Single-Station Positioning Method and System Combining Scene Recognition with Range and Angle Measurement Error Calibration

Abstract

The present invention provides a single-station positioning method and system for joint scene recognition and ranging and angle measurement error calibration. Based on an indoor positioning scenario, waveforms of positioning signals with annotations and real distance and angle information within the indoor positioning scenario are collected to construct a scene feature database; a training pool is constructed based on the scene feature database to train and generate a learning network; based on the generated learning network, variational autoencoders are used for processing to obtain distance codewords, angle codewords, and scene codewords, and the distance codewords, angle codewords, and scene codewords are input into corresponding calibrators to calibrate newly input ranging and angle measurement values, outputting calibrated positioning results, and simultaneously updating the scene feature database. The present invention solves the problems of separate design for ranging and angle measurement error calibration and high annotation costs in the prior art, and realizes joint correction of ranging and angle measurement errors of a single base station.

Description

Technical Field

[0001] The present invention relates to the technical field of positioning and calibration, and particularly to a single-station positioning method and system that combines scene recognition with ranging and angle measurement error calibration. Background Art

[0002] At present, most of the existing technologies achieve high-precision indoor positioning by combining multiple base stations with non-line-of-sight error calibration algorithms, including: TOA ranging error correction method based on line-of-sight and non-line-of-sight discrimination, non-line-of-sight error calibration method based on neural network, and ultra-wideband non-line-of-sight discrimination and elimination method based on LS-SVM. In indoor positioning scenarios, multipath and non-line-of-sight (NLoS, Non-Line-of-Sight) effects are one of the main sources of positioning errors, especially in the application of multi-antenna single-station positioning technology. In terms of azimuth (AoA, Angle-of-Arrival) estimation, multipath may cause misdetection of the direct angle component, while non-line-of-sight results in the absence of the direct angle, unable to correctly provide the angle of the target relative to the base station; in terms of distance (ToF, Time-of-Flight) estimation, the multipath effect may cause misjudgment of the direct component, while non-line-of-sight will cause a positive deviation in distance estimation, resulting in a relatively large estimated distance between the target and the base station compared to the true value. If the ranging and angle measurement values of the base station are directly used for positioning, it will lead to serious deviation of the positioning result. Therefore, in actual measurement, it is necessary to calibrate the measurement values to improve the positioning accuracy.

[0003] The current multipath non-line-of-sight ranging and angle measurement error calibration technology has the following problems: simple non-line-of-sight angle measurement error calibration lacks other information support and has low reliability; indoor positioning scenarios vary greatly, and only using line-of-sight / non-line-of-sight scene recognition is difficult to provide a scalable positioning solution; ranging and angle measurement error calibration are all based on different theories, frameworks, and algorithms, without considering the correlation between the two in the actual environment and lacking an integrated joint design; indoor positioning scenarios are quite different, and the full-supervised scheme requires a large amount of labeled data, and the cost of label acquisition is relatively high. Summary of the Invention

[0004] The present invention provides a single-station positioning method and system that combines scene recognition with ranging and angle measurement error calibration to solve the problems of separate design of ranging and angle measurement error calibration and high annotation cost in the prior art, and to achieve single-base-station ranging and angle measurement error correction.

[0005] The present invention provides a single-station positioning method that combines scene recognition with ranging and angle measurement error calibration, including:

[0006] Based on the indoor positioning scenario, collect the waveforms of the positioning signals with annotations and the real distance and angle information in the indoor positioning scenario, and construct a scene feature database;

[0007] Construct a training pool based on the scene feature database and train to generate a learning network;

[0008] Based on the generated learning network, process it through a variational autoencoder to obtain distance codewords, angle codewords, and scene codewords, and input the distance codewords, angle codewords, and scene codewords into the corresponding calibrators to calibrate the newly input ranging and angle measurement values, output the calibrated positioning result, and simultaneously update the scene feature database.

[0009] According to a single-station positioning method for joint scene recognition and ranging / angle measurement error calibration provided by the present invention, based on the indoor positioning scene, collect the waveforms of the labeled positioning signals and the true distance and angle information in the indoor positioning scene, and construct a scene feature database, specifically including:

[0010] Select the environment and area of the scene. The scene is distinguished by room type and internal partitions, define the scene label, and grid the scene. There are base stations and target points in the scene. Measure the ranging value, angle measurement value, waveform, waveform feature, true distance, and angle value between the base station and the target point until no new recognizable features are added;

[0011] Change the grid point positions of the base station and the target point, measure the ranging value, angle measurement value, waveform, waveform feature, true distance, and angle value between the base station and the target point until no new recognizable features are added, and measure all grid point positions;

[0012] Replace the scene and repeat the grid point position measurement until all scenes in the application environment are covered.

