WO2019134549A1 - 基于深度学习的定位方法、装置、计算机设备及存储介质 - Google Patents

基于深度学习的定位方法、装置、计算机设备及存储介质 Download PDF

Info

Publication number
WO2019134549A1
WO2019134549A1 PCT/CN2018/123111 CN2018123111W WO2019134549A1 WO 2019134549 A1 WO2019134549 A1 WO 2019134549A1 CN 2018123111 W CN2018123111 W CN 2018123111W WO 2019134549 A1 WO2019134549 A1 WO 2019134549A1
Authority
WO
WIPO (PCT)
Prior art keywords
cell signal
neural network
point
location
points
Prior art date
Application number
PCT/CN2018/123111
Other languages
English (en)
French (fr)
Inventor
张宗源
王连臣
Original Assignee
中兴通讯股份有限公司
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 中兴通讯股份有限公司 filed Critical 中兴通讯股份有限公司
Publication of WO2019134549A1 publication Critical patent/WO2019134549A1/zh

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
    • G01S11/00—Systems for determining distance or velocity not using reflection or reradiation
    • G01S11/02—Systems for determining distance or velocity not using reflection or reradiation using radio waves
    • 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
    • G01S11/00—Systems for determining distance or velocity not using reflection or reradiation
    • G01S11/02—Systems for determining distance or velocity not using reflection or reradiation using radio waves
    • G01S11/04—Systems for determining distance or velocity not using reflection or reradiation using radio waves using angle measurements
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Definitions

