positioning apparatus provided by the embodiment of the present invention may be disposed in a base station in a wireless network, or may be an additional network device independent of the base station.
  • the processor 72 is specifically configured to randomly select an initial weight and an initial offset of each neuron in the initial ANN as the initial position of the particle of the PSO algorithm.
  • the initial velocity of each particle establishes a population of the PSO algorithm; in each generation of the PSO algorithm iterative process, the fitness of each particle in the PSO algorithm is sequentially calculated, and the most of the population of the PSO algorithm is updated.
  • the processor 72 is specifically configured to randomly select an initial weight and an initial offset of each neuron in the initial ANN as the initial position of the particle of the PSO algorithm.
  • the initial velocity of each particle establishes a population of the PSO algorithm; in each generation of the PSO algorithm iterative process, the fitness of each particle in the PSO algorithm is sequentially calculated, and the most of the population of the PSO algorithm is updated.
  • the processor 72 is specifically configured to use a position of each particle in the PSO algorithm as a weight and an offset in the initial ANN; and the at least one test
  • the fingerprint information of the point is used as an input layer neuron to calculate an output layer neuron of the initial ANN; an error of the geographic location information of the output layer neuron and the at least one test point is calculated, and the error is used as the adaptation of the particle
  • the smallest particle is used as the best particle in the population of the PSO algorithm.
  • the ANN is a backward propagating ANN or a generalized regression linear network.
  • the fingerprint information is RSSI.
  • the aforementioned program can be stored in a computer readable storage medium.
  • the program when executed, performs the steps including the foregoing method embodiments; and the foregoing storage medium includes various media that can store program codes, such as a ROM, a RAM, a magnetic disk, or an optical disk.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Remote Sensing (AREA)
  • Radar, Positioning & Navigation (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Position Fixing By Use Of Radio Waves (AREA)

Abstract

本发明实施例提供一种提高人工神经网络定位性能的方法和装置，该方法包括：建立初始ANN；使用PSO算法对初始ANN进行优化，得到优化后的初始ANN；对优化后的初始ANN进行训练，得到训练后的ANN。本发明实施例提供的定位方法和装置，通过使用PSO算法对ANN中各神经元的权重和偏移量进行优化，提高了采用ANN进行定位的性能。