[0013] According to a single-station positioning method for joint scene recognition and ranging / angle measurement error calibration provided by the present invention, the construction of a training pool based on the scene feature database and the training to generate a learning network specifically include:

[0014] Construct a training pool, and the training pool includes a labeled data set and an unlabeled data set The labeled data set includes the scene feature database, and the unlabeled data set includes the high-quality results selected during the working calibration process, where x (i) is the waveform feature data, Δd (i) is the ranging error, Δθ (i) is the angle measurement error, and k (i) is the scene label;

[0015] Use the data in the training pool to train and generate a learning network. For the labeled data set in the training pool construct a supervised loss term, and for the unlabeled data set in the training pool Construct an unsupervised loss term, and through the supervised loss term and the unsupervised loss term, construct an overall network loss function for the training pool to generate a learning network.

[0016] According to a single-station positioning method for joint scene recognition and ranging and angle measurement error calibration provided by the present invention, based on the generated learning network, it is processed through a variational autoencoder to obtain a distance codeword, an angle codeword, and a scene codeword, and the distance codeword, the angle codeword, and the scene codeword are input into corresponding calibrators, specifically including:

[0017] Input the waveform data x into the variational autoencoder to generate three Gaussian latent variable representations, including: the distance codeword y d , the angle codeword y θ , and the scene environment codeword y k ;

[0018] Based on the unsupervised loss term, substitute the data in the unlabeled dataset into the variational autoencoder for training to learn its latent variable mapping relationship with the labeled dataset;

[0019] Based on the supervised loss term, substitute the distance codeword y d , the angle codeword y θ , and the scene environment codeword y k generated from the labeled data into the corresponding distance estimator, angle estimator, and scene recognizer for training;

[0020] Based on the overall loss function, jointly optimize the labeled and unlabeled data through iteration, and output the ranging, angle measurement, and scene parameter estimations.

[0021] According to a single-station positioning method for joint scene recognition and ranging and angle measurement error calibration provided by the present invention, calibrate the newly input ranging and angle measurement values, output the calibrated positioning result, and update the scene feature database at the same time, specifically including:

[0022] An online training stage and an offline calibration stage;

[0023] In the online training stage, input the training pool dataset, learning rate, batch processing length, initial encoding parameters, initial decoding parameters, and initial sub-module parameters;

[0024] Extract a batch of samples from the training pool for training, and use the gradient descent algorithm to train the network encoder and network decoder;

[0025] Repeat the training process until convergence, and output the learning result.

[0026] A single - station positioning method combining scene recognition and ranging and angle - measurement error calibration provided by the present invention calibrates the newly input ranging and angle - measurement values, outputs the calibrated positioning result, and updates the scene feature database at the same time, specifically including:

[0027] In the offline calibration stage, input the real - time observed channel impulse response waveform data, as well as the coding parameters, distance calibrator parameters, angle calibrator parameters, and scene recognizer parameters obtained by training and learning based on the data set;

[0028] Based on the coding parameters, encode the observed data to generate distance codewords, angle codewords, and scene environment codewords;

[0029] Input the distance codeword y d into the distance calibrator to obtain the distance deviation estimate value;

[0030] Input the angle codeword y θ into the angle calibrator to obtain the angle deviation estimate value;

[0031] Input the scene recognition codeword y k into the scene recognizer to obtain the scene estimate value;

[0032] Output the distance deviation estimate value, angle deviation estimate value, and scene estimate value, and screen the non - standard data in the training pool to update the scene feature database.

[0033] The present invention also provides a single - station positioning system combining scene recognition and ranging and angle - measurement error calibration, and the system includes:

[0034] A scene feature database generation module, which is used to collect the waveforms of the labeled positioning signals and the real - distance and angle information in the indoor positioning scene based on the indoor positioning scene, and construct a scene feature database;

[0035] A learning network generation module, which is used to construct a training pool based on the scene feature database and train and generate a learning network;

[0036] A calibration module, which is used to process through a variational auto - encoder based on the generated learning network to obtain distance codewords, angle codewords, and scene codewords, and input the distance codewords, angle codewords, and scene codewords into the corresponding calibrators to calibrate the newly input ranging and angle - measurement values, output the calibrated positioning result, and update the scene feature database at the same time.