  • the present disclosure relates to, but is not limited to, the field of wireless positioning technology.
  • a depth learning-based positioning method includes: receiving, by a mobile terminal, a plurality of first cell signal parameters to be located, where the first cell signal parameter does not include the mobile terminal Actual latitude and longitude; input the first cell signal parameter reported by the same mobile terminal into a pre-trained cluster model for classification; and input each of the classified first cell signal parameters into the pre-trained feedforward nerve
  • the network current latitude and longitude data corresponding to the first cell signal parameter is obtained; the current latitude and longitude data is input into the pre-trained circulating neural network to obtain latitude and longitude data of the next moment.
  • a depth learning-based positioning apparatus comprising: a receiving module, configured to receive, by a mobile terminal, a plurality of first cell signal parameters to be located, the first cell signal parameter Excluding the actual latitude and longitude of the mobile terminal; the classification module is configured to input the first cell signal parameter reported by the same mobile terminal into the pre-trained cluster model for classification; and the processing module is configured to classify each class The first cell signal parameters are respectively input into the pre-trained feedforward neural network to obtain current latitude and longitude data corresponding to the first cell signal parameter; the processing module is further configured to input the current latitude and longitude data into the pre-trained In the cyclic neural network, the latitude and longitude data of the next moment is obtained.
  • a computer apparatus comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the methods described herein.
  • a computer readable storage medium having stored thereon is a computer program that, when executed by a processor, implements the steps in the methods described herein.
  • FIG. 1 is a flowchart of a depth learning based positioning method, in accordance with one embodiment of the present disclosure
  • FIG. 2 is a schematic diagram of an implementation principle of a depth learning based positioning method according to an embodiment of the present disclosure
  • FIG. 3 is a schematic diagram of data conversion in a positioning process according to an embodiment of the present disclosure.
  • FIG. 4 is a schematic structural diagram of a feedforward neural network according to an embodiment of the present disclosure.
  • FIG. 5 is a schematic structural diagram of a cyclic neural network according to an embodiment of the present disclosure.
  • FIG. 6 is a schematic diagram of training error convergence of a feedforward neural network and a cyclic neural network according to an embodiment of the present disclosure
  • FIG. 7 is a schematic diagram of positioning accuracy verification according to an embodiment of the present disclosure.
  • FIG. 8 is an exemplary structural block diagram of a depth learning based positioning apparatus according to an embodiment of the present disclosure.
  • FIG. 9 is a schematic diagram of the internal structure of a computer device in accordance with one embodiment of the present disclosure.
  • the positioning is mainly performed by using a certain signal indicator received by the terminal, but due to multipath propagation, complex field conditions, measurement errors of the signal itself, etc.
  • the positioning accuracy of the positioning method is not satisfactory, and the positioning system designed by this is not stable enough. Therefore, there is an urgent need for a stable and accurate positioning method to improve the positioning effect.
  • the present disclosure particularly provides a deep learning based positioning method, apparatus, computer device, and storage medium that substantially obviate one or more of the problems due to the limitations and disadvantages of the related techniques. According to the deep learning-based positioning method, device, computer device and storage medium provided by the present disclosure, the stability and accuracy of positioning can be improved.
  • FIG. 1 is a flowchart of a depth learning based positioning method according to an embodiment of the present disclosure
  • FIG. 2 is a schematic diagram of an implementation principle of a depth learning based positioning method according to an embodiment of the present disclosure; Schematic diagram of data conversion in the positioning process of one embodiment
  • FIG. 4 is a schematic structural diagram of a feedforward neural network according to an embodiment of the present disclosure
  • FIG. 5 is a schematic structural diagram of a cyclic neural network according to an embodiment of the present disclosure
  • 6 is a schematic diagram of training error convergence of a feedforward neural network and a cyclic neural network according to an embodiment of the present disclosure
  • FIG. 7 is a schematic diagram of positioning accuracy verification according to an embodiment of the present disclosure.
  • the method includes the following steps S101 to S104.
  • the first cell signal parameters to be located reported by the mobile terminal are received, and the first cell signal parameter does not include the actual latitude and longitude of the mobile terminal.
  • the first cell signal parameter includes Reference Signal Receiving Power (RSRP), Reference Signal Receiving Quality (RSRQ), Angle of Arrival, and Time Advance.
  • RSRP Reference Signal Receiving Power
  • RSRQ Reference Signal Receiving Quality
  • TA Time Advance
  • the first cell signal parameter reported by the same mobile terminal is input into a pre-trained cluster model for classification.
  • the clustering model includes the optimal number of the first cell signal parameters to be divided, and the location points corresponding to the respective first cell signal parameters are divided into each class in various manners, that is, determined.
  • the first cell signal parameter included in each class includes the optimal number of the first cell signal parameters to be divided, and the location points corresponding to the respective first cell signal parameters are divided into each class in various manners, that is, determined. The first cell signal parameter included in each class.
  • the clustering model may classify the first cell signal parameters using a K-Means clustering algorithm.
  • the classified first cell signal parameters of each class are respectively input into the pre-trained feedforward neural network to obtain current latitude and longitude data corresponding to the first cell signal parameter.
  • the trained feedforward neural network is as shown in FIG. 2, and the number of feedforward neural networks is the same as the number of classes divided in step S102, and the clusters obtained by clustering are obtained.
  • the data (for convenience of description, each class divided into a cluster) is input into the corresponding feedforward neural network, and the obtained current latitude and longitude data is the calculated movement corresponding to the first cell signal parameter of each position point.
  • the latitude and longitude data of the terminal is as shown in FIG. 2, and the number of feedforward neural networks is the same as the number of classes divided in step S102, and the clusters obtained by clustering are obtained.
  • the data (for convenience of description, each class divided into a cluster) is input into the corresponding feedforward neural network, and the obtained current latitude and longitude data is the calculated movement corresponding to the first cell signal parameter of each position point.
  • the latitude and longitude data of the terminal is as shown in FIG. 2, and the number of feedforward neural networks is the same as the number of classes divided in step S102, and the cluster
  • the feedforward neural network may use a double-layer BP (Back Propagation) neural network, which is a relatively mature neural network and can be created by determining an optimal K value.
  • BP Back Propagation
  • step S104 the current latitude and longitude data is input into the pre-trained cyclic neural network to obtain latitude and longitude data of the next moment, and the data conversion during the positioning process is as shown in FIG. 3.
  • the trained cyclic neural network is as shown in FIG. 2.
  • the number of the circulating neural network is the same as the number of the feedforward neural network and is corresponding to the feedforward neural network, and is also divided with the position point.
  • the optimal number of classes is the same, and is used to output the latitude and longitude of the next position point corresponding to each position point according to the latitude and longitude of each position point input by the feedforward neural network.
  • the cyclic neural network may use a bidirectional LSTM (Long Short-Term Memory) neural network, which is a relatively mature technology, and can implement the bidirectional LSTM by using a library provided by the Tensorflow tool.
  • LSTM Long Short-Term Memory
  • the neural network, this application is not exhaustive.
  • the method further includes: receiving, by the mobile terminal, a second cell signal parameter of the plurality of location points, where the second cell signal parameter includes an actual latitude and longitude of the mobile terminal; and the same mobile terminal according to the reported time sequence
  • the reported second cell signal parameters are sorted; according to the sorted second cell signal parameters, a clustering algorithm is used to classify the position points corresponding to the second cell signal parameters whose similarity is higher than the preset value;
  • the clustering model is determined by determining the optimal number of classes to which the location points are divided and the location points included in each class.
  • the second cell signal parameter further includes a reference signal received power RSRP, a reference signal received quality RSRQ, an angle of arrival AOA, and a timing advance TA; the method further includes calculating, between the two location points, by using the following formula distance:
  • RSRP 1 represents the reference signal received power of the location point S 1
  • RSRP 2 represents the reference signal received power of the location point S 2
  • RSRQ 1 represents the reference signal reception quality of the location point S 1
  • RSRQ 2 represents the reference of the location point S 2
  • TA 1 represents the timing advance of the position point S 1
  • TA 2 represents the timing advance of the position point S 2
  • AOA 1 represents the arrival angle of the position point S 1
  • AOA 2 represents the arrival angle of the position point S 2
  • D represents the calculated distance between the position point S 1 and the position point S 2 , and the smaller the calculated D value, the higher the similarity between the position point S 1 and the position point S 2 .
  • the method further includes: when receiving the input K value, determining the number of classes in which the location points are divided into K classes; dividing the location points into seed points and other location points, randomly Determining K seed points in the location point, calculating the distance from all other location points to each of the seed points; dividing each of the other location points into a class in which the calculated closest seed point is located, completing the first clustering Calculating an average value of the second cell signal parameters of the position points divided into the same class after the first clustering, the second cell signal parameters including an average value of the reference signal received power, an average value of the reference signal receiving quality, The average of the timing advance and the average of the angle of arrival; the calculated point corresponding to the average of the second cell signal parameters as the new seed point, the class to which the other location points are divided, and the second after the division
  • the average value of the cell signal parameters is iteratively calculated until the calculated average value of the second cell signal parameter does not change or the difference of the change is less than a preset threshold, and the location point is determined
  • the step of determining the optimum K value comprises: receiving a plurality of K enumeration value input, the value of K represents the number of classes in all positions divided points; determining a value wherein K enumeration of I Calculate the average contour coefficient when the position point is divided into K i by the following formula:
  • S(i) represents the calculated average contour coefficient corresponding to K i
  • a(n) represents the average value of the distance from the position point n to all other points in the same class as the position point n
  • b ( n) represents the average of the distance from the position point n to all other position points of the different class with the position point n
  • the K value corresponding to the calculated maximum average contour coefficient is determined as the optimal K value.
  • the step of training the feedforward neural network comprises: constructing a feedforward neural network, constructing a feedforward after determining the optimal number of classes to be divided at the location point and the location points included in each class
  • the number of neural networks is the same as the optimal number of ones, and the feedforward neural network can use a two-layer BP neural network, and ten double-layer BP neural networks can be implemented by using Tensorflow;
  • the second cell signal parameter of the location point is normalized to obtain a feature vector corresponding to each location point; the feature vector of the location point belonging to the same class is used as an input of the corresponding feedforward neural network, and the received