Description

提高人工神经网络定位性能的方法和装置 技术领域
本发明实施例涉及无线通信技术领域，尤其涉及一种提高人工神经网络定位性能的方法和装置。
背景技术
定位技术是为了确定终端的地理位置而采用的技术，可以利用无线通信网络的资源来直接或间接地得到终端的地理位置信息。
当前，定位技术主要关注在室外定位，随着无线技术的发展，室内定位受到越来越多的关注。室内定位被应用于紧急情况下的定位，公共安全，商业和军事应用等。现在业界已经研究了很多的定位方法。依赖于卫星的定位系统在室外环境中获得了巨大的成功。但是，由于建筑物的遮挡，这些系统很难应用于室内环境。
一种实现室内定位的可行方法是使用无线传感网络(Wireless Sensor Networks，WSN)或者无线局域网(Wireless Local Area Networks，WLAN)。这些都是一些低成本且稳定的无线技术。目前主要由三种室内定位的方法，包括同距离无关的方法，基于距离的方法及指纹匹配方法。同距离无关的方法能效比高且成本低，但是仅适用于具有高连接性的网络，这种方法精度较差。基于距离的方法主要基于测量到的信号，然后，再应用信道的数学模型来进行定位。因为确定多径传播和快衰环境下的信道模型是非常困难的，因此这些方法是很难实现较高精度的定位。指纹匹配定位方法具有非常好的潜力来获得较好的定位精度。这种方法基于每个位置都有它特有的标记，例如RSSI等。一些位置对应的标记信息被预先存储在一个数据库中。当定位时，通过匹配数据库中的标记信息来确定UE的具体位置。
一般地，有两种指纹匹配的定位方法。一种是经典的非智能方法，例如最相邻(Nearest Neighbors，NN)方法，及K个最邻近的(K Nearest Neighbors，KNN)方法，及K个加权最邻近(K Weighted Nearest Neighbors，KWNN)方法等。这些方法成本较低，但是精度较差。第二种方法就是众所周知的智能方法。智能方法也成为指纹匹配方法。人工神经网络(Artificial Neutral  Network，ANN)就是指纹匹配方法的一种，它在室内定位技术上有很好的潜力。
将ANN应用与定位技术时，分为训练阶段和定位阶段两个阶段，其中训练阶段是使用已知地点的地理位置信息和指纹信息对人工神经网络进行训练，然后在定位阶段将待定位点测量到的指纹信息输入训练后的人工神经网络，从而得到待定位点的地理位置信息。其中在训练阶段是先随即生成一个初始人工神经网络，然后再对该初始人工神经网络进行训练从而得到训练后的人工神经网络。
ANN在WSN定位系统中性能优异。但是ANN的缺点在于搜索全局最优方面，特别是在具有不完整或非理想的信息的场景下。因此，在实际的室内环境下，由于多径和快衰因素的存在，使得该方法很难有较好的性能。也即，在训练阶段对随即生成的初始人工神经网络进行训练时，训练后的人工神经网络可能不是最优的人工神经网络，导致使用该训练后的人工神经网络进行定位可能影响定位性能。
因此，在应用ANN进行定位时，如何在训练阶段得到最优的人工神经网络，以提高定位性能是亟需解决的关键问题。
发明内容
本发明实施例提供一种提高人工神经网络定位性能的方法和装置，通过使用PSO算法对ANN中各神经元的权重和偏移量进行优化，提高了采用ANN进行定位的性能。
第一方面提供一种提高人工神经网络定位性能的方法，包括：
获取至少一个测试点的地理位置信息和指纹信息；
以所述至少一个测试点的指纹信息作为输入层神经元，以所述至少一个测试点的地理位置信息作为输出层神经元，随机选取输入层、输出层、隐藏层中每一神经元的初始权值和初始偏移量，建立初始ANN；
使用PSO算法对所述初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN；
对所述优化后的初始ANN进行训练，得到训练后的ANN。
结合第一方面，在第一方面第一种可能的实现方式中，所述使用PSO算 法对所述初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN，包括：
将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；
在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法的迭代代数结束；
将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；
使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
结合第一方面，在第一方面第二种可能的实现方式中，所述使用PSO算法对所述初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN，包括：
将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；
在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法中所有粒子的适应度之和小于预设阈值；
将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；
使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
结合第一方面第一种或第二种可能的实现方式，在第一方面第三种可能的实现方式中，所述计算所述PSO算法中每一粒子的适应度，包括：
使用所述PSO算法中每一粒子的位置作为初始ANN中的权值和偏移量；
以所述至少一个测试点的指纹信息作为输入层神经元，计算初始ANN的输出层神经元；
计算所述输出层神经元和所述至少一个测试点的地理位置信息的误差， 将所述误差作为所述粒子的适应度；
所述更新所述PSO算法的种群中的最佳粒子，包括：
将误差最小的粒子作为所述PSO算法的种群中的最佳粒子。
结合第一方面至第一方面第三种可能的实现方式中任一种可能的实现方式，在第一方面第四种可能的实现方式中，所述ANN为后向传播ANN或广义回归线性网络。
结合第一方面至第一方面第四种可能的实现方式中任一种可能的实现方式，在第一方面第五种可能的实现方式中，所述指纹信息为RSSI。
第二方面提供一种提高人工神经网络定位性能的装置，包括：
获取模块，用于获取至少一个测试点的地理位置信息和指纹信息；
初始信息模块，用于以所述至少一个测试点的指纹信息作为输入层神经元，以所述至少一个测试点的地理位置信息作为输出层神经元，随机选取输入层、输出层、隐藏层中每一神经元的初始权值和初始偏移量，建立初始ANN；
优化模块，用于使用PSO算法对所述初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN；
训练模块，用于对所述优化后的初始ANN进行训练，得到训练后的ANN。
结合第二方面，在第二方面第一种可能的实现方式中，所述优化模块，具体用于将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法的迭代代数结束；将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
结合第二方面，在第二方面第二种可能的实现方式中，所述优化模块，具体用于将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒 子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法中所有粒子的适应度之和小于预设阈值；将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