[0037] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the single - station positioning method combining scene recognition and ranging and angle - measurement error calibration as described in any one of the above.

[0038] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration as described in any one of the above.

[0039] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration as described in any one of the above.

[0040] The single-station positioning method and system for joint scene recognition and ranging and angle measurement error calibration provided by the present invention realize the joint calibration of non-line-of-sight ranging and angle measurement errors by combining small-sample labeled data and a large amount of unlabeled data obtained from actual applications during the self-learning process, reducing the data acquisition cost. At the same time, by combining real-time unstandardized data, it ensures that the regression model can be adaptively optimized, improving the algorithm adaptability.

[0041] Learning based on waveform features, jointly realizing scene recognition and eliminating ranging and angle measurement errors, reducing the loss caused by information processing during the positioning process. Among them, there is a strong coupling between the positive deviations of the angle and distance caused by non-line-of-sight, and the scene material and physical obstruction conditions directly affect their probability distribution. Therefore, it can significantly improve the positioning performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0043] Figure 1 is one of the flow schematic diagrams of the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration provided by the present invention;

[0044] Figure 2 is another flow schematic diagram of the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration provided by the present invention;

[0045] Figure 3 is yet another flow schematic diagram of the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration provided by the present invention;

[0046] Figure 4 is the scene grid diagram provided by the present invention;

[0047] Figure 5 is the variational autoencoder network diagram provided by the present invention;

[0048] Figure 6 is the flowchart of the online selection algorithm provided by the present invention;

[0049] Figure 7 is the flowchart of the offline recognition and calibration provided by the present invention;

[0050] Figure 8 is the schematic connection diagram of the single-station positioning system module for the joint scene recognition and ranging and angle measurement error calibration provided by the present invention;

[0051] Figure 9 is the schematic structural diagram of the electronic device provided by the present invention.

[0052] Reference numerals:

[0053] 810: scene feature database generation module; 820: learning network generation module; 830: calibration module; 910: processor; 920: communication interface; 930: memory; 940: communication bus. Detailed implementation manners

[0054] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without making creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0055] The following combines Figures 1 - 3 to describe the single-station positioning method for the joint scene recognition and ranging and angle measurement error calibration of the present invention, including:

[0056] S100. Based on the indoor positioning scene, collect the waveforms of the labeled positioning signals and the real distance and angle information in the indoor positioning scene, and construct a scene feature database; the scene feature database includes scene labels, real distances, measured distances, received signal energies, maximum amplitudes, rise times, average time delays, root mean square spreads, and kurtoses;

[0057] S200. Based on the scene feature database, construct a training pool and train to generate a learning network; the learning network includes four sub-network modules: a variational autoencoder, a ranging error calibrator, an angle measurement error calibrator, and a scene recognizer;

[0058] S300. Based on the generated learning network, process through the variational autoencoder to obtain distance codewords, angle codewords, and scene codewords, input the distance codewords, angle codewords, and scene codewords into the corresponding calibrators to calibrate the newly input ranging and angle measurement values, output the calibrated positioning results, and update the scene feature database at the same time.

[0059] The scene feature database is constructed by acquiring the waveforms of the labeled positioning signals and the true distance and angle information in the target scene. The scene feature database includes scene labels, true distances, measured distances, received signal energies, maximum amplitudes, rise times, average time delays, root mean square spreads, and kurtosis.

[0060] By combining small-sample labeled data and a large amount of unlabeled data obtained from actual applications during the self-learning process, joint calibration of non-line-of-sight ranging and angle measurement errors is achieved, reducing the data acquisition cost. At the same time, the combination of real-time unlabeled data ensures that the regression model can be adaptively optimized, improving the algorithm adaptability.

[0061] Based on the indoor positioning scenario, acquire the waveforms of the labeled positioning signals and the true distance and angle information in the indoor positioning scenario, and construct a scene feature database, specifically including:

[0062] S101. Select the experimental environment and area of the designated scene. The scenes are mainly distinguished by room types and internal partitions. Define the scene labels and grid the scenes. Room types include but are not limited to corridors, offices, and meeting rooms; internal partitions include but are not limited to glass, wood, sponge, metal, etc. Define the scene labels And grid the scenes As Figure 4 Shown. Place base station A and target point B at any two grid points (p i , p j ) in the scene, i, j ∈ [1, N g , i ≠ j, record the ranging value, angle measurement value, waveform, and waveform characteristics between single base station A and target point B, and record the true distance and angle values between target point B and base station A until no new recognizable features are added.

[0063] S102. Under the scene label, change the grid point positions of base station A or target B, continue to measure and record the ranging value, angle measurement value, waveform, and waveform characteristics between single base station A and target B, and record the true distance and angle values between target B and base station A until no new recognizable features are added;

[0064] S103. Repeat steps S101 - S102 until all grid point combinations (p i , p j ), i, j ∈ [1, N g , i ≠ j are recorded;

[0065] S104. Change the scene (different room types or partitions), define the scene label Repeat steps S102 - S103 until the scene label set It covers up to the scene types in the specified application environment. Specifically, it basically covers the house types or major material categories respectively.

[0066] By traversing all points in the scene, the acquisition of all elements of the scene is realized, ensuring that no recognizable features are missed. After collecting all waveform features, ranging errors, angle measurement errors, and scene label information, a complete scene feature database is established, which helps with subsequent training and learning.

[0067] Based on the scene feature database, a training pool is constructed, and a learning network is trained and generated, specifically including:

[0068] Construction of the training pool, where the training pool contains two types of data sets, namely the labeled data set and the unlabeled data set Among them, the labeled data set is composed of the scene feature database, while the unlabeled data set is composed of high-quality results selected during the system working calibration process, where x (i) is waveform-related data (waveform features), Δd (i) is the ranging error, Δθ (i) is the angle measurement error, and k (i) is the scene label. The waveform-related data includes waveforms and waveform features such as received signal energy, maximum amplitude, rise time, average excess delay, root mean square spread, and kurtosis:

[0069] Received signal energy: ε r =∫ T |r(t)| 2 dt

[0070] Maximum amplitude:

[0071] Rise time: t rise =min{t:|r(t)|≥βr max}-min{t:|r(t)|≥ασ n}

[0072] Average excess delay:

[0073] Root mean square spread:

[0074] Kurtosis:

[0075] Training and generation of the learning network, define the encoder f(·; φ): x→y d ,y θ ,y k ,where the variational posterior probability q of the three latent variablesφ (y d ,y θ ,y k (x|y) can be learned through the encoder f(; φ); define the decoder g(·; ψ): y d ,y θ ,y k →x, where the likelihood distribution p ψ (x|y d ,y θ ,y k ) can be learned through the decoder g(·; ψ). Since the three latent variables are independent of each other, the encoding network can be decomposed into three sub-modules, the ranging error calibrator the azimuth error calibrator and the scene recognizer Among them, the relevant variational posterior probability can be learned through the corresponding module:

[0076]

[0077]

[0078]

[0079] Based on the above variational distribution of latent variables, the evidence lower bound (ELBO, evidence lower bound) for the data x in the training pool can be expressed as:

[0080]

[0081] For the labeled dataset and the unlabeled dataset in the training pool, different reconstruction loss terms will be used when generating the autoencoder:

[0082] For the unlabeled dataset with low confidence in the dataset the unsupervised loss term is constructed as:

[0083]

[0084] For the labeled dataset with high confidence in the dataset the supervised loss term is constructed as:

[0085]

[0086] Among them, L label is the loss term for labeled data constructed for the parameters of the ranging error calibrator, the angle error calibrator, and the scene recognizer and is specifically expressed as:

[0087]

[0088] After establishing the learning network, learning is directly carried out based on waveform features, jointly realizing scene recognition and ranging and angle measurement error elimination, reducing the loss caused by information processing during the positioning process. Among them, there is a strong coupling between the positive deviations of the angles and distances caused by non-line-of-sight, and the scene material and physical blocking conditions directly affect their probability distributions. Therefore, this joint processing method has a significant gain in improving the positioning performance.

[0089] As Figure 5 shown, the generated learning network includes: a variational autoencoder, a ranging error calibrator, an angle error calibrator, and a scene recognizer. The variational autoencoder is used for latent variable encoding and signal reconstruction. The ranging error calibrator is used to calibrate the ranging measurement results. The angle error calibrator is used to calibrate the angle measurement. The scene recognizer is used to identify and classify the usage scenarios.

[0090] Referring to Figure 6 , based on the generated learning network, through the variational autoencoder for processing, distance codewords, angle codewords, and scene codewords are obtained, and the distance codewords, angle codewords, and scene codewords are input into the corresponding calibrators, specifically including:

[0091] S301. Input the waveform data x into the variational autoencoder to generate three Gaussian latent variable representations, namely the distance codeword y d , the angle codeword y θ , and the scene environment codeword y k ;

[0092] S302. Based on the unsupervised loss term, substitute the data in the unlabeled dataset into the variational autoencoder for training to learn its latent variable mapping relationship with the labeled dataset;

[0093] S303. Based on the supervised loss term, substitute the distance codeword y d , the angle codeword y θ , and the scene environment codeword y k generated from the labeled data into the corresponding distance estimator, angle estimator, and scene recognizer for training;

[0094] S304. Based on the overall loss function, jointly optimize the labeled and unlabeled data through iteration, and output the ranging, angle measurement, and scene parameter estimations.

[0095] After completing the learning process through the unsupervised loss term and the supervised loss term function, the learning network is output, and the loss caused by information processing during the positioning process is reduced through the learning network for calibration.

[0096] Referring to Figure 7, calibrate the newly input ranging and angle measurement values, output the calibrated positioning result, and update the scene feature database at the same time, specifically including:

[0097] In the online training stage, input the training pool dataset Learning rate α, batch processing length m, initial encoding parameter φ0, initial decoding parameter ψ0, and initial sub-module parameter

[0098] Extract a batch of samples from the training pool for training, and use the gradient descent algorithm to train and generate a network encoder and a network decoder, where:

[0099] If the data x (i) comes from the unlabeled dataset Then:

[0100]

[0101]

[0102] φ ← φ + α * Adam(φ, f φ )

[0103] ψ ← ψ + α * Adam(ψ, g ψ )

[0104] If the data x (i) comes from the labeled dataset Then extract the corresponding labeled data Δd (i) , Δθ (i) , k (i) , and substitute them into the training:

[0105]

[0106]

[0107]

[0108] φ ← φ + α * Adam(φ, f φ )

[0109] ψ ← ψ + α * Adam(ψ, g ψ )

[0110]

[0111] Repeat the above training until convergence, and output the learning results φ * , ψ * ,

[0112] In the offline calibration stage, input the real-time observed channel impulse response (CIR) waveform data x ob , as well as the coding parameters φ * and ψ * obtained through training and learning based on the data set, and the parameters of the distance calibrator, angle calibrator, and scene recognizer

[0113] Based on the coding parameters φ * and ψ * , encode the observed data x ob to generate the variable codeword y d , y θ , y k ;

[0114] Based on the parameters , input the distance codeword y d into the distance calibrator to obtain the distance deviation estimate Based on the parameters , input the angle codeword y θ into the angle calibrator to obtain the angle deviation estimate Based on the parameters , input the scene recognition codeword y k into the scene recognizer to obtain the scene estimate

[0115] Output the distance deviation estimate the angle deviation estimate and the scene estimate and screen the non-standard data in the training pool, and at the same time update the scene feature database.

[0116] This embodiment proposes a single-station positioning method that combines scene recognition with ranging and angle measurement error calibration, reducing the negative impact brought by the non-line-of-sight effect in the single-base station positioning scenario;

[0117] In the self-learning process, combined with small-sample labeled data and a large amount of unlabeled data obtained from actual applications, realize the joint calibration of non-line-of-sight ranging and angle measurement errors, reduce the data acquisition cost, and at the same time, combined with real-time non-standard data to ensure that the regression model can be adaptively optimized, improving the algorithm adaptability;

[0118] Directly learn based on waveform features, jointly realize scene recognition and eliminate ranging and angle measurement errors, reducing the loss caused by information processing during the positioning process. Among them, there is a strong coupling between the positive deviations of the angle and distance caused by non-line-of-sight, and the scene material and physical blocking conditions directly affect their probability distributions. Therefore, this joint processing method has a significant gain in improving the positioning performance;

[0119] A learning framework based on variational autoencoders is proposed, and the ELBO (evidence lower bound) is derived. The algorithm performance is better evaluable. At the same time, the gradient descent algorithm designed based on the ELBO (evidence lower bound) has good convergence, and the overall scheme has strong robustness.

[0120] Reference Figure 8 , the present invention also discloses a single-station positioning system for joint scene recognition and ranging and angle measurement error calibration, and the system includes:

[0121] A scene feature database generation module 801, configured to collect waveforms of location signals with annotations and real distance and angle information in an indoor positioning scene based on the indoor positioning scene, and construct a scene feature database;

[0122] A learning network generation module 802, configured to construct a training pool based on the scene feature database and train to generate a learning network;

[0123] A calibration module 803, configured to process through a variational autoencoder based on the generated learning network to obtain a distance codeword, an angle codeword, and a scene codeword, input the distance codeword, the angle codeword, and the scene codeword into corresponding calibrators, calibrate newly input ranging and angle measurement values, output a calibrated positioning result, and update the scene feature database at the same time.

[0124] The scene feature database generation module 801 selects the experimental environment and area of a specified scene, and the scenes are mainly distinguished by room types and internal partitions. The room types include but are not limited to corridors, offices, and meeting rooms; the internal partitions include but are not limited to glass, wood, sponge, metal, etc. Define scene labels And grid the scene Place base station A and target point B at any two grid points (p i , p j ) in the scene, i, j ∈ [1, N g , i ≠ j, record the ranging value, angle measurement value, waveform, and waveform characteristics between single-station A and target point B, and record the true distance and angle values between target point B and base station A until no new recognizable features are added.

[0125] In the label l scene, change the grid point positions of base station A or target B, continue to measure and record the ranging value, angle measurement value, waveform, and waveform characteristics between single-station A and target B, and record the true distance and angle values between target B and base station A until no new recognizable features are added;

[0126] Repeat the above steps until all grid point combinations (p i , p j ) are recorded, i, j ∈ [1, N g , i ≠ j;

[0127] Change the scenario (different room types or partitions), and define the scenario tags until the set of scenario tags covers the scenario types in the specified application environment. Specifically, it basically covers different room types or major material categories respectively.

[0128] The learning network generation module 802 includes: a variational autoencoder, a ranging error calibrator, an angle error calibrator, and a scenario recognizer. The variational autoencoder is used for latent variable encoding and signal reconstruction. The ranging error calibrator is used to calibrate the distance measurement results. The angle error calibrator is used to calibrate the angle measurement. The scenario recognizer is used to identify and classify the usage scenarios.

[0129] The process of learning network generation is: constructing a training pool, where the training pool contains two types of data sets, namely the labeled data set and the unlabeled data set Among them, the labeled data set is composed of the scenario feature database, while the unlabeled data set is composed of the high-quality results selected during the system working calibration process, where x (i) is waveform-related data (waveform features), Δd (i) is the ranging error, Δθ (i) is the angle measurement error, k (i) is the scenario tag. The waveform-related data includes the waveform and waveform features such as received signal energy, maximum amplitude, rise time, average excess delay, root mean square spread, kurtosis, etc.:

[0130] Received signal energy: ε r = ∫ T |r(t)| 2 dt

[0131] Maximum amplitude:

[0132] Rise time: t rise = min{t: |r(t)| ≥ βr max}-min{t: |r(t)| ≥ ασ n}

[0133] Average excess delay:

[0134] Root mean square spread:

[0135] Kurtosis:

[0136] The learning network is trained and generated, and the encoder f(·; φ) is defined: x → yd , y θ , y k , where the variational posterior probabilities q φ (y d , y θ , y k |x) of the three latent variables can be learned through the encoder f(·; φ); define the decoder g(·; ψ): y d , y θ , y k →x, where the likelihood distribution p ψ (x|y d , y θ , y k ) can be learned through the decoder g(·; ψ). Since the three latent variables are independent of each other, the encoding network can be decomposed into three sub-modules, the ranging error calibrator the angle measurement error calibrator and the scene recognizer Among them, the relevant asymptotic posterior probabilities can be learned through the corresponding modules:

[0137]

[0138]

[0139]

[0140] Based on the above variational distribution of the latent variables, the evidence lower bound (ELBO, evidence lower bound) for the data x in the training pool can be expressed as:

[0141]

[0142] For the labeled dataset and the unlabeled dataset in the training pool respectively, different reconstruction loss terms will be used when the autoencoder is generated:

[0143] For the unlabeled dataset with low confidence in the dataset the unsupervised loss term is constructed as:

[0144]

[0145] For the labeled dataset with high confidence in the dataset the supervised loss term is constructed as:

[0146]

[0147] Among them, L labelFor the parameters of the ranging error calibrator, angle error calibrator, and scene recognizer The constructed annotation data loss term is specifically expressed as:

[0148]

[0149] Calibration module 803, in the offline calibration stage, inputs the real-time observed channel impulse response (CIR) waveform data x ob , and the coding parameters φ * and ψ * obtained through training and learning based on the data set, and the parameters of the distance calibrator, angle calibrator, and scene recognizer

[0150] Based on the coding parameters φ * and ψ * , encode the observed data xob to generate the variable codeword y d , y θ , y k ;

[0151] Based on the parameter Input the distance codeword y d into the distance calibrator to obtain the distance deviation estimate Based on the parameter Input the angle codeword y θ into the angle calibrator to obtain the angle deviation estimate Based on the parameter Input the scene recognition codeword y k into the scene recognizer to obtain the scene estimate value

[0152] Output the estimation result And screen the non-standard data in the training pool, and update the scene feature database at the same time.

[0153] Figure 9 Illustrates a schematic diagram of the physical structure of an electronic device, as Figure 9 shown. The electronic device may include: a processor 910, a communication interface 920, a memory 930, and a communication bus 940. Among them, the processor 910, the communication interface 920, and the memory 930 complete mutual communication through the communication bus 940. The processor 910 can call the logical instructions in the memory 930 to execute the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration. The method includes: based on the indoor positioning scene, collecting the waveforms of the labeled positioning signals and the real distance and angle information in the indoor positioning scene, and constructing a scene feature database;

[0154] Construct a training pool based on the scene feature database and train to generate a learning network;

[0155] Based on the generated learning network, process it through a variational autoencoder to obtain distance codewords, angle codewords, and scene codewords, and input the distance codewords, angle codewords, and scene codewords into the corresponding calibrators to calibrate the newly input ranging and angle measurement values, output the calibrated positioning result, and update the scene feature database at the same time.

[0156] In addition, when the logical instructions in the above-mentioned memory 930 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0157] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration provided by the above-mentioned various methods. The method includes: based on the indoor positioning scene, collect the waveforms of the labeled positioning signals and the true distance and angle information in the indoor positioning scene, and construct a scene feature database;

[0158] Construct a training pool based on the scene feature database and train to generate a learning network;

[0159] Based on the generated learning network, process it through a variational autoencoder to obtain distance codewords, angle codewords, and scene codewords, and input the distance codewords, angle codewords, and scene codewords into the corresponding calibrators to calibrate the newly input ranging and angle measurement values, output the calibrated positioning result, and update the scene feature database at the same time.

[0160] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements a single-station positioning method for joint scene recognition and ranging and angle measurement error calibration provided by the above-mentioned various methods. The method includes: based on the indoor positioning scene, collecting the waveforms of the marked positioning signals and the real distance and angle information in the indoor positioning scene, and constructing a scene feature database;

[0161] Constructing a training pool based on the scene feature database and training to generate a learning network;

[0162] Based on the generated learning network, processing through a variational autoencoder to obtain a distance codeword, an angle codeword, and a scene codeword, and inputting the distance codeword, the angle codeword, and the scene codeword into the corresponding calibrator to calibrate the newly input ranging and angle measurement values, outputting the calibrated positioning result, and simultaneously updating the scene feature database.

[0163] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0164] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A single-station positioning method combining scene recognition and ranging and angle measurement error calibration, characterized in that, Including: Based on the indoor positioning scenario, collect the waveforms of the labeled positioning signals and the real distance and angle information in the indoor positioning scenario, and construct a scene feature database; Construct a training pool based on the scene feature database and train to generate a learning network; Based on the generated learning network, process it through a variational autoencoder to obtain distance codewords, angle codewords, and scene codewords, and input the distance codewords, angle codewords, and scene codewords into the corresponding calibrators to calibrate the newly input ranging and angle measurement values, output the calibrated positioning results, and update the scene feature database at the same time; Among them, the processing based on the generated learning network through a variational autoencoder to obtain distance codewords, angle codewords, and scene codewords, and inputting the distance codewords, angle codewords, and scene codewords into the corresponding calibrators specifically includes: Input the waveform data x into the variational autoencoder to generate three Gaussian latent variable representations, including: distance codeword y d , angle codeword y θ , scene environment codeword y k ; Based on the unsupervised loss term, substitute the data in the unlabeled dataset into the variational autoencoder for training to learn its latent variable mapping relationship with the labeled dataset; Based on the supervised loss term, the distance codeword y generated from the labeled data d , the angle codeword y θ , and the scene environment codeword y k are substituted into the corresponding distance estimator, angle estimator, and scene recognizer for training; Based on the overall loss function, jointly optimize the labeled and unlabeled data through iteration to output ranging, angle measurement, and scene parameter estimation.

2. The single-station positioning method for joint scenario recognition and ranging and angle measurement error calibration according to claim 1, characterized in that The collection of the waveforms of the labeled positioning signals and the real distance and angle information in the indoor positioning scenario based on the indoor positioning scenario to construct a scene feature database specifically includes: Select the environment and area of the scene. The scene is distinguished by room type and internal partitions, define the scene label, and grid the scene. There are base stations and target points in the scene. Measure the ranging value, angle measurement value, waveform, waveform feature, real distance, and angle value between the base station and the target point until no new recognizable features are added; Change the grid point positions of the base station and the target point, measure the ranging value, angle measurement value, waveform, waveform feature, real distance, and angle value between the base station and the target point until no new recognizable features are added, and measure all grid point positions; Change the scene and repeat the grid point position measurement until all scenes in the application environment are covered.

3. The single-station positioning method for joint scene recognition and ranging and angle measurement error calibration according to claim 1, wherein The construction of a training pool based on the scene feature database and the training to generate a learning network specifically includes: Construct a training pool, which includes an annotated data set and a non-standard data set The annotated data set includes a scene feature database, and the non-standard data set includes high-quality results selected during the work calibration process, where x (i) is waveform feature data, Δd (i) is the ranging error, and Δθ (i) is the angle measurement error, and k (i) is the scene label; Train the learning network using the data in the training pool, and for the labeled data sets in the training pool respectively Construct the supervised loss term, and for the unlabeled data sets in the training pool Construct the unsupervised loss term. Through the supervised loss term and the unsupervised loss term, construct the overall network loss function for the training pool to generate the learning network.

4. The single-station positioning method for joint scene recognition and ranging and angle measurement error calibration according to claim 1, characterized in that, The calibration of the newly input ranging and angle measurement values to output the calibrated positioning results and update the scene feature database at the same time specifically includes: The online training stage and the offline calibration stage, In the online training stage, input the training pool dataset, learning rate, batch processing length, initial encoding parameters, initial decoding parameters, and initial sub-module parameters; Extract batch samples from the training pool for training, and use the gradient descent algorithm to train the network encoder and network decoder; Repeat the training process until convergence and output the learning results.

5. The single-station positioning method for joint scene recognition and ranging and angle measurement error calibration according to claim 4, characterized in that, The calibration of the newly input ranging and angle measurement values to output the calibrated positioning results and update the scene feature database at the same time specifically includes: In the offline calibration stage, input the real-time observed channel impulse response waveform data and the encoding parameters, distance calibrator parameters, angle calibrator parameters, and scene recognizer parameters obtained by training and learning based on the dataset; Based on the encoding parameters, encode the observed data to generate distance codewords, angle codewords, and scene environment codewords; The distance codeword y d is input into a distance calibrator to obtain a distance deviation estimate value; Input the angle codeword y θ into the angle calibrator to obtain an estimated value of the angle deviation; Input the scene recognition codeword y k into the scene recognizer to obtain a scene estimate value; Output the estimated values of distance deviation, angle deviation, and scene, screen the non-standard data in the training pool, and update the scene feature database.

6. A single-station positioning system for joint scene recognition and ranging and angle measurement error calibration, characterized in that, The system includes: A scene feature database generation module, configured to construct a scene feature database by acquiring the waveforms of the labeled positioning signals and the true distance and angle information in the target scene; A learning network generation module, configured to construct a training pool based on the scene feature database and train to generate a learning network; A calibration module, configured to process through a variational autoencoder based on the generated learning network to obtain distance codewords, angle codewords, and scene codewords, input the distance codewords, angle codewords, and scene codewords into the corresponding calibrators, calibrate the newly input ranging and angle measurement values, output the calibrated positioning results, and update the scene feature database at the same time; Among them, the processing through the variational autoencoder based on the generated learning network to obtain distance codewords, angle codewords, and scene codewords, and input the distance codewords, angle codewords, and scene codewords into the corresponding calibrators specifically includes: Input the waveform data x into the variational autoencoder to generate three Gaussian latent variable representations, including: distance codeword y d , angle codeword y θ , scene environment codeword y k ; Based on the unsupervised loss term, substitute the data in the unlabeled dataset into the variational autoencoder for training to learn its latent variable mapping relationship with the labeled dataset; Based on the supervised loss term, the distance codeword y generated from the labeled data d , the angle codeword y θ , and the scene environment codeword y k are substituted into the corresponding distance estimator, angle estimator, and scene recognizer for training; Based on the overall loss function, jointly optimize the labeled and unlabeled data through iteration, and output the estimation of ranging, angle measurement, and scene parameters.

7. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration according to any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration according to any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the single-station positioning method for joint scene recognition and ranging and angle measurement error calibration according to any one of claims 1 to 5.

Images