  • the actual latitude and longitude of the position point is used as the output of the corresponding feedforward neural network, and each of the feedforward neural networks is trained until the output error of the feedforward neural network converges to meet the preset first threshold, and the trained feedforward nerve is obtained.
  • the internet The structure of the feedforward neural network according to one embodiment of the present disclosure is as shown in FIG.
  • the second cell signal parameter further includes a reference signal received power RSRP, a reference signal received quality RSRQ, an angle of arrival AOA, and a timing advance TA.
  • the step of normalizing the second cell signal parameters respectively for each type of included location point includes: reporting the reference signal to each cluster (or each of the classified) position points by the following formula
  • the received power RSRP, the reference signal received quality RSRQ, the angle of arrival AOA, and the timing advance TA are normalized:
  • ⁇ is the average value of the corresponding data in the same cluster
  • ⁇ is the corresponding value in the corresponding cluster The standard deviation of the data.
  • each cluster is composed of [RSRP, RSRQ, AOA, TA]
  • the four-dimensional input vector is input into the feedforward neural network corresponding to each cluster (class), and can be trained according to the actual latitude and longitude of the mobile terminal reported by the mobile terminal.
  • the step of training the cyclic neural network comprises: constructing a cyclic neural network, which may use a bidirectional LSTM neural network, which may be implemented by Tensorflow; according to the next point of the position after sorting The actual latitude and longitude of the position point trains the cyclic neural network until the output error of the circulating neural network converges to meet the preset second threshold, and a trained cyclic neural network model is obtained.
  • the structure of a cyclic neural network according to an embodiment of the present disclosure is as shown in FIG.
  • the method further includes: the first cell signal parameter reported by the same mobile terminal according to a sequence of reporting times Perform sorting; input the sorted first cell signal parameters into the cluster model for classification.
  • the trained clustering model since the trained clustering model has determined the optimal K value, that is, the optimal category to be divided, when the first cell signal parameter is classified by the clustering model, the first cell is determined.
  • the optimal number of classes to which the signal parameters are divided, and then according to the clustering model, the location points (or the first cell signal parameters included in each class) included in each class are determined by the similarity degree.
  • the method divides the signal parameters of the first cell.
  • a usage scenario of an offline training includes the following steps 1 to 12.
  • Step 1 Deriving the terminal data of the reported AGPS (Assisted Global Positioning System), including IMSI (International Mobile Subscriber Identification Number), time stamp, RSRP of the primary cell, and primary The RSRQ of the cell, the primary cell TA, the primary cell AOA, and the AGPS reported by the primary cell, these values are spliced by commas, and the data reported at the same time is identified as one piece of data.
  • AGPS Assisted Global Positioning System
  • Step 2 Excluding the data with the reported value as null, using the IMSI as the identifier, traversing all the data, and splicing the same data of the IMSI into a single file in time increment order.
  • Step 3 Implement the K-Means clustering algorithm by using the function provided by Tensorflow.
  • the Tensorflow is an application. The specific method is as follows: the original data obtained in steps 1 and 2 is the position point S (s 1 , s 2 ,... , s n ), where each s i represents a valid data identified by IMSI, and can also be understood as a location point, and is represented by RSRP, RSRQ, TA, AOA values as feature values. Then the K-Means clustering algorithm is used for classification: The purpose of K-means clustering is to divide the raw data into k classes given the number k of classification groups.
  • the method of calculating the distance is the Euclidean distance of each eigenvalue of the data.
  • the following is the formula for calculating the Euclidean distance:
  • the threshold is generally determined by the experimental experience configuration.
  • Step 4 Select one of the user's bill data, input it into the module implemented in step 3, enumerate the k value, take 2 to 10, and each iterate the corresponding steps until the error is less than a specific threshold. It is generally confirmed by experiments to save the data at this time.
  • Step 5 For the data obtained in step 4, traverse the average contour coefficient, and the average contour coefficient formula is Where a(n) is the average of the distances of the n vector to all other points in the cluster to which it belongs, and b(n) is the minimum of the average distance of the n vector to the points of all clusters that are not themselves.
  • the k average contour coefficients are obtained, and the k value corresponding to the value with the largest contour coefficient is selected as the optimal cluster cluster number, and the optimal k value determines the number of neural networks that need to be created in step 6.
  • Step 6 Construct a BP neural network with k double-hidden layer m nodes using the Tensorflow tool, and use the Relu function as the activation function, with the RSRP of the primary cell in the data, the RSRQ of the primary cell, the primary cell TA, the primary cell AOA, and the primary cell.
  • the longitude and the latitude of the primary cell form an input vector.
  • Step 7 normalize the k group data obtained in step 5, and normalize by using the Z-Score method, and the formula is Where x is the original value of the data, ⁇ is the data mean, and ⁇ is the data standard deviation.
  • the statistical dimension here is the cluster, that is, the above values of each cluster are calculated separately.
  • Step 8 Please refer to Figure 4, for the normalized data obtained in step 7, the RSRP column represents x 1 , the RSRQ represents x 2 , and so on, and the k-cluster [primary RSRP, primary RSRQ, TA, AOA] four-dimensional input vector is obtained. .
  • Step 9 Take the data obtained in step 8 as input, report the latitude and longitude as the output, train the BP neural network, select the gradient descent method optimization, and train the n times in steps x until the error convergence meets the threshold.
  • the training error convergence of the feedforward neural network according to one embodiment of the present disclosure is as shown in the curve of the upper portion in FIG.
  • Step 10 Cache the latitude and longitude of the same CDRs in time increments to obtain time series latitude and longitude data of [k, 2] dimension.
  • Step 11 Implement a two-way LSTM neural network using the Tensorflow framework.
  • the neural network is a double hidden layer. The number of specific nodes in each layer is determined by the experimental process parameters.
  • the library provided by Tensorflow is used, and then k BP neural networks are cached.
  • the output in terms of longitude and latitude, form two sets of input vectors, respectively, using the longitude and latitude reported in the data as the output of the neural network to form an output vector.
  • Step 12 Using the data obtained in step 10, the neural network obtained in step 11 is trained, and the ADAM algorithm is selected to be optimized, and the specific step is trained n times until the convergence error satisfies the threshold.
  • the training error convergence of the cyclic neural network according to an embodiment of the present disclosure is as shown in the curve of the lower part in FIG. 6, and the training parameters of the k+1 neural networks created in steps 6 and 10 are saved, and the offline model is obtained. .
  • a usage scenario of online prediction includes the following steps 1 to 6.
  • Step 1 Export the terminal data that fails to report AGPS.
  • the data includes the IMSI, the timestamp, the RSRP of the primary cell, the RSRQ of the primary cell, the primary cell TA, and the primary cell AOA.
  • Step 2 Eliminate the data with the reported value as empty, and splicing all the data with the IMSI as the identifier to form the CDRs of the respective users and sort them by time.
  • Step 3 Classify the user's data into a clustering algorithm module generated by offline training.
  • Step 4 The classified data is input into the k BP neural networks trained in the offline phase step 9.
  • Step 5 Repeat step 4 to buffer the output to form feature data of [m, k] dimensions.
  • m indicates the formation of a time series
  • the output of the buffered BP neural network can be obtained, saved by time, and m can be adjusted according to experiments.
  • Step 6 Input the data obtained in step 5 into the trained LSTM network to finally obtain the latitude and longitude values.
  • This embodiment utilizes the greatest feature of the end user's movement trajectory, that is, continuity. That is to say, the position of the user at a certain moment depends on the location of the user at the previous moment, and the latitude and longitude of the trajectory point can be regarded as a set of sequences strongly related to time.
  • the related positioning technology does not explore the above characteristics, but discrete positioning of the terminal position. At the same time, only a single signal feature is utilized in the positioning process. For example, only RSSI or TA is used, and it is difficult to model a complex and complex environment. .
  • the technical problem to be solved by the present disclosure is to provide a novel positioning method, which can fully utilize the characteristics of the continuity of the moving track and various indicators reported by the terminal, thereby achieving more accurate and effective positioning.
  • the circulatory neural network of the present embodiment introduces an directional loop that can handle the problem of correlation between inputs, i.e., the current output of a sequence is also related to the previous output.
  • the specific form of expression is that the network memorizes the previous information and applies it to the calculation of the current output, that is, the nodes between the hidden layers are no longer connected but connected, and the input of the hidden layer includes not only the output of the input layer. It also includes the output of the hidden layer at the previous moment.
  • the present disclosure introduces a cyclic neural network, which forms a deep learning module with a traditional feedforward neural network, and combines a clustering algorithm to utilize RSRP, RSRQ, TA, AOA, AGPS, cell latitude and longitude data of the primary cell in the data reported by the terminal. According to the following key steps, the positioning of the end user motion trajectory that fails to report AGPS data is realized.
  • the working principle of the key deep learning module in the positioning process in this embodiment is:
  • the k-cluster data output by clustering is trained by the feedforward neural network.
  • the k-group feedforward neural network hyperparameters are output, which is used to describe the latitude and longitude of each cluster and the RSRP and other signal indicators in the reported data. relationship.
  • the clustering algorithm is used to preprocess a large amount of data, so that the correlation between the data in each cluster is higher.
  • the feedforward neural network is used to learn the preprocessed data, which is easier to converge, thus reducing the training time and improving the performance of the model.
  • the cyclic neural network is used to fit the time series, that is, the output of the feedforward neural network in a continuous continuous time is used as the input of the cyclic neural network, and the latitude and longitude of the current time point is trained as the output of the cyclic network.
  • the obtained model can determine the position of the next moment according to the trajectory of the CDR movement.
  • This embodiment mainly consists of an offline training phase and an online prediction phase.
  • the offline training phase mainly includes the following seven steps.
  • the effective report data is obtained, and the IMSI is mainly built, and the CDRs of each user are spliced.
  • k identical multi-layer feedforward neural networks are constructed.
  • the number of hidden layers of the neural network and the number of nodes of each hidden layer are confirmed by the specific experimental tuning process.
  • the k-cluster data set is normalized in step two.
  • the k cluster CDR data is input into each feedforward neural network, and p iterations are performed until the error convergence meets the threshold.
  • the output sequence of the fifth step is buffered to form the [t, m, k] dimension feature data as an input of the cyclic neural network, and the purpose of the step buffer is mainly to form a track sequence.
  • a cyclic neural network is generated, and the data obtained in the sixth step is taken as an input, and the AGPS data finally reported by the terminal is used as an output, and p iterations are performed until the error convergence meets the threshold value, and an offline model is obtained.
  • the online prediction phase can include the following three steps.
  • the AGPS data is not reported, and the IMSI is mainly constructed, and the data of the same IMSI number is associated with the data to be located.
  • the data obtained in the first step is clustered according to the optimal k value obtained by the offline training phase.
  • the k-cluster data is input into an offline model obtained by deep learning to obtain a latitude and longitude output.
  • the labels of the foregoing steps S101-S104 are not used to limit the sequence of the steps in the embodiment, and the numbering of each step is only for the convenience of referring to the labels of the steps in the description of each step. It is to be noted that as long as the order in which the steps are performed does not affect the logical relationship of the embodiment, it is intended to be within the scope of the claimed application.
  • the deep learning-based positioning method provided by the embodiment uses the clustering algorithm to learn a large number of seemingly disordered data, and extracts similar data, and then uses the deep learning framework to perform feature learning on each cluster data, and finally generates offline.
  • Model which can effectively locate terminals that have not reported location information.
  • the model has strong versatility, high anti-interference ability, rapid positioning and positioning accuracy has been greatly improved, reaching 70% within 100 meters.
  • the positioning accuracy verification according to one embodiment of the present disclosure is as shown in FIG.
  • the deep learning-based positioning apparatus 100 may include a receiving module 11, a classification module 12, and a processing module 13.
  • the receiving module 11 is configured to receive a plurality of first cell signal parameters to be located reported by the mobile terminal, where the first cell signal parameter does not include the actual latitude and longitude of the mobile terminal.
  • the classification module 12 is configured to input the first cell signal parameter reported by the same mobile terminal into a pre-trained cluster model for classification.
  • the processing module 13 is configured to input each of the classified first cell signal parameters into the pre-trained feedforward neural network to obtain current latitude and longitude data corresponding to the first cell signal parameter.
  • the processing module 13 is further configured to input the current latitude and longitude data into the pre-trained cyclic neural network to obtain latitude and longitude data of the next moment.
  • the receiving module 11 is further configured to receive a second cell signal parameter of the several location points reported by the mobile terminal, where the second cell signal parameter includes an actual latitude and longitude of the mobile terminal.
  • the deep learning-based positioning apparatus 100 may further include: a sorting module configured to sort the second cell signal parameters reported by the same mobile terminal according to the reported time sequence; the clustering module is configured to be according to the sorted The second cell signal parameter is used to classify the location points corresponding to the second cell signal parameters whose similarity is higher than the preset value by using a clustering algorithm; the clustering module is further configured to determine the optimal division by the location point.
  • the cluster model is determined by the number of classes and the location points contained in each class.
  • the second cell signal parameter further includes a reference signal received power RSRP, a reference signal received quality RSRQ, an angle of arrival AOA, and a timing advance TA;
  • the depth learning based positioning apparatus 100 may further include a calculation module, where The calculation module calculates the distance between two position points by the following formula:
  • RSRP 1 represents the reference signal received power of the location point S 1
  • RSRP 2 represents the reference signal received power of the location point S 2
  • RSRQ 1 represents the reference signal reception quality of the location point S 1
  • RSRQ 2 represents the reference of the location point S 2
  • TA 1 represents the timing advance of the position point S 1
  • TA 2 represents the timing advance of the position point S 2
  • AOA 1 represents the arrival angle of the position point S 1
  • AOA 2 represents the arrival angle of the position point S 2
  • D represents the calculated distance between the position point S 1 and the position point S 2 , and the smaller the calculated D value, the higher the similarity between the position point S 1 and the position point S 2 .
  • the clustering module is further configured to: when receiving the input K value, determine the number of classes in which the location points are divided into K classes; the calculation module is further configured to divide the location point into seed points. And other location points, randomly determining K seed points in the location point, calculating distances of all other location points to each of the seed points; the clustering module is further configured to divide each of the other location points into calculated distances
  • the first cluster is completed in the class in which the recent seed point is located;
  • the calculation module is further configured to calculate an average value of the second cell signal parameters of the location points divided into the same class after the first cluster, the first
  • the second cell signal parameter includes an average value of the reference signal received power, an average value of the reference signal received quality, an average of the time advance amount, and an average value of the angle of arrival; the calculation module is further configured to calculate the second cell signal parameter
  • the position point corresponding to the average value is used as a new seed point, and the class of the other position points and the average value of the signal parameters of the second cell after the division
  • the clustering module may further include: a receiving unit configured to receive the input of the plurality of enumerated K values, the K value indicating the number of classes in which all the location points are divided; the calculating module is further configured to wherein a determination enumeration value K i is calculated by the following equation into this position is the point average profile of the coefficient K i categories:
  • S(i) represents the calculated average contour coefficient corresponding to K i
  • a(n) represents the average value of the distance from the position point n to all other points in the same class as the position point n
  • b ( n) represents an average of the distance from the position point n to all other position points of the different class with the position point n
  • the calculation module is further configured to traverse all of the K values enumerated, and calculate An average contour coefficient corresponding to the K value
  • the clustering module is further configured to determine a K value corresponding to the calculated maximum average contour coefficient as the optimal K value.
  • the deep learning-based positioning apparatus 100 may further include: a feedforward neural network construction module configured to determine an optimal number of classes to be divided at the location point and a location point included in each class After that, a feedforward neural network is constructed, and the number of the feedforward neural network is the same and one-to-one correspondence with the optimal class; the processing module is configured to respectively perform the second cell signal parameter for each type of included location point.
  • a feedforward neural network construction module configured to determine an optimal number of classes to be divided at the location point and a location point included in each class After that, a feedforward neural network is constructed, and the number of the feedforward neural network is the same and one-to-one correspondence with the optimal class
  • the processing module is configured to respectively perform the second cell signal parameter for each type of included location point.
  • a first training module configured to use a feature vector of a location point belonging to the same class as an input of a corresponding feedforward neural network, the location to be received The actual latitude and longitude of the point is used as the output of the corresponding feedforward neural network, and each of the feedforward neural networks is trained until the output error of the feedforward neural network converges to meet the preset first threshold, and the trained feedforward neural network is obtained.
  • the deep learning-based positioning apparatus 100 may further include: a cyclic neural network construction module configured to construct a cyclic neural network; and a second training module configured to be according to the next position of the position point after sorting The actual latitude and longitude of the point trains the cyclic neural network until the output error of the circulating neural network converges to meet the preset second threshold, and a trained cyclic neural network model is obtained.
  • a cyclic neural network construction module configured to construct a cyclic neural network
  • a second training module configured to be according to the next position of the position point after sorting The actual latitude and longitude of the point trains the cyclic neural network until the output error of the circulating neural network converges to meet the preset second threshold, and a trained cyclic neural network model is obtained.
  • the deep learning-based positioning apparatus 100 may further include: a sorting module configured to sort the first cell signal parameters reported by the same mobile terminal according to a sequence of reporting times; the classification module further The method is configured to input the sorted first cell signal parameters into the cluster model for classification.
  • each module included in the depth learning-based positioning apparatus may be implemented in whole or in part by software, hardware, or a combination thereof. Further, each module in the depth learning based positioning device may be a program segment for implementing a corresponding function.
  • the above-described depth learning based positioning device can be implemented in the form of a computer program that can be run on a computer device as shown in FIG.
  • a computer apparatus includes a memory, a processor, and a computer program stored on the memory and operable on the processor, and the processor implements the above-described deep learning-based positioning method when the program is executed .
  • FIG. 9 is a schematic diagram of an internal structure of a computer device, which may be a server, according to an embodiment of the present disclosure.
  • the computer device includes a processor, memory, input device, display screen, and network interface connected by a system bus.
  • the memory includes a non-volatile storage medium and an internal memory, the non-volatile storage medium of the computer device storing an operating system and computer readable instructions that, when executed, cause the processor to perform the present application
  • the processor of the computer device is configured to provide computing and control capabilities to support operation of the entire computer device.
  • the internal memory can store computer readable instructions that, when executed by the processor, cause the processor to perform a deep learning based positioning method.
  • the input device of the computer device is configured for input of various parameters
  • the display screen of the computer device is configured for display
  • the network interface of the computer device is configured for network communication. It will be understood by those skilled in the art that the structure shown in FIG. 9 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation of the computer device to which the solution of the present application is applied.
  • the computer device may include a ratio. More or fewer components are shown in the figures, or some components are combined, or have different component arrangements.
  • the memory in this embodiment can be configured to store software programs as well as various data.
  • the memory may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application required for at least one function, and the like; the storage data area may store data created according to usage of the mobile phone, and the like.
  • the memory may include a high speed random access memory, and may also include a nonvolatile memory such as at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device.
  • the embodiment further provides a computer readable storage medium, on which a computer program is stored, and when the program is executed by the processor, each step in the above-described deep learning-based positioning method is implemented.
  • all or part of the processes in the foregoing embodiment may be completed by a computer program to instruct related hardware, and the program may be stored in a computer readable storage medium, as embodied in the present disclosure.
  • the program can be stored in a storage medium of a computer system and executed by at least one processor in the computer system to implement a process comprising an embodiment of the methods described above.
  • the storage medium includes, but is not limited to, a magnetic disk, a USB flash drive, an optical disk, a read-only memory (ROM), and the like.
  • the depth learning-based positioning method, device, computer device and storage medium respectively input the first cell signal parameters into the training according to the plurality of first cell signal parameters to be located reported by the mobile terminal.
  • a feedforward neural network and a cyclic neural network the latitude and longitude of the current moment of the mobile terminal and the latitude and longitude of the next moment can be obtained, because the signal parameters of the first cell include multiple parameter indicators reported, so that the The positioning information obtained by the first cell signal parameter is more accurate, and since the cyclic neural network has a position prediction function, the latitude and longitude of the next moment can be obtained in real time, so that the stability of the positioning function is stronger.