结合第二方面第一种或第二种可能的实现方式，在第二方面第三种可能的实现方式中，所述优化模块，具体用于使用所述PSO算法中每一粒子的位置作为初始ANN中的权值和偏移量；以所述至少一个测试点的指纹信息作为输入层神经元，计算初始ANN的输出层神经元；计算所述输出层神经元和所述至少一个测试点的地理位置信息的误差，将所述误差作为所述粒子的适应度；将误差最小的粒子作为所述PSO算法的种群中的最佳粒子。
结合第二方面至第二方面第三种可能的实现方式中任一种可能的实现方式，在第二方面第四种可能的实现方式中，所述ANN为后向传播ANN或广义回归线性网络。
结合第二方面至第二方面第四种可能的实现方式中任一种可能的实现方式，在第二方面第五种可能的实现方式中，所述指纹信息为指示RSSI。
本实施例提供的提高人工神经网络定位性能的方法和装置，在应用人工神经网络进行无线定位时，当使用随机的初始权重和偏移量建立了初始人工神经网络后，使用PSO算法对初始人工神经网络各层的初始权重和初始偏移量进行优化，得到优化后的初始人工神经网络，对优化后的人工神经网络进行训练得到训练后的人工神经网络，由于使用PSO算法对初始权重和初始偏移量进行优化能够得到全局最优的初始权重和初始偏移量，因此应用本发明实施例提供的人工神经网络的定位性能能够得到提高。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作一简单地介绍，显而易见地，下面描述中的附图是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为人工神经网络算法的结构示意图；
图2为本发明实施例提供的提高人工神经网络定位性能的方法实施例一的流程图；
图3为本发明实施例提供的提高人工神经网络定位性能的方法实施例二的流程图；
图4为本发明实施例提供的提高人工神经网络定位性能的方法实施例三的流程图；
图5为本发明实施例提供的提高人工神经网络定位性能的方法实施例四的流程图；
图6为本法明实施例提供的提高人工神经网络定位性能的装置实施例一的结构示意图；
图7为本发明实施例提供的提高人工神经网络定位性能的装置实施例二的结构示意图。
具体实施方式
为使本发明实施例的目的、技术方案和优点更加清楚，下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
ANN是一种非常出名的模式匹配算法，应用于多层和多连接的人工神经网络中。人工神经元通过使用激活功能来模拟生物的神经元。每个神经元都有激活功能，该激活功能负责映射神经元的输入到它的输出。神经网络的结构依赖于不同层的人工神经元相互连接的方式。每个神经元有他自己的权重和偏移量。权重和偏移量在训练阶段可以进行调整。这种学习的过程被称作监督下的学习，用来找到一个最佳的输入到输出的映射函数。
当人工神经网络计算出的输出低于一个设定的误差门限后，人工神经网络就结束了他的训练阶段。然后，经过训练的网络可以用来解决模式识别问题。
图1为人工神经网络算法的结构示意图。如图1所示，整个人工神经网络分为输入层、隐藏层和输出层。其中，输入层、隐藏层、输出层的各节点 (也称为神经元)相互之间具有联系，每个节点代表一种特定的函数输出，每两个节点之间的连接都代表一个通过该连接对信号的加权值，即权重。
对人工神经网络进行训练的目的就是找到各层节点对应的函数关系以及权重和偏移量，从而得到输入层节点到输出层节点的非线性的函数关系。
将ANN应用于无线定位技术中时，首先在ANN的训练阶段，需要在一些测量点对指纹信息进行测量。以指纹信息为RSSI为例，在已知地理位置信息(例如经纬度信息)的测量点，测量接收到的各无线接入点(Access Point，AP)的RSSI。将一个测量点的多个RSSI作为人工神经网络的输入神经元，将该测量点的位置信息作为输出神经元，使用多个测量点的RSSI和位置信息对人工神经网络进行训练，直到人工神经网络的输出低于设定的误差门限。这样即完成了人工神经网络的训练过程。
在使用训练完的人工神经网络进行定位时，将待定位点接收到的各AP的RSSI输入训练完的人工神经网络，即可得到带定位点的位置信息，从而完成了定位。
但是，传统的人工神经网络中，在对网络进行训练的过程中，首先需要随机选取各节点(神经元)的初始权重和初始偏移量，然后以初始权重和初始偏移量开始对人工神经网络进行训练。但是随机选取的初始权重和初始偏移量可能会与最优的权重和偏移量差距较大，在应用ANN对网络进行训练时，可能会使权重和偏移量收敛到一个局部最优值。而使用局部最优的权重和偏移量进行无线定位时，将会对定位性能产生影响。
图2为本发明实施例提供的提高人工神经网络定位性能的方法实施例一的流程图，如图2所示，本实施例的方法包括：
步骤S201，获取至少一个测试点的地理位置信息和指纹信息。
具体地，本实施例提供的定位方法基于图1所示的ANN算法。本实施例提供的定位方法包括训练阶段和定位阶段两个阶段。其中训练阶段是根据已知的一些测量点的位置信息和指纹信息对人工神经网络进行训练，得到人工神经网络各层的权重和偏移量。然后在定位阶段，将测量到的指纹信息输入至训练后的人工神经网络即可得到待定位地点的位置信息。
在本步骤中，首先进行定位的训练阶段，为了对人工神经网络进行训练，需要获取至少一个测试点的地理位置信息和指纹信息，测试点的数量越多， 训练得到的人工神经网络性能更优。每个测试点的指纹信息可以是该测试点测量到的任一种能够表征该测试点的特定标记的测量量，例如RSSI等。由于在无线网络中，每个测试点可能接收到多个AP的信号，因此每个测试点的指纹信息可能是由多个测量量组成的。而测试点的地理位置信息可以采用任一种已知的定位方法确定，例如使用全球定位系统(Global Positioning System，GPS)等定位系统确定。测试点的地理位置信息一般为一个经纬度值。
步骤S202，以至少一个测试点的指纹信息作为输入层神经元，以至少一个测试点的地理位置信息作为输出层神经元，随机选取输入层、输出层、隐藏层中每一神经元的初始权值和初始偏移量，建立初始人工神经网络。
具体地，当获取到至少一个测试点的地理位置信息和指纹信息后，需要进行初始人工神经网络的建立。其中，分别将至少一个测试点的指纹信息作为输入层神经元，分别将至少一个测试点的地理位置信息作为输出层神经元。然后随机选取输入层、输出层、隐藏层中每一神经元的初始权值和初始偏移量，建立初始人工神经网络。初始人工神经网络的输入层神经元的数量根据各测试点所测量到的测量量的数量不同而有所区别，初始人工神经网络的输入层神经元的数量为两个，即测试点的地理位置的经度和维度值。
本步骤的初始人工神经网络建立方法和现有的将ANN应用于无线定位技术的方法相同。
步骤S203，使用粒子群优化(Particle Swarm Optimization，PSO)算法对初始人工神经网络每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始人工神经网络。
具体地，为了解决将人工神经网络应用于无线定位技术时，在对人工神经网络进行训练时，可能无法得到最优的权值和偏移量的问题，本实施例将PSO算法应用于人工神经网络的训练阶段，使用PSO算法对随机选取的初始人工神经网络的初始权值和初始偏移量进行优化，从而得到优化后的初始人工神经网络。
PSO算法属于粒子群理论，是在通过模拟简化的社会模型时发现的。它模拟了鸟群和鱼群的特征。在一个粒子群优化中，有一群候选方案成为个体或粒子。同时，这些个体或粒子通过合作或竞争进行演进。
PSO算法的原理是，定义N维空间中的粒子xi＝{x1,x2,...,xN}，各粒子在空间中的飞行速度为Vi＝{V1,V2,...,VN}。每个粒子都有一个由目标函数决定的适应值(fitness value)，并且每个粒子都知道自己到目前为止发现的最好为止(pbest)以及现在位置。每个粒子还知道到目前为止整个粒子群体中所有粒子发现的最好位置(gbest)。每个粒子都追随整个粒子群中的最优粒子在空间中进行搜索，经过多次迭代最终找到整个空间中的最优解(整个空间中的最好位置)。
PSO算法在寻找全局最优解的方面具有优势，因此本实施例将PSO算法应用于人工神经网络中，使用PSO算法去寻找初始人工神经网络的最优初始权重和最优初始偏移量，从而得到最优的初始人工神经网络。
在本实施例中，将初始人工神经网络的初始权重和初始偏移量作为PSO算法的粒子。通过粒子间的竞争和合作，这些粒子能够在搜索空间中找到它们的最佳位置。在PSO算法的每次迭代过程结束后，将各粒子对应的迭代后的权重和偏移量代入人工神经网络中，使用各测试点的指纹信息作为输入，然后计算各测试点经人工神经网络处理后得到的输出和该测试点的实际地理位置信息之间的误差。误差最小权重和偏移量即为本轮PSO算法迭代的最佳位置(gbest)。
将各测试点的上述误差综合在一起即为本次PSO算法迭代过程的总误差。使用PSO算法对初始权重和初始偏移量进行优化的目标是使上述总误差小于一个预设的阈值，若计算的总误差未小于该预设的阈值，则进行PSO算法的下一次计算，否则将得到优化后的初始权重和初始偏移量，从而得到优化后的初始人工神经网络。或者使用PSO算法对初始权重和初始偏移量进行优化的目标是当PSO算法的迭代次数到达预设次数，当迭代次数到达预设次数，则结束优化过程，从而得到优化后的初始人工神经网络。
步骤S204，对优化后的初始人工神经网络进行训练，得到训练后的人工神经网络。
具体地，在得到优化后的初始人工神经网络后，将使用ANN算法对优化后的初始人工神经网络进行训练。对优化后的初始人工神经网络进行训练所使用的ANN算法可以为任一种ANN算法，例如后向传播-ANN(Back-Propagation-ANN，BP-ANN)广义回归线性网络(General Regression  Neural Network，GRNN)等。
基于优化后的初始人工神经网络中各层神经元优化后的初始权重和初始偏移量，各神经元进行相互的学习，最终得到训练后的人工神经网络。从而完成人工神经网络的训练过程。
在完成了人工神经网络的训练过程后，即可使用训练后的人工神经网络进行定位。当得到训练后的人工神经网络后，即相当于得到了从位置点的指纹信息到地理位置信息的映射函数。当需要对网络中的设备进行定位时，首先获取该设备所处待定位点的指纹信息，这里的指纹信息需要与步骤S201中在训练阶段获取的指纹信息是相同的测量量。当获取到设备所处待定位点的指纹信息后，将其作为训练后的人工神经网络的输入层神经元，经训练后的人工神经网络处理后的输入层神经元即为该设备所处待定位点的地理位置信息。从而完成了对带定位点的定位处理。
本实施例提供的提高人工神经网络定位性能的方法，在应用人工神经网络进行无线定位时，当使用随机的初始权重和偏移量建立了初始人工神经网络后，使用PSO算法对初始人工神经网络各层的初始权重和初始偏移量进行优化，得到优化后的初始人工神经网络，对优化后的人工神经网络进行训练得到训练后的人工神经网络，由于使用PSO算法对初始权重和初始偏移量进行优化能够得到全局最优的初始权重和初始偏移量，因此应用本实施例提供的定位方法生成的人工神经网络的定位性能较高。
图3为本发明实施例提供的提高人工神经网络定位性能的方法实施例二的流程图，具体地，本实施例为图2所示实施例中步骤S203的具体执行过程。如图3所示，本实施例的方法包括：
步骤S301，将初始人工神经网络中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立PSO算法的种群。
首先，将初始人工神经网络中每一神经元的初始权重和初始偏移量作为PSO算法中各粒子的初始位置。随机选取各粒子的初始速度，完成PSO算法种群的建立。
具体地，本实施例的方法中，PSO算法结束的目标是迭代代数结束，即迭代代数到达预定的代数。在每一次迭代过程中，可以依次循环执行步骤 S302-S305：
步骤S302，计算PSO算法中每一粒子的适应度。
步骤S303，更新PSO算法的种群中的最佳粒子。
具体地，在得到PSO种群中每一粒子的适应度之后，需要将每一粒子对应的位置代入初始人工神经网络中，作为初始人工神经网络中各神经元的权重和偏移量，然后将各测试点的指纹信息输入该初始人工神经网络，计算该初始人工神经网络的输出与各测试点的地理位置信息之间的误差，将上述误差最小的权重和偏移量作为PSO算法中的最佳粒子。在得到PSO算法中的最佳粒子后，更新PSO算法的种群中的最佳位置(gbest)。
步骤S304，更新PSO算法种群中各粒子位置及其速度。
具体地，在更新了PSO算法种群的最佳位置后，其他所有粒子都向该最佳粒子以一定速度移动，从而得到各粒子的最新位置。然后还要更新每一粒子的速度，得到PSO算法的新的种群。
步骤S305，判断PSO算法的迭代代数是否结束。
具体地，当PSO算法的迭代代数还未结束，则重复执行步骤S302，否则执行步骤S306。
步骤S306，将PSO算法种群中各粒子的最佳位置作为初始人工神经网络每一神经元的优化后权值和优化后偏移量。
具体地，当PSO算法的迭代代数结束，将PSO算法中各粒子的最佳位置(pbest)作为该粒子对应的初始人工神经网络每一神经元的优化后权值和优化后偏移量。
步骤S307，使用初始人工神经网络每一神经元的优化后权值和优化后偏移量建立优化后的初始人工神经网络。
图4为本发明实施例提供的提高人工神经网络定位性能的方法实施例三的流程图，具体地，本实施例为图2所示实施例中步骤S203的另一种具体执行过程。如图4所示，本实施例的方法包括：
步骤S401，将初始人工神经网络中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立PSO算法的种群。
首先，将初始人工神经网络中每一神经元的初始权重和初始偏移量作为 PSO算法中各粒子的初始位置。随机选取各粒子的初始速度，完成PSO算法种群的建立。
具体地，本实施例的方法中，PSO算法结束的目标是所有粒子的适应度之和小于预设阈值。在每一次迭代过程中，可以依次循环执行步骤S402-S405：
步骤S402，计算PSO算法中每一粒子的适应度。
步骤S403，更新PSO算法的种群中的最佳粒子。
具体地，在得到PSO种群中每一粒子的适应度之后，需要将每一粒子对应的位置代入初始人工神经网络中，作为初始人工神经网络中各神经元的权重和偏移量，然后将各测试点的指纹信息输入该初始人工神经网络，计算该初始人工神经网络的输出与各测试点的地理位置信息之间的误差，将上述误差最小的权重和偏移量作为PSO算法中的最佳粒子。在得到PSO算法中的最佳粒子后，更新PSO算法的种群中的最佳位置(gbest)。
步骤S404，更新PSO算法种群中各粒子位置及其速度。
具体地，在更新了PSO算法种群的最佳位置后，其他所有粒子都向该最佳粒子以一定速度移动，从而得到各粒子的最新位置。然后还要更新每一粒子的速度，得到PSO算法的新的种群。
步骤S405，判断PSO算法所有粒子的适应度之和是否小于预设阈值。
具体地，PSO算法中各粒子的适应度表示其与最佳位置的误差，适应度越小则误差越小。因此，当PSO算法的每一粒子的适应度都很小时，则每一粒子都与最佳位置较接近。当PSO算法中所有粒子的适应度之和小于预设阈值时，则经PSO算法优化的各粒子与最佳位置的平均距离将较近，这样使用优化后的各粒子生成的初始人工神经网络中各神经元的优化后权值和优化后偏移量也将是较佳的。因此，当PSO算法的每一次迭代结束后，都计算所有粒子的适应度之和，若PSO算法所有粒子的适应度之和不小于预设阈值，则重复执行步骤S402，否则执行步骤S406。该预设阈值可以根据经验在系统中预设，该预设阈值越小，则PSO算法优化的结果越好，但优化时间将越长。
步骤S406，将PSO算法种群中各粒子的最佳位置作为初始人工神经网络每一神经元的优化后权值和优化后偏移量。
具体地，当PSO算法所有粒子的适应度之和小于预设阈值，将PSO算法中各粒子的最佳位置(pbest)作为该粒子对应的初始人工神经网络每一神经 元的优化后权值和优化后偏移量。
步骤S407，使用初始人工神经网络每一神经元的优化后权值和优化后偏移量建立优化后的初始人工神经网络。
图5为本发明实施例提供的提高人工神经网络定位性能的方法实施例四的流程图，具体地，本实施例为图3所示实施例中步骤S302和步骤S303(或图4所示实施例中步骤S402和步骤S403)的一种具体执行过程。如图5所示，本实施例的方法包括：
步骤S501，使用PSO算法中每一粒子的位置作为初始人工神经网络中的权值和偏移量。
步骤S502，以至少一个测试点的指纹信息作为输入层神经元，计算初始人工神经网络的输出层神经元。
步骤S503，计算输出层神经元和至少一个测试点的地理位置信息的误差，将误差作为粒子的适应度。
步骤S504，将误差最小的粒子作为PSO算法的种群中的最佳粒子。
具体地，本实施例为计算PSO算法中每一粒子的适应度以及选择PSO算法中最佳粒子的具体方法。也就是将测试点的指纹信息和地理位置信息作为PSO算法的优化目标，使用PSO算法对粒子进行优化处理。
图6为本法明实施例提供的提高人工神经网络定位性能的装置实施例一的结构示意图，如图6所示，本实施例的方法包括：
获取模块61，用于获取至少一个测试点的地理位置信息和指纹信息。
初始信息模块62，用于以至少一个测试点的指纹信息作为输入层神经元，以至少一个测试点的地理位置信息作为输出层神经元，随机选取输入层、输出层、隐藏层中每一神经元的初始权值和初始偏移量，建立初始ANN。
优化模块63，用于使用PSO算法对初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN。
训练模块64，用于对优化后的初始ANN进行训练，得到训练后的ANN。
本实施例提供的定位装置用于执行图2所示方法实施例的技术方案，其实现原理和技术效果类似，此处不再赘述。
进一步地，在图6所示实施例中，优化模块63，具体用于将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置， 随机选取各粒子的初始速度，建立所述PSO算法的种群；在所述PSO算法迭代过程的每一代中，依计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法的迭代代数结束；将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
进一步地，在图6所示实施例中，优化模块63，具体用于将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法中所有粒子的适应度之和小于预设阈值；将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
进一步地，在图6所示实施例中，优化模块63，具体用于使用所述PSO算法中每一粒子作为初始ANN中的权值和偏移量；以所述至少一个测试点的指纹信息作为输入层神经元，计算初始ANN的输出层神经元；计算所述输出层神经元和所述至少一个测试点的地理位置信息的误差，将所述误差作为所述粒子的适应度；将误差最小的粒子作为所述PSO算法的种群中的最佳粒子。
进一步地，在图6所示实施例中，人工神经网络为BP-ANN或GRNN。
进一步地，在图6所示实施例中，指纹信息为RSSI。
本发明实施例提供的定位装置可以设置于无线网络中的基站中，也可以是独立于基站的一个额外的网络设备。
图7为本发明实施例提供的提高人工神经网络定位性能的装置实施例二的结构示意图，如图7所示本实施例的定位装置包括：接收器71和处理器72。可选的，该定位装置还可以包括存储器73。其中，接收器71、处理器72和存储器73可以通过系统总线或其他方式相连，图7中以系统总线相连为例；系统总线可以是工业标准结构(Industrial Standard Architecture，ISA) 总线、外部设备互联(Peripheral Component Interconnect，PCI)总线或扩展工业标准结构(Extended Industrial Standard Architecture，EISA)总线等。所述系统总线可以分为地址总线、数据总线、控制总线等。为便于表示，图7中仅用一条线表示，但并不表示仅有一根总线或一种类型的总线。
接收器71，用于获取至少一个测试点的地理位置信息和指纹信息。
处理器72，用于以所述至少一个测试点的指纹信息作为输入层神经元，以所述至少一个测试点的地理位置信息作为输出层神经元，随机选取输入层、输出层、隐藏层中每一神经元的初始权值和初始偏移量，建立初始ANN；使用PSO算法对所述初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN；对所述优化后的初始ANN进行训练，得到训练后的ANN。
存储器73，用于存储接收器71接收到的数据，并提供给数据给处理器72进行处理。
本发明实施例提供的定位装置可以设置于无线网络中的基站中，也可以是独立于基站的一个额外的网络设备。
在图7所示实施例的一种实现方式中，处理器72，具体用于将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法的迭代代数结束；将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
在图7所示实施例的一种实现方式中，处理器72，具体用于将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法中所有粒子的适应度之和小于预设阈值；将所述PSO 算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
在图7所示实施例的一种实现方式中，处理器72，具体用于使用所述PSO算法中每一粒子的位置作为初始ANN中的权值和偏移量；以所述至少一个测试点的指纹信息作为输入层神经元，计算初始ANN的输出层神经元；计算所述输出层神经元和所述至少一个测试点的地理位置信息的误差，将所述误差作为所述粒子的适应度；将误差最小的粒子作为所述PSO算法的种群中的最佳粒子。
在图7所示实施例的一种实现方式中，所述ANN为后向传播ANN或广义回归线性网络。
在图7所示实施例的一种实现方式中，所述指纹信息为RSSI。
本领域普通技术人员可以理解：实现上述各方法实施例的全部或部分步骤可以通过程序指令相关的硬件来完成。前述的程序可以存储于一计算机可读取存储介质中。该程序在执行时，执行包括上述各方法实施例的步骤；而前述的存储介质包括：ROM、RAM、磁碟或者光盘等各种可以存储程序代码的介质。
最后应说明的是：以上各实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述各实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分或者全部技术特征进行等同替换。因此，本发明的保护范围应以权利要求的保护范围为准。

Claims (12)

  1. 一种提高人工神经网络定位性能的方法，其特征在于，包括：
    获取至少一个测试点的地理位置信息和指纹信息；
    以所述至少一个测试点的指纹信息作为输入层神经元，以所述至少一个测试点的地理位置信息作为输出层神经元，随机选取输入层、输出层、隐藏层中每一神经元的初始权值和初始偏移量，建立初始人工神经网络ANN；
    使用粒子群优化PSO算法对所述初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN；
    对所述优化后的初始ANN进行训练，得到训练后的ANN。
  2. 根据权利要求1所述的方法，其特征在于，所述使用PSO算法对所述初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN，包括：
    将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；
    在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法的迭代代数结束；
    将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；
    使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
  3. 根据权利要求1所述的方法，其特征在于，所述使用PSO算法对所述初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN，包括：
    将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；
    在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法中所有粒子的适应度之和小于预设阈值；
    将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；
    使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
  4. 根据权利要求2或3所述的方法，其特征在于，所述计算所述PSO算法中每一粒子的适应度，包括：
    使用所述PSO算法中每一粒子的位置作为初始ANN中的权值和偏移量；
    以所述至少一个测试点的指纹信息作为输入层神经元，计算初始ANN的输出层神经元；
    计算所述输出层神经元和所述至少一个测试点的地理位置信息的误差，将所述误差作为所述粒子的适应度；
    所述更新所述PSO算法的种群中的最佳粒子，包括：
    将误差最小的粒子作为所述PSO算法的种群中的最佳粒子。
  5. 根据权利要求1～4任一项所述的方法，其特征在于，所述ANN为后向传播ANN或广义回归线性网络。
  6. 根据权利要求1～5任一项所述的方法，其特征在于，所述指纹信息为接收的信号强度指示RSSI。
  7. 一种提高人工神经网络定位性能的装置，其特征在于，包括：
    获取模块，用于获取至少一个测试点的地理位置信息和指纹信息；
    初始信息模块，用于以所述至少一个测试点的指纹信息作为输入层神经元，以所述至少一个测试点的地理位置信息作为输出层神经元，随机选取输入层、输出层、隐藏层中每一神经元的初始权值和初始偏移量，建立初始人工神经网络ANN；
    优化模块，用于使用粒子群优化PSO算法对所述初始ANN每一神经元的初始权值和初始偏移量进行优化，得到优化后的初始ANN；
    训练模块，用于对所述优化后的初始ANN进行训练，得到训练后的ANN。
  8. 根据权利要求7所述的提高人工神经网络定位性能的装置，其特征在于，所述优化模块，具体用于将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；在所述PSO算法迭代过程的每一代中，依次计算所述 PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法的迭代代数结束；将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
  9. 根据权利要求7所述的提高人工神经网络定位性能的装置，其特征在于，所述优化模块，具体用于将所述初始ANN中每一神经元的初始权值和初始偏移量作为PSO算法的粒子初始位置，随机选取各粒子的初始速度，建立所述PSO算法的种群；在所述PSO算法迭代过程的每一代中，依次计算所述PSO算法中每一粒子的适应度、更新所述PSO算法的种群中的最佳粒子、更新所述PSO算法种群中各粒子位置及其速度，直到所述PSO算法中所有粒子的适应度之和小于预设阈值；将所述PSO算法种群中各粒子的最佳位置作为所述初始ANN每一神经元的优化后权值和优化后偏移量；使用所述初始ANN每一神经元的优化后权值和优化后偏移量建立所述优化后的初始ANN。
  10. 根据权利要求8或9所述的提高人工神经网络定位性能的装置，其特征在于，所述优化模块，具体用于使用所述PSO算法中每一粒子的位置作为初始ANN中的权值和偏移量；以所述至少一个测试点的指纹信息作为输入层神经元，计算初始ANN的输出层神经元；计算所述输出层神经元和所述至少一个测试点的地理位置信息的误差，将所述误差作为所述粒子的适应度；将误差最小的粒子作为所述PSO算法的种群中的最佳粒子。
  11. 根据权利要求7～10任一项所述的提高人工神经网络定位性能的装置，其特征在于，所述ANN为后向传播ANN或广义回归线性网络。
  12. 根据权利要求7～11任一项所述的提高人工神经网络定位性能的装置，其特征在于，所述指纹信息为接收的信号强度指示RSSI。
PCT/CN2015/079553 2015-05-22 2015-05-22 提高人工神经网络定位性能的方法和装置 WO2016187746A1 (zh)

Priority Applications (5)

Application Number Priority Date Filing Date Title
PCT/CN2015/079553 WO2016187746A1 (zh) 2015-05-22 2015-05-22 提高人工神经网络定位性能的方法和装置
JP2017560712A JP6596516B2 (ja) 2015-05-22 2015-05-22 人工ニューラルネットワークの測位性能を改善するための方法および装置
EP15892831.7A EP3282781A4 (en) 2015-05-22 2015-05-22 Method and device for improving positioning performance of artificial neural network
CN201580069876.4A CN107113764B (zh) 2015-05-22 2015-05-22 提高人工神经网络定位性能的方法和装置
US15/818,979 US20180089566A1 (en) 2015-05-22 2017-11-21 Method and apparatus for positioning of artificial neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2015/079553 WO2016187746A1 (zh) 2015-05-22 2015-05-22 提高人工神经网络定位性能的方法和装置

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US15/818,979 Continuation US20180089566A1 (en) 2015-05-22 2017-11-21 Method and apparatus for positioning of artificial neural network

Publications (1)

Publication Number Publication Date
WO2016187746A1 true WO2016187746A1 (zh) 2016-12-01

Family

ID=57393787

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/079553 WO2016187746A1 (zh) 2015-05-22 2015-05-22 提高人工神经网络定位性能的方法和装置

Country Status (5)

Country Link
US (1) US20180089566A1 (zh)
EP (1) EP3282781A4 (zh)
JP (1) JP6596516B2 (zh)
CN (1) CN107113764B (zh)
WO (1) WO2016187746A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106941663A (zh) * 2017-05-16 2017-07-11 重庆邮电大学 一种融合凸优化与多目标粒子群的uwb定位方法
CN107124761A (zh) * 2017-03-10 2017-09-01 华南理工大学 融合pso和ss‑elm的蜂窝网无线定位方法
CN107566979A (zh) * 2017-08-25 2018-01-09 中山大学深圳研究院 基于共轭梯度与模拟原子跃迁的室内定位方法
CN107862970A (zh) * 2017-11-20 2018-03-30 无锡开放大学 一种用于翻转课堂的教学质量评价模型

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR102535411B1 (ko) * 2017-11-16 2023-05-23 삼성전자주식회사 메트릭 학습 기반의 데이터 분류와 관련된 장치 및 그 방법
CN108537376B (zh) * 2018-04-02 2022-02-11 西安建筑科技大学 基于rs-pso-grnn的埋地油气管道土壤腐蚀预测方法
KR102145717B1 (ko) * 2018-08-28 2020-08-19 (주)제이엠피시스템 머신러닝을 이용한 실내 측위 방법 및 시스템
CN109814066B (zh) * 2019-01-24 2023-08-18 西安电子科技大学 基于神经网络学习的rssi室内定位测距方法、室内定位平台
CN109922432B (zh) * 2019-03-28 2021-07-16 南京邮电大学 无线通信环境下通过优化指纹元素个数的目标定位方法
DE102019110998A1 (de) 2019-04-29 2020-10-29 Tdk Electronics Ag Anordnungen eines Fahrzeug-Positionsbestimmungssystems, System, Verwendung und Verfahren
CN110232444B (zh) * 2019-06-17 2023-03-28 武汉轻工大学 地质监测bp神经网络的优化方法、装置、设备及存储介质
CN110336637A (zh) * 2019-07-15 2019-10-15 北京航空航天大学 一种无人机干扰信号特征选择方法
CN111182460A (zh) * 2020-01-15 2020-05-19 深圳市微埃智能科技有限公司 混合型室内定位方法、装置、计算机设备及存储介质
JP7410510B2 (ja) 2020-02-10 2024-01-10 株式会社グルーヴノーツ 位置推定装置、位置推定方法及び位置推定プログラム
CN114208221A (zh) 2020-07-28 2022-03-18 蜂图志科技控股有限公司 一种定位方法、装置、移动终端、存储介质
CN113188794B (zh) * 2021-04-29 2021-11-23 东北电力大学 一种基于改进pso-bp神经网络齿轮箱故障诊断方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2005032189A1 (en) * 2003-09-26 2005-04-07 Universite Du Quebec En Abitibi-Temiscamingue (Uqat) Method and system for indoor geolocation using an impulse response fingerprinting technique
CN101778399A (zh) * 2010-01-18 2010-07-14 哈尔滨工业大学 基于fcm和最小二乘曲面拟合法的wlan室内优化ann定位方法
CN102932738A (zh) * 2012-10-31 2013-02-13 北京交通大学 一种改进的基于分簇神经网络的室内指纹定位方法
CN104361393A (zh) * 2014-09-06 2015-02-18 华北电力大学 基于粒子群优化算法改进的神经网络模型用于数据预测方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP3165391B2 (ja) * 1996-03-22 2001-05-14 松下電器産業株式会社 移動体無線通信システムとその移動局の位置検出方法
US20130162481A1 (en) * 2009-10-01 2013-06-27 Parviz Parvizi Systems and methods for calibration of indoor geolocation
CN103401625B (zh) * 2013-08-23 2015-04-08 西安电子科技大学 基于粒子群优化算法的协作频谱感知优化方法
CN104636508A (zh) * 2013-11-07 2015-05-20 大连市勘察测绘研究院有限公司 一种基于粒子群优化算法和质心算法的定位方法
CN103945533B (zh) * 2014-05-15 2016-08-31 济南嘉科电子技术有限公司 基于大数据的无线实时位置定位方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2005032189A1 (en) * 2003-09-26 2005-04-07 Universite Du Quebec En Abitibi-Temiscamingue (Uqat) Method and system for indoor geolocation using an impulse response fingerprinting technique
CN101778399A (zh) * 2010-01-18 2010-07-14 哈尔滨工业大学 基于fcm和最小二乘曲面拟合法的wlan室内优化ann定位方法
CN102932738A (zh) * 2012-10-31 2013-02-13 北京交通大学 一种改进的基于分簇神经网络的室内指纹定位方法
CN104361393A (zh) * 2014-09-06 2015-02-18 华北电力大学 基于粒子群优化算法改进的神经网络模型用于数据预测方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3282781A4 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107124761A (zh) * 2017-03-10 2017-09-01 华南理工大学 融合pso和ss‑elm的蜂窝网无线定位方法
CN107124761B (zh) * 2017-03-10 2021-02-19 华南理工大学 融合pso和ss-elm的蜂窝网无线定位方法
CN106941663A (zh) * 2017-05-16 2017-07-11 重庆邮电大学 一种融合凸优化与多目标粒子群的uwb定位方法
CN106941663B (zh) * 2017-05-16 2020-10-20 重庆邮电大学 一种融合凸优化与多目标粒子群的uwb定位方法
CN107566979A (zh) * 2017-08-25 2018-01-09 中山大学深圳研究院 基于共轭梯度与模拟原子跃迁的室内定位方法
CN107862970A (zh) * 2017-11-20 2018-03-30 无锡开放大学 一种用于翻转课堂的教学质量评价模型

Also Published As

Publication number Publication date
JP2018521311A (ja) 2018-08-02
CN107113764A (zh) 2017-08-29
US20180089566A1 (en) 2018-03-29
EP3282781A4 (en) 2018-05-16
JP6596516B2 (ja) 2019-10-23
CN107113764B (zh) 2020-02-14
EP3282781A1 (en) 2018-02-14

Similar Documents

Publication Publication Date Title
WO2016187746A1 (zh) 提高人工神经网络定位性能的方法和装置
KR102116824B1 (ko) 딥러닝 기반 측위 시스템 및 이의 구축 방법
US9239990B2 (en) Hybrid location using pattern recognition of location readings and signal strengths of wireless access points
CN102395192B (zh) 一种无线传感终端的定位方法和装置
US8738032B2 (en) Hybrid location using a weighted average of location readings and signal strengths of wireless access points
Yan et al. An improved multihop-based localization algorithm for wireless sensor network using learning approach
CN105245627B (zh) 一种基于网络坐标系统的ip定位方法
Gupta et al. Study of range free centroid based localization algorithm and its improvement using particle swarm optimization for wireless sensor networks under log normal shadowing
Li et al. Unsupervised learning of indoor localization based on received signal strength
Shit et al. AI-enabled fingerprinting and crowdsource-based vehicle localization for resilient and safe transportation systems
KR102304466B1 (ko) 실내 측위용 데이터베이스 구축, 실내 측위 방법 및 시스템
Zhang et al. Domain adversarial graph convolutional network based on rssi and crowdsensing for indoor localization
Liouane et al. Improved Two Hidden Layers Extreme Learning Machines for Node Localization in Range Free Wireless Sensor Networks.
Vahidnia et al. A hierarchical signal-space partitioning technique for indoor positioning with WLAN to support location-awareness in mobile map services
CN112738709B (zh) 一种基于uwb信号的粒子群搜索室内定位方法及系统
Fahama et al. An Experimental Comparison of RSSI-Based Indoor Localization Techniques Using ZigBee Technology
Mozamir et al. Indoor localization estimation techniques in wireless sensor network: A review
Nguyen Lora localisation in cities with neural networks
CN117241215A (zh) 一种基于图神经网络的无线传感器网络分布式节点协作定位方法
Cao A localization algorithm based on particle swarm optimization and quasi-newton algorithm for wireless sensor networks
CN107197519B (zh) 基于改进的最小二乘支持向量机的水下目标定位方法
Górak et al. Malfunction immune Wi–Fi localisation method
Yan et al. Online detection of Wi-Fi fingerprint alteration strength via deep learning
Liu Research on wsn node localization algorithm based on rssi iterative centroid estimation
CN111194000B (zh) 基于蓝牙融合混合滤波与神经网络的测距方法与系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15892831

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2017560712

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE