CN110264154B - Crowd-sourced signal map construction method based on self-encoder - Google Patents
Crowd-sourced signal map construction method based on self-encoder Download PDFInfo
- Publication number
- CN110264154B CN110264154B CN201910454208.3A CN201910454208A CN110264154B CN 110264154 B CN110264154 B CN 110264154B CN 201910454208 A CN201910454208 A CN 201910454208A CN 110264154 B CN110264154 B CN 110264154B
- Authority
- CN
- China
- Prior art keywords
- signal map
- signal
- signals
- encoder
- self
- Prior art date
- Legal status (The legal status 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 status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/29—Geographical information databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/10—Office automation; Time management
- G06Q10/101—Collaborative creation, e.g. joint development of products or services
Abstract
The invention discloses a crowdsourcing signal map construction method based on a self-encoder, which comprises the following steps: (1) offline training phase: training a model in an offline mode by using an incomplete historical signal map collected in a crowdsourcing mode; (2) an online reconstruction stage: and carrying out missing value inference on the collected incomplete historical signals by using the model completed by offline training so as to reconstruct a complete signal map. Compared with the traditional signal map collecting mode, the method does not need professional on-site measurement; compared with the common crowdsourcing mode, the invention does not need a plurality of crowdsourcing participants; compared with the previous missing value inference method, the method has better reconstruction precision, strong adaptability and high precision.
Description
Technical Field
The invention relates to a crowdsourcing signal map construction method, in particular to a crowdsourcing signal map construction method based on a self-encoder.
Background
Signal maps, which consist of signal strengths at different locations, play an important role in site spectrum monitoring, location Based Services (LBS), network construction and network optimization. For example, a signal map may be used for fingerprint-based indoor location to provide better location services, while a user may use the signal map to learn about current network conditions to select a suitable location to experience better mobile network services. ISPs can use signal maps to learn about network conditions and optimize network architecture to provide better service coverage. Especially in the prospect of 5G networks, the signal map may also be used to analyze the current 4G network to guide the deployment of the 5G network. Despite its importance, building a signal map by a professional's hand-held device is time consuming and laborious. To solve this problem, the chinese patent publication No. CN106157342a proposes an automatic calibration method and system for signal map, which uses a mobile signal receiver to collect signals and construct a signal map, and the method effectively reduces manpower but has high cost and small moving range. With the development of electronic technology, mobile devices are becoming a necessary tool for people, and meanwhile, the mobile devices are also provided with a plurality of sensors, so that the crowdsourcing mode for constructing a signal map becomes possible. However, crowdsourcing often requires a large number of participants to construct a signal map, but users are more passive to this, which results in fewer participants. In order to solve the problem, the Chinese patent with publication number of CN104380294A 'Wi-Fi signal map construction device and method' proposes to construct a database of a user resident location and implicitly collect signals, and the method can effectively solve the problem of a small number of participants, but the construction of the location database involves user privacy and needs frequent updating of the database. While document 1 (Steering Crowdsourced Signal Map Construction via Bayesian Compressive Sensing) proposes to infer an incomplete signal map by means of bayesian compressed sensing to obtain a complete signal map, the method has high calculation cost, certain requirements on the structure of the original signal map and certain limitations on the applicability of the method.
Disclosure of Invention
The invention aims to: the invention aims to provide a crowdsourcing signal map construction method based on a self-encoder, which solves the problem that the crowdsourcing signal map cannot be used due to incomplete collected signal maps under the situation of constructing the signal map in a crowdsourcing mode.
The technical scheme is as follows: the invention provides a crowdsourcing signal map construction method based on a self-encoder, which comprises the following steps:
(1) Offline training stage: training a model in an offline mode by using an incomplete historical signal map collected in a crowdsourcing mode;
(2) And (3) an online reconstruction stage: and carrying out missing value inference on the incomplete historical signals collected at present by using the model completed by offline training so as to reconstruct a complete signal map.
Further, the offline training of step (1) includes the following steps:
(1) Geographic grid division is carried out according to the range of the collected signal map, and historical crowdsourcing signals uploaded in the same time interval are divided into different grids according to GPS coordinates so as to form a plurality of incomplete signal maps at different moments;
(2) Flattening the signal maps at different moments into one-dimensional vectors, forming a historical signal map with behavior moments listed as grid numbers, and inputting the historical signal map into a self-encoder for training;
(3) The network structure, parameters and the model with the minimum test error of the encoder are selected through a cross-validation method.
The online reconstruction of the step (2) comprises the following steps:
(1) Calculating the importance of the signals in different grids to the accuracy of the reconstructed signal map and sequencing;
(2) According to the size of the budget cost, setting the number of required crowdsourcing signals and formulating a corresponding excitation mechanism, and preferentially selecting grids with higher importance to collect signals and upload the signals so as to form an incomplete signal map;
(3) Taking the collected incomplete signal map as input, and inputting the incomplete signal map into a self-encoder model after off-line training for fine adjustment to obtain a complete signal map;
(4) And calculating the reconstruction precision of the collected signals by using a cross-validation method, comparing the reconstruction precision with a preset threshold value, and repeating the process until the threshold value requirement is met if the threshold value requirement is not met.
Further, the method for calculating the importance of the step (1) includes the following steps:
(1) Taking a preset value of a missing value in the signal map as input of a self-encoder for offline training, and outputting the preset value as a default signal map;
(2) Comparing the average value in the historical signal with a default signal map under the condition of no crowdsourcing signal, wherein the difference degree is used as the importance of the next round of signal collection;
(3) And under the condition of collecting a plurality of signals, comparing the two adjacent reconstructed signal maps, and taking the difference degree as the importance of the signal collected in the next round.
Further, in the step (3), when only one round of signals is collected, the signals are compared with a default signal map.
Further, the network structure is the number of hidden layers and the number of neurons in each layer.
Further, the incentive mechanism is a game mechanism or a cash mechanism.
The invention fully utilizes a large number of incomplete historical signal maps in a crowdsourcing mode and utilizes the self-encoder to extract the time characteristics thereof, thereby guiding the collection process of the signal map, and further deducing the missing value so as to effectively reconstruct the complete signal map. The method is different from the traditional signal map construction method, and the method learns the internal rules existing in a large number of historical incomplete signal maps by means of the self-encoder, and deduces the missing value of the incomplete signal map obtained at this time so as to obtain a complete signal map. Therefore, the crowdsourcing signal map construction method of the present invention does not require a prior complete signal map and does not require a large number of crowdsourcing participants.
The beneficial effects are that: compared with the traditional signal map collecting mode, the method does not need professional on-site measurement; compared with the common crowdsourcing mode, the invention does not need a plurality of crowdsourcing participants; compared with the prior missing value inference method, the method has better reconstruction accuracy.
Drawings
FIG. 1 is a system frame diagram of the present invention;
FIG. 2 is a flow chart of the present invention;
fig. 3 is a block diagram of a self-encoder of the present invention.
Detailed Description
As shown in fig. 1, the present embodiment includes two stages, an off-line training stage and an on-line reconstruction stage.
1. The off-line training stage is divided into the following steps (fig. 2):
(1) Determining reasonable geographical grid division according to the range of the collected signal map, adopting simulation data in the embodiment, wherein the simulation space is 24 multiplied by 20m 2 Dividing the grid into 20 multiplied by 20cm 2 Forming 120×100 grids, dividing the simulation signals into different grids according to coordinate ranges, and generating 5000 grids with 120×100 sizeA signal map, which is an incompleteness of the analog history signal map, so that 50% of grid signals are randomly selected from each signal map to be set to-150 (a non-collected signal default value is selected, and the value is only required to be the largest integer smaller than the minimum value of the signal map), and thus the 5000 multiplied by 120 multiplied by 100-dimensional history signal map is finally generated;
(2) The 5000×120×100-dimensional incomplete historical signal map is flattened into 5000×12000-dimensional vectors to form a historical signal map matrix with behavior moments listed as grid numbers. The matrix is then used as input to a self-encoder (fig. 3) for training, the self-encoder structure being as shown in fig. 3, wherein the black circles represent the uncollected signals;
a self-encoder is an unsupervised neural network model that includes two parts: the encoder may learn implicit features present in the input data; the decoder uses the learned implicit features to reconstruct the original input data. In this embodiment, the encoder is composed of a single-layer neural network, i.e., an input layer (incomplete signal pattern) and a single-layer hidden layer; the decoder is also composed of a single-layer neural network, i.e. a single-layer hidden layer and an output layer (reconstructing the complete signal diagram). Wherein the number of neurons in the hidden layer should be no less than the number of known elements in the input data, which can be determined by a cross-validation method.
(3) Selecting a nearest self-encoder model, namely a model with the smallest test error, by adopting a cross validation method of 80-20, wherein 1 hidden layer is arranged, and the loss function is as follows:
where M is the number of rows and Ω is the indication matrix (if the signal corresponding to the grid is collected, its value is 1; otherwise 0), the product of the elements is represented by, I.I F F norm of expression x i Representing the corresponding input row, sigma (E) ,W (E) ,b (E) Representing the activation function (tanh () function), weight matrix and offset value (default random initialization), σ, respectively, in the coding layer (D) ,W (D) ,b (D) Respectively represent an activation function (identity () function), a weight matrix (W (E) And offset values (default random initialization), λ is the super-parameter (selected by the cross-validation method) to produce the best self-encoder model, i, j is the row, column number, respectively.