Landscapes

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

Abstract

一种基于深度学习的定位方法，包括：接收移动终端上报的若干个待定位的第一小区信号参数，第一小区信号参数不包括移动终端的实际经纬度（S101）；将同一移动终端上报的第一小区信号参数输入至预先训练好的聚类模型中进行分类（S102）；将分类后的每类第一小区信号参数分别输入到预先训练好的前馈神经网络中，得到与第一小区信号参数相对应的当前经纬度数据（S103）；将当前经纬度数据输入到预先训练好的循环神经网络中，得到下一时刻的经纬度数据（S104）。还涉及一种基于深度学习的定位装置、计算机设备及存储介质。

Description

基于深度学习的定位方法、装置、计算机设备及存储介质 技术领域
本公开涉及但不限于无线定位技术领域。
背景技术
近年来，无论是基于移动蜂窝系统的网络连接还是WIFI的网络连接，无线定位服务逐渐成为信息服务的热点，各种定位技术层出不穷，比如基于RSSI(Received Signal Strength Indication，接收的信号强度指示)的定位方法、基于TDOA(Time difference of Arrival，到达时间差)的定位方法以及基于AOA(Angle of Arrival，到达角)的定位方法等。
发明内容
根据本公开的一个方面，提供的一种基于深度学习的定位方法，该方法包括：接收移动终端上报的若干个待定位的第一小区信号参数，该第一小区信号参数不包括该移动终端的实际经纬度；将同一移动终端上报的该第一小区信号参数输入至预先训练好的聚类模型中进行分类；将分类后的每类该第一小区信号参数分别输入到预先训练好的前馈神经网络中，得到与该第一小区信号参数相对应的当前经纬度数据；将该当前经纬度数据输入到预先训练好的循环神经网络中，得到下一时刻的经纬度数据。
根据本公开的另一个方面，提供的一种基于深度学习的定位装置，该装置包括：接收模块，用于接收移动终端上报的若干个待定位的第一小区信号参数，该第一小区信号参数不包括该移动终端的实际经纬度；分类模块，用于将同一移动终端上报的该第一小区信号参数输入至预先训练好的聚类模型中进行分类；处理模块，用于将分类后的每类该第一小区信号参数分别输入到预先训练好的前馈神经网络中，得到与该第一小区信号参数相对应的当前经纬度数据；该处理模块还用于将该当前经纬度数据输入到预先训练好的循环神经网络中， 得到下一时刻的经纬度数据。
根据本公开的又一个方面，提供的一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，该处理器执行该程序时实现本文所述的方法。
根据本公开的再一个方面，提供的一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现本文所述的方法中的步骤。
附图说明
图1为根据本公开的一个实施例的基于深度学习的定位方法的流程图；
图2为根据本公开的一个实施例的基于深度学习的定位方法的实现原理示意图；
图3为根据本公开的一个实施例的定位过程中的数据转化示意图；
图4为根据本公开的一个实施例的前馈神经网络的结构示意图；
图5为根据本公开的一个实施例的循环神经网络的结构示意图；
图6为根据本公开的一个实施例的前馈神经网络及循环神经网络的训练误差收敛示意图；
图7为根据本公开的一个实施例的定位精度验证示意图；
图8为根据本公开的一个实施例的基于深度学习的定位装置的示范性结构框图；
图9为根据本公开的一个实施例的计算机设备的内部结构示意图。
本公开目的的实现、功能特点及优点将结合实施例，参照附图做进一步说明。
具体实施方式
为了使本公开所要解决的技术问题、技术方案及有益效果更加清楚、明白，以下结合附图和实施例，对本公开进行进一步详细说明。 应当理解，此处所描述的具体实施例仅仅用以解释本公开，并不用于限定本公开。
基于RSSI定位方法、基于TDOA的定位方法以及基于AOA的定位方法等，主要利用终端接收到的某一种信号指标进行定位，但是由于多径传播、复杂的现场状况、信号本身测量误差等，这些定位方法的定位精度并不能令人满意，由此设计的定位系统表现也不够稳定。所以，迫切需要一种稳定、精确的定位方法，以提高定位效果。
因此，本公开特别提供了基于深度学习的定位方法、装置、计算机设备及存储介质，其实质上避免了由于相关技术的局限和缺点所导致的问题中的一个或多个。根据本公开提供的基于深度学习的定位方法、装置、计算机设备及存储介质，可以提高定位的稳定性及精确性。
图1为根据本公开的一个实施例的基于深度学习的定位方法的流程图；图2为根据本公开的一个实施例的基于深度学习的定位方法的实现原理示意图；图3为根据本公开的一个实施例的定位过程中的数据转化示意图；图4为根据本公开的一个实施例的前馈神经网络的结构示意图；图5为根据本公开的一个实施例的循环神经网络的结构示意图；图6为根据本公开的一个实施例的前馈神经网络及循环神经网络的训练误差收敛示意图；图7为根据本公开的一个实施例的定位精度验证示意图。下面，结合图1至图7来详细描述根据本公开的一个实施例的基于深度学习的定位方法，该方法应用于计算机设备或服务器，用于对移动终端进行定位，如图1所示，该方法包括以下步骤S101至S104。
在步骤S101处，接收移动终端上报的若干个待定位的第一小区信号参数，该第一小区信号参数不包括该移动终端的实际经纬度。
在一个实施例中，该第一小区信号参数包括参考信号接收功率RSRP(Reference Signal Receiving Power)、参考信号接收质量RSRQ(Reference Signal Receiving Qual ity)、到达角AOA(Angle of Arrival)及时间提前量TA(Time Advance)。
在步骤S102处，将同一移动终端上报的该第一小区信号参数输 入至预先训练好的聚类模型中进行分类。
在该实施例中，聚类模型包括了该第一小区信号参数被划分的最优类数，以及以各种方式将各个第一小区信号参数对应的位置点划分到每一类中，即确定每一类所包含的第一小区信号参数。
在一个实施例中，该聚类模型可以采用K-Means聚类算法对该第一小区信号参数进行分类。
在步骤S103处，将分类后的每类该第一小区信号参数分别输入到预先训练好的前馈神经网络中，得到与该第一小区信号参数相对应的当前经纬度数据。
在该实施例中，该训练好的前馈神经网络如图2所示，前馈神经网络的个数与步骤S102中所划分的类数相同且一一对应，通过将聚类得到的各簇数据(为了方便说明，被划分的每一类称为一簇)输入到对应的前馈神经网络中，得到的当前经纬度数据为计算出来的与各个位置点的第一小区信号参数相对应的移动终端的经纬度数据。
在一个实施例中，该前馈神经网络可以选用双层BP(Back Propagation,反向传播)神经网络，BP神经网络是目前比较成熟的神经网络，通过确定的最优K值即可进行创建，本申请就不过多赘述了。
在步骤S104处，将该当前经纬度数据输入到预先训练好的循环神经网络中，得到下一时刻的经纬度数据，其在定位过程中的数据转化如图3所示。
在该实施例中，该训练好的循环神经网络如图2所示，循环神经网络的个数与前馈神经网络的个数相同并与前馈神经网络一一对应，也与位置点被划分的最优类数相同，用于根据前馈神经网络输入的各个位置点的经纬度，输出与各个位置点对应的下一个位置点的经纬度。
在一个实施例中，该循环神经网络可以选用双向LSTM(Long Short-Term Memory,长短期记忆)神经网络，该LSTM神经网络是目前比较成熟的技术，可以利用Tensorflow工具提供的库实现该双向LSTM神经网络，本申请就不过多赘述了。
在一个实施例中，该方法还包括：接收移动终端上报的若干个位置点的第二小区信号参数，该第二小区信号参数包括该移动终端的实际经纬度；按照上报的时间顺序对同一移动终端上报的该第二小区信号参数进行排序；根据排序后的该第二小区信号参数，利用聚类算法将相似度高于预设值的第二小区信号参数对应的位置点归为一类；以及通过确定该位置点被划分的最优类数以及每类所包含的位置点，确定该聚类模型。
在一个实施例中，该第二小区信号参数还包括参考信号接收功率RSRP、参考信号接收质量RSRQ、到达角AOA及时间提前量TA；该方法还包括通过以下公式计算两个位置点之间的距离：
Figure PCTCN2018123111-appb-000001
其中，RSRP 1表示位置点S 1的参考信号接收功率，RSRP 2表示位置点S 2的参考信号接收功率，RSRQ 1表示位置点S 1的参考信号接收质量，RSRQ 2表示位置点S 2的参考信号接收质量，TA 1表示位置点S 1的时间提前量，TA 2表示位置点S 2的时间提前量，AOA 1表示位置点S 1的到达角，AOA 2表示位置点S 2的到达角，D表示计算出的所述位置点S 1与位置点S 2的距离，计算出的D值越小表示位置点S 1与位置点S 2的相似度越高。
在一个实施例中，该方法还包括：接收到输入的K值时，将所有该位置点被划分的类数确定为K类；将该位置点分为种子点及其它位置点，随机在该位置点中确定K个种子点，计算所有其他位置点到每个该种子点的距离；将每个该其他位置点划分到计算的距离最近的种子点所在的类中，完成第一次聚类；计算第一次聚类后被划分在同一类中的位置点的第二小区信号参数的平均值，该第二小区信号参数包括参考信号接收功率的平均值、参考信号接收质量的平均值、时间提前量的平均值及到达角的平均值；以计算的该第二小区信号参数的平均值对应的位置点作为新的种子点，对该其他位置点被划分的类以及划分后该第二小区信号参数的平均值进行迭代计算，直至计算的该 第二小区信号参数的平均值不再变化或变化的差值小于预设的门限时，确定该位置点被划分为K类时每类所包含的位置点，以得到该聚类模型。
在一个实施例中，确定该最优K值的步骤包括：接收输入的若干个枚举的K值，该K值表示所有位置点被划分的类数；确定枚举的其中一个K i的值，通过以下公式计算该位置点被分为K i类时的平均轮廓系数：
Figure PCTCN2018123111-appb-000002
其中，S(i)表示计算的与K i对应的平均轮廓系数，a(n)表示位置点n到所有与所述位置点n划分在同一类的其它位置点的距离的平均值，b(n)表示位置点n到所有与所述位置点n划分在不同类的其它位置点的距离的平均值；遍历枚举的所有所述K值，计算与每个所述K值对应的平均轮廓系数；将与计算的最大平均轮廓系数对应的K值确定为所述最优K值。
在一个实施例中，训练该前馈神经网络的步骤包括：在确定该位置点被划分的最优类数以及每类所包含的位置点的步骤之后，构建前馈神经网络，构建的前馈神经网络的个数与该最优类数相同且一一对应，其中，该前馈神经网络可以选用双层BP神经网络，可以利用Tensorflow实现k个双层BP神经网络；分别对每类所包含的位置点的第二小区信号参数进行归一化处理，得到与每个位置点相对应的特征向量；将属于同一类的位置点的特征向量作为对应前馈神经网络的输入，将接收的该位置点的实际经纬度作为对应前馈神经网络的输出，对每个该前馈神经网络进行训练，直至该前馈神经网络的输出误差收敛满足预设的第一门限，得到训练好的前馈神经网络。根据本公开的一个实施例的前馈神经网络的结构如图4所示。
在一个实施例中，该第二小区信号参数还包括参考信号接收功率RSRP、参考信号接收质量RSRQ、到达角AOA及时间提前量TA。上述分别对每类所包含的位置点的第二小区信号参数进行归一化处理的步骤包括：通过以下公式对每一簇(或者说是被划分的每一类)位 置点上报的该参考信号接收功率RSRP、参考信号接收质量RSRQ、到达角AOA及时间提前量TA进行归一化处理：
Figure PCTCN2018123111-appb-000003
其中，x为数据(参考信号接收功率RSRP、参考信号接收质量RSRQ、到达角AOA或时间提前量TA原始值)的原始值，μ为同一簇中对应数据的平均值，σ为对应簇中对应数据的标准差。
对每一簇位置点上报的第二小区信号参数中包含的上述四种数据进行归一化处理之后，得到K簇输入向量，且每一簇为由[RSRP、RSRQ、AOA、TA]组成的四维输入向量，将这K簇四维输入向量分别输入到与每一簇(类)相对应的前馈神经网络中，并根据移动终端上报的该移动终端的实际经纬度即可对其进行训练。
在一个实施例中，训练该循环神经网络的步骤包括：构建循环神经网络，该循环神经网络可以选用双向LSTM神经网络，可以利用Tensorflow实现该双向LSTM神经网络；根据排序后该位置点的下一位置点的实际经纬度对该循环神经网络进行训练，直至该循环神经网络的输出误差收敛满足预设的第二门限，得到训练好的循环神经网络模型。根据本公开的一个实施例的循环神经网络的结构如图5所示。
在一个实施例中，在接收移动终端上报的若干个待定位的第一小区信号参数的步骤之后，该方法还包括：按照上报时刻的先后顺序，对同一移动终端上报的该第一小区信号参数进行排序；将排序后的第一小区信号参数输入至该聚类模型中进行分类。
在该实施例中，由于训练好的聚类模型已经确定了最优K值，即被划分的最优类别，通过聚类模型对第一小区信号参数进行分类时，即确定了该第一小区信号参数被划分的最优类数，再根据聚类模型中确定每一类所包含的位置点(或者说是每一类所包含的第一小区信号参数)的按相似度进行位置点划分的办法，对该第一小区信号参数进行划分。
例如，根据本实施例的一个离线训练的使用场景包括以下步骤1至步骤12。
步骤1：导出上报的AGPS(Assisted Global Positioning System，辅助全球卫星定位系统)的终端数据，数据中包括IMSI(International Mobile Subscriber Identification Number，国际移动用户识别码)、时间戳、主小区的RSRP、主小区的RSRQ、主小区TA、主小区AOA、主小区上报的AGPS，这些值以逗号拼接，同一时刻上报的数据标识为一条数据。
步骤2：剔除上报值为空的数据，以IMSI为标识，遍历所有数据，把IMSI相同的数据按时间递增顺序拼接为单个文件保存。
步骤3：利用Tensorflow提供的函数实现K-Means聚类算法，其中，Tensorflow为一种应用程序，具体方法为：假设步骤1、2得到的原始数据为位置点S(s 1,s 2,…,s n)，其中每个s i代表一条以IMSI为标识的有效的数据，也可以理解为一个位置点，并且以RSRP、RSRQ、TA、AOA值作为特征值进行表示。然后利用K-Means聚类算法进行分类：K-means聚类的目的是，在给定分类组数k的条件下，将原始数据分成k类。随机在数据中选取k个种子点，然后对剩余的所有数据求到这k个种子点的距离，计算距离的方法是对数据的各个特征值的欧式距离，以下是欧式距离的计算公式：
Figure PCTCN2018123111-appb-000004
距离值越小表示两个数据的相似度越高；
对于本例来说，假设两条数据分别为s 1(RSRP 1，RSRQ 1，TA 1，AOA 1)，s 2(RSRP 2，RSRQ 2，TA 2，AOA 2)，则它们之间的欧式距离为：
Figure PCTCN2018123111-appb-000005
按相似程度(距离)对数据分完组后，分别计算k分组中的四个数据的均值，并以这K个均值作为新的质心(或者说是种子点)开始下一次迭代，重复该步骤，直至各个分组中的数据均值不再变化或变化差值小于一定的门限，该门限一般由实验经验配置确定。
步骤4：任选其中某一用户的话单数据，将其输入步骤3实现的模块，枚举k值，取2到10，各自迭代相应的步骤，直至误差小于特定门限，此门限值可配，一般由实验确认，保存此时的数据。
步骤5：对步骤4中得到的数据，遍历计算平均轮廓系数，平均 轮廓系数公式为
Figure PCTCN2018123111-appb-000006
其中a(n)为n向量到所有它属于的簇中其它点的距离的平均值，b(n)为n向量到所有非本身所在簇的点的平均距离的最小值。得到k个平均轮廓系数，选择轮廓系数最大的值对应的k值作为最优聚类集群数，最优k值决定了步骤6中需要创建的神经网络个数。
步骤6：利用Tensorflow工具构造k个双隐层m节点的BP神经网络，使用Relu函数作为激活函数，以数据中主小区的RSRP、主小区的RSRQ、主小区TA、主小区AOA、主小区的经度、主小区的纬度形成输入向量。
步骤7：对步骤5得到的k组数据进行归一化处理，利用Z-Score方法进行归一化，公式为
Figure PCTCN2018123111-appb-000007
其中x为数据原始值，μ为数据均值，σ为数据标准差，这里的统计维度为簇，即分别计算各簇的上述值。
步骤8：请见附图4，对步骤7得到的归一化数据，RSRP列代表x 1，RSRQ代表x 2，依次类推，得到k簇[主RSRP，主RSRQ，TA，AOA]四维输入向量。
步骤9：将步骤8得到的数据作为输入，上报的经纬度作为输出，对BP神经网络进行训练，选择梯度下降法优化，以步长x，训练n次直至误差收敛满足门限。根据本公开的一个实施例的前馈神经网络的训练误差收敛如图6中位于上部分的曲线所示。
步骤10：对同一话单按时间递增上报的经纬度进行缓存，得到[k,2]维的时间序列经纬度数据。
步骤11：利用Tensorflow框架实现一个双向LSTM神经网络，该神经网络为双隐层，每层具体节点数由实验过程调参确定，此处利用的是Tensorflow提供的库，然后缓存k个BP神经网络输出，分别以经度、纬度为维度，形成两组输入向量，分别以数据中上报的经度、纬度作为神经网络的输出，形成输出向量。
步骤12：利用步骤10得到的数据对步骤11得到的神经网络进 行训练，选择ADAM算法优化，以特定步长训练n次直至收敛误差满足门限值。根据本公开的一个实施例的循环神经网络的训练误差收敛如图6中位于下部分的曲线所示所示，保存步骤6和步骤10创建的k+1个神经网络的训练参数，得到离线模型。
例如，根据本实施例的一个在线预测的使用场景包括以下步骤1至步骤6。
步骤1：导出未能上报AGPS的终端数据，数据中包括IMSI、时间戳、主小区的RSRP、主小区的RSRQ、主小区TA、主小区AOA。
步骤2：剔除上报值为空的数据，以IMSI为标识对所有数据进行拼接，形成各自用户的话单，并按时间排序。
步骤3：将用户的数据输入离线训练生成的聚类算法模块进行分类。
步骤4：分类后的数据输入离线阶段步骤9中训练好的k个BP神经网络。
步骤5：重复步骤4，缓存输出，形成[m,k]维的特征数据。其中，m表示形成时间序列，缓存BP神经网络的输出就可以得到，按时间保存下来，m可以根据实验进行调整。
步骤6：将步骤5得到的数据输入训练好的LSTM网络，最终得到经纬度值。
本实施例利用终端用户移动轨迹的最大特点，也就是连续性。也就是说，某个时刻的用户位置取决于上一时刻用户所在位置，轨迹点的经纬度可以被认为是与时间强相关的一组序列。
而相关定位技术没有发掘上述特性，只是对终端位置进行离散的定位，同时，在定位过程中仅仅利用了单一的信号特征，比如只是用到了RSSI或者TA，很难对现实复杂的环境进行建模。
本公开所要解决的技术问题在于，提供一种新型的定位方法，可以充分利用移动轨迹连续性的特点和终端上报的各种指标，实现更为精确有效的定位。
不同于传统的神经网络，本实施例的循环神经网络引入了定向循环，能够处理那些输入之间前后关联的问题，即一个序列当前的输 出与前面的输出也有关。具体的表现形式为网络会对前面的信息进行记忆并应用于当前输出的计算中，即隐藏层之间的节点不再无连接而是有连接的，并且隐藏层的输入不仅包括输入层的输出还包括上一时刻隐藏层的输出。所以本公开中引入了循环神经网络，与传统前馈神经网络组成深度学习模块，结合了聚类算法，利用终端上报数据中的主小区的RSRP、RSRQ、TA、AOA、AGPS、小区经纬度等数据，按以下关键步骤，实现了对未能上报AGPS数据的终端用户运动轨迹的定位。
请见附图2，本实施例中定位流程中关键的深度学习模块工作原理为：
(1)由前馈神经网络对聚类输出的k簇数据进行训练拟合，此阶段将会输出k组前馈神经网络超参数，用来描述各簇经纬度与上报数据中RSRP等信号指标的关系。用聚类算法对大量数据进行预处理，使得各簇中数据的相关性更高，用前馈神经网络来学习预处理过后的数据，更容易收敛，从而减少训练时间，提高模型的性能表现；
(2)循环神经网络用于对时间序列进行拟合，也就是将过去某个连续时间内前馈神经网络的输出作为循环神经网络的输入，当前的时间点的经纬度作为循环网络的输出进行训练，得到的模型就能够依据话单移动的轨迹进行下一时刻位置点的判定。
本实施例主要由离线训练阶段和在线预测阶段组成。
离线训练阶段主要包括以下七个步骤。
在第一步，获取有效上报数据，以IMSI为主建，拼接各个用户的话单。
在第二步，通过枚举，令k从2到一个固定值如10，在每个k值上重复运行数次K均值算法，避免局部最优解，并计算当前k值的平均轮廓系数，最后选取轮廓系数最大的值对应的k作为最终的集群数目。
在第三步，构建k个相同的多层前馈神经网络，神经网络的隐层数与各个隐层的节点数由具体实验调参过程确认。
在第四步，对步骤二得到k簇数据集进行归一化。
在第五步，分别将k簇话单数据输入各前馈神经网络，进行p 次迭代直至误差收敛满足门限值。
在第六步，缓存第五步的输出序列，形成[t，m，k]维特征数据作为循环神经网络的输入，该步缓存的目的主要是形成轨迹序列。
在第七步，生成一个循环神经网络，将第六步得到的数据作为输入，终端最后上报的AGPS数据作为输出，进行p次迭代直至误差收敛满足门限值，得到离线模型。
在线预测阶段可以包括以下三个步骤。
在第一步，获取没有上报AGPS数据，以IMSI为主建，关联相同IMSI号的数据，形成待定位话单数据。
在第二步，对第一步得到的数据按由离线训练阶段得到的最优k值进行聚类。
在第三步，将k簇数据输入通过深度学习得到的离线模型，得到经纬度输出。
根据本实施例的一个示例，上述步骤S101-S104的标号并不用于限定本实施例中各个步骤的先后顺序，各个步骤的编号只是为了使得描述各个步骤时可以通用引用该步骤的标号进行便捷的指代，只要各个步骤执行的顺序不影响本实施例的逻辑关系即表示在本申请请求保护的范围之内。
本实施例提供的基于深度学习的定位方法通过利用聚类算法对上报的大量看似杂乱无章的数据进行学习，挖掘出相似的数据后又利用深度学习框架对各簇数据进行特征学习，最终生成离线模型，利用此模型可以有效对未上报位置信息的终端进行定位。该模型通用性强，具有很高的抗干扰能力，定位迅速并且定位精度得到了很大的提高，100米内能达到70％。根据本公开的一个实施例的定位精度验证如图7所示。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件来实现，但很多情况下前者是更佳的实施方式。基于这样的理解，本公开的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存 储在一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，空调器，或者网络设备等)执行本公开各个实施例所述的方法。
图8为根据本公开的一个实施例的基于深度学习的定位装置的示范性结构框图，下面结合图8来详细描述根据本公开的一个实施例的基于深度学习的定位装置，如图8所示，该基于深度学习的定位装置100可以包括接收模块11、分类模块12以及处理模块13。
接收模块11配置为接收移动终端上报的若干个待定位的第一小区信号参数，该第一小区信号参数不包括该移动终端的实际经纬度。
分类模块12配置为将同一移动终端上报的该第一小区信号参数输入至预先训练好的聚类模型中进行分类。
处理模块13配置为将分类后的每类该第一小区信号参数分别输入到预先训练好的前馈神经网络中，得到与该第一小区信号参数相对应的当前经纬度数据。
该处理模块13还配置为将该当前经纬度数据输入到预先训练好的循环神经网络中，得到下一时刻的经纬度数据。
在一个实施例中，该接收模块11还配置为接收移动终端上报的若干个位置点的第二小区信号参数，该第二小区信号参数包括该移动终端的实际经纬度。
该基于深度学习的定位装置100还可以包括：排序模块，其配置为按照上报的时间顺序对同一移动终端上报的该第二小区信号参数进行排序；聚类模块，其配置为根据排序后的该第二小区信号参数，利用聚类算法将相似度高于预设值的第二小区信号参数对应的位置点归为一类；该聚类模块还配置为通过确定该位置点被划分的最优类数以及每类所包含的位置点，确定该聚类模型。
在一个实施例中，该第二小区信号参数还包括参考信号接收功率RSRP、参考信号接收质量RSRQ、到达角AOA及时间提前量TA；该基于深度学习的定位装置100还可以包括计算模块，该计算模块通过以下公式计算两个位置点之间的距离：
Figure PCTCN2018123111-appb-000008
其中，RSRP 1表示位置点S 1的参考信号接收功率，RSRP 2表示位置点S 2的参考信号接收功率，RSRQ 1表示位置点S 1的参考信号接收质量，RSRQ 2表示位置点S 2的参考信号接收质量，TA 1表示位置点S 1的时间提前量，TA 2表示位置点S 2的时间提前量，AOA 1表示位置点S 1的到达角，AOA 2表示位置点S 2的到达角，D表示计算出的所述位置点S 1与位置点S 2的距离，计算出的D值越小表示位置点S 1与位置点S 2的相似度越高。
在一个实施例中，该聚类模块还配置为接收到输入的K值时，将所有该位置点被划分的类数确定为K类；上述计算模块还配置为将该位置点分为种子点及其它位置点，随机在该位置点中确定K个种子点，计算所有其他位置点到每个该种子点的距离；该聚类模块还配置为将每个该其他位置点划分到计算的距离最近的种子点所在的类中，完成第一次聚类；上述计算模块还配置为计算第一次聚类后被划分在同一类中的位置点的第二小区信号参数的平均值，该第二小区信号参数包括参考信号接收功率的平均值、参考信号接收质量的平均值、时间提前量的平均值及到达角的平均值；上述计算模块还配置为以计算的该第二小区信号参数的平均值对应的位置点作为新的种子点，对该其他位置点被划分的类以及划分后该第二小区信号参数的平均值进行迭代计算，直至计算的该第二小区信号参数的平均值不再变化或变化的差值小于预设的门限时，确定该位置点被划分为K类时每类所包含的位置点，以得到该聚类模型。
在一个实施例中，该聚类模块还可以包括：接收单元，其配置为接收输入的若干个枚举的K值，该K值表示所有位置点被划分的类数；上述计算模块还配置为确定枚举的其中一个K i的值，通过以下公式计算该位置点被分为K i类时的平均轮廓系数：
Figure PCTCN2018123111-appb-000009
其中，S(i)表示计算的与K i对应的平均轮廓系数，a(n)表示位置点n到所有与所述位置点n划分在同一类的其它位置点的距离的平均值，b(n)表示位置点n到所有与所述位置点n划分在不同类的其它 位置点的距离的平均值；上述计算模块还配置为遍历枚举的所有所述K值，计算与每个所述K值对应的平均轮廓系数；该聚类模块还配置为将与计算的最大平均轮廓系数对应的K值确定为所述最优K值。
在一个实施例中，该基于深度学习的定位装置100还可以包括：前馈神经网络构建模块，其配置为在确定该位置点被划分的最优类数以及每类所包含的位置点的步骤之后，构建前馈神经网络，构建的前馈神经网络的个数与该最优类数相同且一一对应；处理模块，其配置为分别对每类所包含的位置点的第二小区信号参数进行归一化处理，得到与每个位置点相对应的特征向量；第一训练模块，其配置为将属于同一类的位置点的特征向量作为对应前馈神经网络的输入，将接收的该位置点的实际经纬度作为对应前馈神经网络的输出，对每个该前馈神经网络进行训练，直至该前馈神经网络的输出误差收敛满足预设的第一门限，得到训练好的前馈神经网络。
在一个实施例中，该基于深度学习的定位装置100还可以包括：循环神经网络构建模块，其配置为构建循环神经网络；第二训练模块，其配置为根据排序后该位置点的下一位置点的实际经纬度对该循环神经网络进行训练，直至该循环神经网络的输出误差收敛满足预设的第二门限，得到训练好的循环神经网络模型。
在一个实施例中，该基于深度学习的定位装置100还可以包括：排序模块，其配置为按照上报时刻的先后顺序，对同一移动终端上报的该第一小区信号参数进行排序；该分类模块还配置为将排序后的第一小区信号参数输入至该聚类模型中进行分类。
根据本公开实施例，该基于深度学习的定位装置中包括的各个模块可全部或部分通过软件、硬件或其组合来实现。进一步地，该基于深度学习的定位装置中的各个模块可以是用于实现对应功能的程序段。
上述基于深度学习的定位装置可以实现为一种计算机程序的形式，计算机程序可以在如图9所示的计算机设备上运行。
需要说明的是，上述装置实施例与方法实施例属于同一构思，其实现过程详见方法实施例，且方法实施例中的技术特征在装置实施 例中均对应适用，这里不再赘述。
根据本公开的一个实施例提供的一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，该处理器执行该程序时实现上述基于深度学习的定位方法。
图9为根据本公开的一个实施例的计算机设备的内部结构示意图，该计算机设备可以为服务器。参照图9，该计算机设备包括通过系统总线连接的处理器、存储器、输入装置、显示屏和网络接口。该存储器包括非易失性存储介质和内存储器，该计算机设备的非易失性存储介质可存储操作系统和计算机可读指令，该计算机可读指令被执行时，可使得处理器执行本申请各实施例的一种基于深度学习的定位方法，该方法的实现过程可参考图1至图7各实施例的内容，在此不再赘述。该计算机设备的处理器配置为提供计算和控制能力，支撑整个计算机设备的运行。该内存储器中可储存有计算机可读指令，该计算机可读指令被处理器执行时，可使得处理器执行一种基于深度学习的定位方法。计算机设备的输入装置配置为用于各个参数的输入，计算机设备的显示屏配置为进行显示，计算机设备的网络接口配置为进行网络通信。本领域技术人员可以理解，图9中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
本实施例中的存储器可配置为存储软件程序以及各种数据。存储器可主要包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、至少一个功能所需的应用程序等；存储数据区可存储根据手机的使用所创建的数据等。此外，存储器可以包括高速随机存取存储器，还可以包括非易失性存储器，例如至少一个磁盘存储器件、闪存器件、或其他易失性固态存储器件。
本实施例另提供一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现上述基于深度学习的定位方法中的各个步骤。
根据本实施例的一个示例，上述实施例方法中的全部或部分流 程，可以通过计算机程序来指令相关的硬件来完成，所述程序可存储于一计算机可读取存储介质中，如本公开实施例中，该程序可存储于计算机系统的存储介质中，并被该计算机系统中的至少一个处理器执行，以实现包括如上述各方法的实施例的流程。该存储介质包括但不限于磁碟、U盘、光盘、只读存储记忆体(Read-Only Memory，ROM)等。
本实施例提供的一种基于深度学习的定位方法、装置、计算机设备及存储介质，通过根据移动终端上报的若干个待定位的第一小区信号参数，将该第一小区信号参数分别输入到训练好的聚类模型、前馈神经网络及循环神经网络中，可以得到移动终端当前时刻的经纬度及下一时刻的经纬度，由于该第一小区信号参数包含了上报的多种参数指标，使得依据该第一小区信号参数得到的定位信息更加准确，且由于该循环神经网络具有位置预测功能，能够实时的得到下一时刻的经纬度，使得定位功能的稳定性更强。
以上参照附图说明了本公开的示例性实施例，并非因此局限本公开的权利范围。本领域技术人员不脱离本公开的范围和实质内所作的任何修改、等同替换和改进，均应在本公开的权利范围之内。

Claims (11)

  1. 一种基于深度学习的定位方法，包括：
    接收移动终端上报的若干个待定位的第一小区信号参数，所述第一小区信号参数不包括所述移动终端的实际经纬度；
    将同一移动终端上报的所述第一小区信号参数输入至预先训练好的聚类模型中进行分类；
    将分类后的每类所述第一小区信号参数分别输入到预先训练好的前馈神经网络中，得到与所述第一小区信号参数相对应的当前经纬度数据；
    将所述当前经纬度数据输入到预先训练好的循环神经网络中，得到下一时刻的经纬度数据。
  2. 根据权利要求1所述的方法，还包括：
    接收移动终端上报的若干个位置点的第二小区信号参数，所述第二小区信号参数包括所述移动终端的实际经纬度；
    按照上报的时间顺序对同一移动终端上报的所述第二小区信号参数进行排序；
    根据排序后的所述第二小区信号参数，利用聚类算法将相似度高于预设值的第二小区信号参数对应的位置点归为一类；
    通过确定所述位置点被划分的最优类数以及每类所包含的位置点，确定所述聚类模型。
  3. 根据权利要求2所述的方法，其中，所述第二小区信号参数还包括参考信号接收功率RSRP、参考信号接收质量RSRQ、到达角AOA及时间提前量TA；所述方法还包括通过以下公式计算两个位置点之间的距离：
    Figure PCTCN2018123111-appb-100001
    其中，RSRP 1表示位置点S 1的参考信号接收功率，RSRP 2表示位置点S 2的参考信号接收功率，RSRQ 1表示位置点S 1的参考信号接收质 量，RSRQ 2表示位置点S 2的参考信号接收质量，TA 1表示位置点S 1的时间提前量，TA 2表示位置点S 2的时间提前量，AOA 1表示位置点S 1的到达角，AOA 2表示位置点S 2的到达角，D表示计算出的所述位置点S 1与位置点S 2的距离，计算出的D值越小表示位置点S 1与位置点S 2的相似度越高。
  4. 根据权利要求3所述的方法，还包括：
    接收到输入的K值时，将所有所述位置点被划分的类数确定为K类；
    将所述位置点分为种子点及其它位置点，随机在所述位置点中确定K个种子点，计算所有其他位置点到每个所述种子点的距离；
    将每个所述其他位置点划分到计算的距离最近的种子点所在的类中，完成第一次聚类；
    计算第一次聚类后被划分在同一类中的位置点的第二小区信号参数的平均值，所述第二小区信号参数包括参考信号接收功率的平均值、参考信号接收质量的平均值、时间提前量的平均值及到达角的平均值；
    以计算的所述第二小区信号参数的平均值对应的位置点作为新的种子点，对所述其他位置点被划分的类以及划分后所述第二小区信号参数的平均值进行迭代计算，直至计算的所述第二小区信号参数的平均值不再变化或变化的差值小于预设的门限时，确定所述位置点被划分为K类时每类所包含的位置点，以得到所述聚类模型。
  5. 根据权利要求4所述的方法，其中，确定所述最优K值的步骤包括：
    接收输入的若干个枚举的K值，所述K值表示所有位置点被划分的类数；
    确定枚举的其中一个K i的值，通过以下公式计算所述位置点被分为K i类时的平均轮廓系数：
    Figure PCTCN2018123111-appb-100002
    其中，S(i)表示计算的与K i对应的平均轮廓系数，a(n)表示位置点n到所有与所述位置点n划分在同一类的其它位置点的距离的平均值，b(n)表示位置点n到所有与所述位置点n划分在不同类的其它位置点的距离的平均值；
    遍历枚举的所有所述K值，计算与每个所述K值对应的平均轮廓系数；
    将与计算的最大平均轮廓系数对应的K值确定为所述最优K值。
  6. 根据权利要求2所述的方法，其中，训练所述前馈神经网络的步骤包括：
    在确定所述位置点被划分的最优类数以及每类所包含的位置点的步骤之后，构建前馈神经网络，构建的前馈神经网络的个数与所述最优类数相同且一一对应；
    分别对每类所包含的位置点的第二小区信号参数进行归一化处理，得到与每个位置点相对应的特征向量；
    将属于同一类的位置点的特征向量作为对应前馈神经网络的输入，将接收的所述位置点的实际经纬度作为对应前馈神经网络的输出，对每个所述前馈神经网络进行训练，直至所述前馈神经网络的输出误差收敛满足预设的第一门限，得到训练好的前馈神经网络。
  7. 根据权利要求6所述的方法，其中，训练所述循环神经网络的步骤包括：
    构建循环神经网络；
    根据排序后所述位置点的下一位置点的实际经纬度对所述循环神经网络进行训练，直至所述循环神经网络的输出误差收敛满足预设的第二门限，得到训练好的循环神经网络模型。
  8. 根据权利要求1所述的方法，其中，在接收移动终端上报的 若干个待定位的第一小区信号参数的步骤之后，所述方法还包括：
    按照上报时刻的先后顺序，对同一移动终端上报的所述第一小区信号参数进行排序；
    将排序后的第一小区信号参数输入至所述聚类模型中进行分类。
  9. 一种基于深度学习的定位装置，包括：
    接收模块，其配置为接收移动终端上报的若干个待定位的第一小区信号参数，所述第一小区信号参数不包括所述移动终端的实际经纬度；
    分类模块，其配置为将同一移动终端上报的所述第一小区信号参数输入至预先训练好的聚类模型中进行分类；
    处理模块，其配置为将分类后的每类所述第一小区信号参数分别输入到预先训练好的前馈神经网络中，得到与所述第一小区信号参数相对应的当前经纬度数据；
    所述处理模块还配置为将所述当前经纬度数据输入到预先训练好的循环神经网络中，得到下一时刻的经纬度数据。
  10. 一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述程序时实现如权利要求1至8中任一项所述的基于深度学习的定位方法。
  11. 一种计算机可读存储介质，其上存储有计算机程序，所述程序被处理器执行时实现如权利要求1至8中任一项所述的基于深度学习的定位方法中的步骤。
PCT/CN2018/123111 2018-01-02 2018-12-24 基于深度学习的定位方法、装置、计算机设备及存储介质 WO2019134549A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810009221.3 2018-01-02
CN201810009221.3A CN109991591B (zh) 2018-01-02 2018-01-02 基于深度学习的定位方法、装置、计算机设备及存储介质

Publications (1)

Publication Number Publication Date
WO2019134549A1 true WO2019134549A1 (zh) 2019-07-11

Family

ID=67129025

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/123111 WO2019134549A1 (zh) 2018-01-02 2018-12-24 基于深度学习的定位方法、装置、计算机设备及存储介质

Country Status (2)

Country Link
CN (1) CN109991591B (zh)
WO (1) WO2019134549A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115038040A (zh) * 2022-06-29 2022-09-09 中国联合网络通信集团有限公司 一种小区定位方法、装置、设备、系统及介质

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11303041B2 (en) * 2019-07-24 2022-04-12 Cypress Semiconductor Corporation Machine learning-based antenna array validation, prototyping and optimization
CN112887897B (zh) * 2019-11-29 2022-07-05 中国电信股份有限公司 终端的定位方法、装置和计算机可读存储介质
CN114222238B (zh) * 2020-09-03 2024-04-26 中国电信股份有限公司 定位方法、装置和计算机可读存储介质
CN115913486A (zh) * 2021-09-18 2023-04-04 维沃移动通信有限公司 信息上报方法、装置、终端及可读存储介质
CN114745674B (zh) * 2022-04-27 2024-06-07 上海鼎为物联技术有限公司 一种基于改进bp神经网络的测距模型定位算法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
CN105163326A (zh) * 2015-09-30 2015-12-16 南京华苏科技股份有限公司 一种基于无线网络话务特征的小区聚类方法和系统
CN105960021A (zh) * 2016-07-07 2016-09-21 济南东朔微电子有限公司 一种改进的位置指纹室内定位方法
CN106960256A (zh) * 2017-03-17 2017-07-18 中山大学 基于时间空间上下文的循环神经网络预测位置的方法
CN107480784A (zh) * 2017-06-28 2017-12-15 青岛科技大学 一种基于深度学习的手机信令数据行人交通轨迹预测方法
CN107517446A (zh) * 2017-09-28 2017-12-26 知谷（上海）网络科技有限公司 基于Wi‑Fi热点的室内定位方法及装置

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8023965B2 (en) * 2008-05-09 2011-09-20 Mitel Networks Corporation Method, system and apparatus for locating a mobile communications device
CN102480677B (zh) * 2010-11-24 2015-11-25 中国移动通信集团公司 一种指纹定位误差的确定方法和设备
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
DE102016007600A1 (de) * 2016-06-21 2017-12-21 Giesecke+Devrient Mobile Security Gmbh Authentifizieren eines tragbaren Endgeräts
CN106604229B (zh) * 2016-12-27 2020-02-18 东南大学 一种基于流形学习与改进支持向量机的室内定位方法
CN107124696B (zh) * 2017-03-20 2020-05-26 东华大学 一种基于蓝牙rssi的无钥匙系统定位方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
CN105163326A (zh) * 2015-09-30 2015-12-16 南京华苏科技股份有限公司 一种基于无线网络话务特征的小区聚类方法和系统
CN105960021A (zh) * 2016-07-07 2016-09-21 济南东朔微电子有限公司 一种改进的位置指纹室内定位方法
CN106960256A (zh) * 2017-03-17 2017-07-18 中山大学 基于时间空间上下文的循环神经网络预测位置的方法
CN107480784A (zh) * 2017-06-28 2017-12-15 青岛科技大学 一种基于深度学习的手机信令数据行人交通轨迹预测方法
CN107517446A (zh) * 2017-09-28 2017-12-26 知谷（上海）网络科技有限公司 基于Wi‑Fi热点的室内定位方法及装置

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115038040A (zh) * 2022-06-29 2022-09-09 中国联合网络通信集团有限公司 一种小区定位方法、装置、设备、系统及介质

Also Published As

Publication number Publication date
CN109991591A (zh) 2019-07-09
CN109991591B (zh) 2023-08-22

Similar Documents

Publication Publication Date Title
WO2019134549A1 (zh) 基于深度学习的定位方法、装置、计算机设备及存储介质
US9239990B2 (en) Hybrid location using pattern recognition of location readings and signal strengths of wireless access points
US8738032B2 (en) Hybrid location using a weighted average of location readings and signal strengths of wireless access points
CN111867049B (zh) 定位方法、装置及存储介质
CN104185275B (zh) 一种基于wlan的室内定位方法
CN113988314B (zh) 一种选择客户端的分簇联邦学习方法及系统
Chidlovskii et al. Semi-supervised variational autoencoder for WiFi indoor localization
CN108882172B (zh) 一种基于hmm模型的室内移动轨迹数据的预测方法
WO2014106363A1 (zh) 移动设备定位系统及方法
CN113993172B (zh) 一种基于用户移动行为预测的超密集网络切换方法
CN110956255A (zh) 难样本挖掘方法、装置、电子设备及计算机可读存储介质
CN101740029B (zh) 应用于基于矢量量化的说话人识别的三粒子协同优化方法
CN113037410A (zh) 信道识别方法、装置、传输方法、传输设备、基站、介质
CN111698695A (zh) 一种基于神经网络的lte指纹式定位方法
CN107290714B (zh) 一种基于多标识指纹定位的定位方法
CN110557829B (zh) 一种融合指纹库的定位方法及定位装置
CN107104747B (zh) 无线时变信道中的多径分量的分簇方法
WO2023061303A1 (zh) 大尺度衰落的建模及估计方法、系统、设备和存储介质
CN110081890A (zh) 一种结合深度网络的动态k最近邻地图匹配方法
CN113033653B (zh) 一种边-云协同的深度神经网络模型训练方法
CN111343664A (zh) 用户定位方法、装置、设备及介质
EP4133815A1 (en) Position estimation
Qiao et al. CWEmd: A Lightweight Similarity Measurement for Resource-Constrained Vehicular Networks
US11962475B2 (en) Estimating properties of units using system state graph models
Musa et al. Optimal sampling for Moving Object Trajectory Tracking in Smart Transportation Systems: A Transformer-based Approach

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: 18898165

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 12.11.2020)

122 Ep: pct application non-entry in european phase

Ref document number: 18898165

Country of ref document: EP

Kind code of ref document: A1