2. The online reconstruction stage is divided into the following steps (figure 2):
(1) The importance of the signals in the different grids to the accuracy of the reconstructed signal map is calculated and ordered. When the crowdsourcing task starts, the average value in the historical signal is calculated
Wherein x is i，j Representing the ith row, jth column element in the input data. ) Comparing the degree of difference (corresponding to the absolute value of the element difference) with a default signal map (a complete signal map generated by taking the uncollected signal default value as input from the encoder) as the importance of the next round of collected signals, comparing the two adjacent reconstructed signal maps in the case of collecting a certain number of signals, and taking the degree of difference (corresponding to the absolute value of the element difference) as the importance of the next round of collected signals (compared with the default signal map when only one round of signals is collected); (2) According to the size of the budget cost, setting the number of required crowdsourcing signals (in the embodiment, the budget cost is set as the number of crowdsourcing signals, and the default is 500), and formulating a corresponding incentive mechanism (such as games and cash), preferentially selecting grids with higher importance to collect signals and upload the signals, so as to form an incomplete signal map x; (3) Taking the collected incomplete signal map as input, and inputting the incomplete signal map into an off-line trained self-encoder model for fine adjustment to obtain the complete signal map +.>
(4) The reconstruction error of the collected signals is calculated by using a cross-validation method
Comparing with preset threshold, if the threshold is not met, repeating the above process until reaching the threshold requirement,>
representing an average value of each column of elements in the history signal; />
Representing the signal map completed by reconstruction from the encoder. />
Claims (5)
1. A crowdsourcing signal map construction method based on a self-encoder is characterized by comprising the following steps of: the method comprises the following steps:
(1) Offline training stage: training a model in an offline mode by using an incomplete historical signal map collected in a crowdsourcing mode;
(2) And (3) an online reconstruction stage: performing missing value inference on the incomplete historical signals collected at present by using a model completed by offline training so as to reconstruct a complete signal map,
the offline training of the step (1) comprises the following steps:
the method comprises the steps of (1.1) carrying out geographic grid division according to the range of a collected signal map, and dividing historical crowdsourcing signals uploaded in the same time interval into different grids according to GPS coordinates so as to form a plurality of incomplete signal maps at different moments;
(1.2) flattening the signal maps at different moments into one-dimensional vectors, forming a historical signal map with behavior moments listed as grid numbers, and then taking the historical signal map as input, and inputting the historical signal map into a self-encoder for training;
(1.3) selecting a model with minimum test error from the network structure, parameters of the encoder by a cross-validation method,
the online reconstruction of the step (2) comprises the following steps:
(2.1) calculating the importance of the signals in the different grids to the accuracy of the reconstructed signal map and ordering;
(2.2) setting the number of required crowdsourcing signals according to the size of the budget cost, formulating a corresponding excitation mechanism, and preferentially selecting grids with higher importance to collect signals and upload the signals so as to form an incomplete signal map;
(2.3) taking the collected incomplete signal map as input, and inputting the incomplete signal map into a self-encoder model which is subjected to off-line training for fine adjustment to obtain a complete signal map;
and (2.4) calculating the reconstruction precision of the collected signals by using a cross-validation method, comparing the reconstruction precision with a preset threshold value, and if the threshold value requirement is not met, repeating the process until the threshold value requirement is met.
2. The self-encoder based crowd-sourced signal map construction method of claim 1, wherein: the method for calculating the importance of the step (2.1) comprises the following steps:
(a) Taking a preset value of a missing value in the signal map as input of a self-encoder for offline training, and outputting the preset value as a default signal map;
(b) Comparing the average value in the historical signal with a default signal map under the condition of no crowdsourcing signal, wherein the difference degree is used as the importance of the next round of signal collection;
(c) And under the condition of collecting a plurality of signals, comparing the two adjacent reconstructed signal maps, and taking the difference degree as the importance of the signal collected in the next round.
3. The self-encoder based crowd-sourced signal map construction method of claim 2, wherein: when only one round of signals is collected in the step (c), the signals are compared with a default signal map.
4. The self-encoder based crowd-sourced signal map construction method of claim 1, wherein: the network structure is the number of hidden layers and the number of neurons in each layer.
5. The self-encoder based crowd-sourced signal map construction method of claim 1, wherein: the incentive mechanism is a game mechanism or a cash mechanism.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910454208.3A CN110264154B (en) | 2019-05-28 | 2019-05-28 | Crowd-sourced signal map construction method based on self-encoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910454208.3A CN110264154B (en) | 2019-05-28 | 2019-05-28 | Crowd-sourced signal map construction method based on self-encoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110264154A CN110264154A (en) | 2019-09-20 |
| CN110264154B true CN110264154B (en) | 2023-06-09 |
Family
ID=67915724
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910454208.3A Active CN110264154B (en) | 2019-05-28 | 2019-05-28 | Crowd-sourced signal map construction method based on self-encoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110264154B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112261719B (en) * | 2020-07-24 | 2022-02-11 | 大连理智科技有限公司 | Area positioning method combining SLAM technology with deep learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8350758B1 (en) * | 2009-10-01 | 2013-01-08 | Lighthouse Signal Systems LLC | Systems and methods for indoor geolocation based on yield of RF signals |
| CN103745489A (en) * | 2014-01-07 | 2014-04-23 | 电子科技大学 | Method for constructing base station signal field intensity map based on compressed sensing |
| CN105263113A (en) * | 2015-10-26 | 2016-01-20 | 深圳大学 | Wi-Fi location fingerprint map building method and system based on crowd-sourcing |
| CN106714110A (en) * | 2017-01-19 | 2017-05-24 | 深圳大学 | Auto building method and system of Wi-Fi position fingerprint map |
| CN106792465A (en) * | 2016-12-26 | 2017-05-31 | 华中科技大学 | A kind of indoor fingerprint map constructing method based on mass-rent fingerprint |
-
2019
- 2019-05-28 CN CN201910454208.3A patent/CN110264154B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8350758B1 (en) * | 2009-10-01 | 2013-01-08 | Lighthouse Signal Systems LLC | Systems and methods for indoor geolocation based on yield of RF signals |
| CN103745489A (en) * | 2014-01-07 | 2014-04-23 | 电子科技大学 | Method for constructing base station signal field intensity map based on compressed sensing |
| CN105263113A (en) * | 2015-10-26 | 2016-01-20 | 深圳大学 | Wi-Fi location fingerprint map building method and system based on crowd-sourcing |
| CN106792465A (en) * | 2016-12-26 | 2017-05-31 | 华中科技大学 | A kind of indoor fingerprint map constructing method based on mass-rent fingerprint |
| CN106714110A (en) * | 2017-01-19 | 2017-05-24 | 深圳大学 | Auto building method and system of Wi-Fi position fingerprint map |
Non-Patent Citations (2)
| Title |
|---|
| iMap: A Crowdsensing Based System for Outdoor Radio Signal Strength Map;Xiaochen Fan;《2016 IEEE Trustcom-BigDataSE-ISPA》;20170209;第1440-1447页 * |
| 基于图像边缘检测的WLAN室内用户运动地图构建;周牧 等;《电子科技大学学报》;20160430;第45卷(第2期);第240-245页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110264154A (en) | 2019-09-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109492822B (en) | Air pollutant concentration time-space domain correlation prediction method | |
| CN109765053B (en) | Rolling bearing fault diagnosis method using convolutional neural network and kurtosis index | |
| CN110533631A (en) | SAR image change detection based on the twin network of pyramid pondization | |
| CN113632101B (en) | Method for predicting atmospheric pollution through vectorization analysis | |
| CN114449452B (en) | Wi-Fi indoor positioning method based on CNN-RNN | |
| CN112287294B (en) | Space-time bidirectional soil water content interpolation method based on deep learning | |
| CN113177937B (en) | Improved YOLOv 4-tiny-based cloth defect detection method | |
| CN112437451B (en) | Wireless network flow prediction method and device based on generation countermeasure network | |
| CN112362756A (en) | Concrete structure damage monitoring method and system based on deep learning | |
| CN109492756A (en) | More element conductor galloping method for early warning and relevant apparatus based on deep learning | |
| CN105868443B (en) | Construction method of micro-terrain microclimate element field | |
| CN110264154B (en) | Crowd-sourced signal map construction method based on self-encoder | |
| CN113379107A (en) | Regional ionized layer TEC forecasting method based on LSTM and GCN | |
| CN110716998B (en) | Fine scale population data spatialization method | |
| CN116976227B (en) | Storm water increasing forecasting method and system based on LSTM machine learning | |
| Wang et al. | Two-dimension monthly river flow simulation using hierarchical network-copula conditional models | |
| CN117313516A (en) | Fermentation product prediction method based on space-time diagram embedding | |
| CN115659249B (en) | Abnormality detection method for intelligent station measurement control system | |
| CN112529270A (en) | Water flow prediction model based on deep learning | |
| CN113902183B (en) | BERT-based non-invasive transformer area charging pile state monitoring and electricity price adjusting method | |
| JP6950647B2 (en) | Data determination device, method, and program | |
| CN113343590A (en) | Wind speed prediction method and system based on combined model | |
| Liu et al. | Cost-effective signal map crowdsourcing with auto-encoder based active matrix completion | |
| CN112307898A (en) | Data glove action segmentation method and device in virtual reality environment | |
| CN115796000B (en) | Short-term air temperature forecast set correction method based on stacked machine learning algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |