Method and apparatus for position parameter estimation

Abstract

The application provides a position parameter estimation method and device, comprising: obtaining a semi-supervised data set and a received waveform at a target time; inputting the semi-supervised data set and the received waveform into a pre-constructed semi-supervised position parameter estimation model to obtain a distribution inference result; wherein the semi-supervised position parameter estimation model is trained based on a deep neural network using a sample data set using a variational inference theory, and at least comprises an encoder and a regressor. The semi-supervised position parameter estimation model provided by the application has the ability of active learning in inference, can effectively utilize the semi-supervised data set for parameter inference, can significantly reduce the communication, calculation and data set acquisition costs generated by the deployment of the machine learning algorithm in the positioning system, and realizes higher precision and higher generalization of the position parameter estimation.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and apparatus for estimating position parameters. Background Technology

[0002] Current mainstream machine learning location parameter estimation schemes mainly consist of three parts: an encoder, a decoder, and a regressor, or only an encoder and a regressor. The encoder primarily receives signal waveform data from the base station and encodes the high-dimensional signal into low-dimensional feature vectors to filter and normalize valuable signal features, improving the efficiency and effectiveness of feature utilization. The decoder is usually used in conjunction with the encoder to remap the low-dimensional feature vectors back to a high-dimensional space as a signal estimation tensor, thereby optimizing the encoder's effectiveness. The regressor, based on various machine learning algorithms, learns the mapping relationship between signal features and location parameters and performs inference tasks in the same data domain, estimating the location parameters based on the input signal.

[0003] However, existing machine learning location parameter estimation algorithms exhibit poor generalization due to the difference between the input data distribution and the training set data distribution. In new positioning scenarios, the accuracy can plummet due to the discrepancy between the input data distribution and the training set data distribution. This difference in data distribution mainly stems from two sources: differences in the high-dimensional waveform distribution of the input, and differences in the output parameter distribution under approximate input waveform conditions. Furthermore, while traditional machine learning algorithms can learn the differences in the high-dimensional waveform distribution of the input through joint learning across multiple scenarios, they struggle to estimate the specific form of the output parameter distribution under approximate input conditions due to a lack of prior information, leading to a severe decrease in accuracy.

[0004] In summary, existing technologies suffer from insufficient accuracy and poor generalization. Summary of the Invention

[0005] This invention provides a method and apparatus for estimating position parameters to address the shortcomings of insufficient accuracy and poor generalization in existing technologies, thereby achieving position parameter estimation with higher accuracy and better generalization.

[0006] This invention provides a method for estimating location parameters, comprising:

[0007] Obtain the received waveforms at the semi-supervised dataset and the target time.

[0008] The semi-supervised dataset and the received waveform are input into a pre-built semi-supervised position parameter estimation model to obtain the distribution inference results;

[0009] The semi-supervised position parameter estimation model is based on a deep neural network trained using variational inference theory on a sample dataset, and includes at least an encoder and a regressor.

[0010] According to a location parameter estimation method provided by the present invention, the semi-supervised dataset and the received waveform are input into a pre-built semi-supervised location parameter estimation model to obtain distribution inference results, specifically including:

[0011] The semi-supervised dataset is dimensionality reduced and aggregated using an encoder to obtain inferred latent variables; wherein the encoder includes a supervised encoder and an unsupervised encoder;

[0012] The distribution inference results are obtained using a regressor based on the inferred latent variables.

[0013] According to the location parameter estimation method provided by the present invention, a semi-supervised location parameter estimation model is obtained by training a deep neural network using a sample dataset and variational inference theory, specifically including:

[0014] Construct a sample dataset;

[0015] Based on the aforementioned sample dataset, waveform sample data pairs are obtained; wherein, the sample dataset includes at least a large number of waveform sample data pairs, and each waveform sample data pair includes waveform signal samples and location labels;

[0016] The waveform sample data pairs are inferred by jointly using an unsupervised encoder and a supervised encoder to obtain a first latent variable distribution and a second latent variable distribution; wherein, the first latent variable distribution is obtained by dimensionality reduction of the input waveform using an unsupervised encoder, and the second latent variable distribution is obtained by dimensionality reduction of the waveform sample data pairs using a supervised encoder; the input waveform is the waveform signal sample included in the waveform sample data pairs;

[0017] The first error, the second error, and the third error are calculated based on the first latent variable distribution and the second latent variable distribution.

[0018] The first and second latent variable distributions are aggregated using an aggregator to obtain the mapping latent variable estimate of the current environment;

[0019] The mapped latent variable estimate and the input waveform are input to a regressor to obtain the position parameter estimation result; the position parameter estimation result and the position label of the input waveform are compared to obtain the fourth error; wherein, the input waveform is a waveform signal sample included in the waveform sample data pair;

[0020] Calculate the lower bound of evidence for the location parameter estimation result; wherein the lower bound of evidence is calculated based on a first preset formula using a first error, a second error, a third error, and a fourth error;

[0021] Gradient descent is performed based on the lower bound of the evidence to obtain a semi-supervised location parameter estimation model.

[0022] According to a location parameter estimation method provided by the present invention, a first error, a second error, and a third error are calculated based on a first latent variable distribution and a second latent variable distribution, specifically including:

[0023] The waveform distribution is reconstructed based on the first latent variable distribution to obtain the reconstructed waveform;

[0024] A first error is calculated based on the reconstructed waveform and the input waveform, and a second error is calculated based on the distribution of the first latent variable.

[0025] The KL divergence between the second latent variable distribution and the first latent variable distribution is measured and summed to obtain the third error.

[0026] According to a location parameter estimation method provided by the present invention, the first preset formula includes:

[0027]

[0028] in, This represents the inherent likelihood relationship of the sample dataset that the semi-supervised location parameter estimation model can describe; Indicates the location label in the sample dataset; This represents the waveform signal sample corresponding to the position label in the sample dataset; This indicates that the sample data is concentrated in M ​​environments, with S unsupervised waveform signal samples collected in each environment; ELBO NPR,m This represents the lower bound of evidence related to NPR in the m-th environment. ELBO VAE,m This represents the lower bound of evidence related to VAE in the m-th environment. Indicates the third error; This represents the aggregated result of the distributions of the N second latent variables in the m-th environment; z represents the variational inference of the posterior distribution of latent variables in the m-th environment using unsupervised sample data; m This represents the hidden variable in the m-th environment; This represents the N location labels collected in the m-th environment of the sample dataset; This represents the waveform signal sample corresponding to the N location labels collected in the m-th environment of the sample dataset; Indicates the fourth error; This represents the estimation of latent variables in mapping; Let f(x) represent the likelihood of N waveform signal sample data with location labels in the m-th environment, assuming the presence of latent variables. This represents the S unlabeled waveform signal samples in the m-th environment of the sample dataset; Indicates the second error; f represents the aggregated result of all distributions of the first latent variable in the m-th environment; θ (z m ) represents the prior of the hidden variable; Indicates the first error; This represents the aggregated result of all reconstructed waveforms in the m-th environment.

[0029] According to a location parameter estimation method provided by the present invention, constructing a sample dataset specifically includes:

[0030] Collect massive amounts of waveform signal samples;

[0031] S1: Extract a waveform signal sample, locate the preset coordinates of the waveform signal sample in the current coordinate system, and obtain the position label of the waveform signal sample;

[0032] S2: The location tag and the waveform signal sample form a waveform sample data pair;

[0033] Repeat steps S1-S2 to obtain a large number of waveform sample data pairs, and use the large number of waveform sample data pairs to construct a sample dataset.

[0034] The present invention also provides a position parameter estimation device, comprising:

[0035] The acquisition unit is used to acquire the received waveforms at the semi-supervised dataset and the target time.

[0036] The inference unit is used to input the semi-supervised dataset and the received waveform into a pre-built semi-supervised position parameter estimation model to obtain the distribution inference result;

[0037] The semi-supervised position parameter estimation model is based on a deep neural network trained using variational inference theory on a sample dataset, and includes at least an encoder and a regressor.

[0038] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the position parameter estimation method as described above.

[0039] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the position parameter estimation method as described above.

[0040] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the position parameter estimation method as described above.

[0041] This invention provides a method and apparatus for estimating location parameters. The method involves acquiring a semi-supervised dataset and the received waveform at a target time. The semi-supervised dataset and the received waveform are then input into a pre-constructed semi-supervised location parameter estimation model to obtain a distribution inference result. The semi-supervised location parameter estimation model is trained using variational inference theory on a sample dataset based on a deep neural network. The model includes at least an encoder and a regressor. The semi-supervised location parameter estimation model provided by this invention has the ability to actively learn during inference and can effectively utilize semi-supervised datasets for parameter inference. This significantly reduces the communication, computational, and dataset acquisition costs associated with deploying machine learning algorithms in positioning systems, achieving higher accuracy and generalization in location parameter estimation. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0043] Figure 1 This is one of the flowcharts illustrating the location parameter estimation method provided by the present invention;

[0044] Figure 2 This is the second flowchart illustrating the location parameter estimation method provided by the present invention;

[0045] Figure 3 This is the third flowchart illustrating the location parameter estimation method provided by this invention;

[0046] Figure 4 This is a schematic diagram of the network structure of a semi-supervised position parameter estimation model according to an embodiment of the position parameter estimation method provided by the present invention;

[0047] Figure 5 This is a schematic diagram of the position parameter estimation device provided by the present invention;

[0048] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0049] Figure label:

[0050] 510: Acquisition unit; 520: Inference unit;

[0051] 610: Processor; 620: Communication interface; 630: Memory; 640: Communication bus. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0053] Due to the difference between the distribution of the input data and the distribution of the data in the training set, existing machine learning location parameter estimation algorithms suffer from insufficient accuracy and poor generalization.

[0054] Based on this, considering that neural processes can encode supervised datasets and obtain complex input-output mapping relationships, thus being applicable to various complex data distributions, neural processes can be used to improve existing machine learning location parameter estimation algorithms.

[0055] Current mainstream neural process algorithms mainly consist of three parts: an encoder, an aggregator, and a regressor. The encoder performs encoding and dimensionality reduction on the labeled dataset. The aggregator reduces the dimensionality of multiple encoding results obtained by the encoder on the dataset to form mapping latent variables, inferring the mapping relationship between input and output from the labeled dataset. The regressor takes the mapping latent variables and waveforms as input, and after obtaining the mapping relationship defined by the mapping latent variables, uses this mapping relationship to infer the position parameter estimation results based on the input waveform. However, neural processes lack flexibility and rely too heavily on labeled data to generate mapping latent variables, thus limiting their application and deployment.

[0056] Based on this, the present invention proposes a location parameter estimation method and apparatus to improve the problems of insufficient accuracy, insufficient generalization, and dependence on labeled datasets when combined with neural process algorithms, which are caused by traditional machine learning parameter estimation algorithms.

[0057] The following is combined with Figures 1-4 The location parameter estimation method of the present invention is described. Figure 1 This is one of the flowcharts illustrating the location parameter estimation method provided by the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0058] Step 110: Obtain the semi-supervised dataset and the received waveform at the target time.

[0059] Semi-supervised datasets include datasets that contain both labeled and unlabeled waveform data. For ease of understanding, semi-supervised datasets can be divided into two types: supervised datasets. and unsupervised datasets Supervised datasets include waveform data with location labels from semi-supervised datasets; unsupervised datasets include waveform data without location labels from semi-supervised datasets. It should be noted that in some embodiments, unsupervised datasets may also include waveform data from supervised datasets after the location labels have been removed.

[0060] In one embodiment, for the new environment m new Obtain a semi-supervised dataset and Received waveform at the target time (denoted as time k) (CIR observation) Its posterior distribution is:

[0061]

[0062] in, This represents the posterior distribution for inferring the position of the waveform in a new environment where a semi-supervised dataset exists. This indicates the position information corresponding to the k-th waveform in the new environment; This represents the k-th waveform in the new environment; This represents a labeled dataset of size N in a new environment; This represents an unlabeled dataset of size S in the new environment; This represents the likelihood function of location information in a new environment under latent variable constraints generated by dimensionality reduction of a semi-supervised dataset. This represents the posterior distribution of latent variables based on a semi-supervised dataset; This represents the likelihood function under a semi-supervised location parameter estimation model. This represents a latent variable in the new environment.

[0063] It should be understood that, in order to better understand this technical solution, the positioning method and system involved in the position parameter estimation method provided by this invention are described using ultra-wideband signals as an example, but this does not mean that the position parameter estimation method provided by this invention is limited.

[0064] Step 120: Input the semi-supervised dataset and the received waveform into the pre-built semi-supervised position parameter estimation model to obtain the distribution inference result;

[0065] The semi-supervised position parameter estimation model is based on a deep neural network trained using variational inference theory on a sample dataset, and includes at least an encoder and a regressor.

[0066] Obtain a semi-supervised dataset and and the received waveform at the target time The data is then fed into a semi-supervised location parameter estimation model, which includes at least an encoder and a regressor. The encoder can be either supervised or unsupervised. The semi-supervised dataset is then processed through the encoder. Dimensionality reduction and aggregation are performed to complete the inference of latent variables and obtain the inferred latent variables. Based on this, parameter estimation is performed on the received waveform, and the latent variables are inferred using a regressor to obtain the distribution inference results:

[0067]

[0068] in, This represents the variational inference result based on the posterior distribution of semi-supervised latent variables, the likelihood of user location and waveform, i.e., the distribution inference result; This represents variational inference of the posterior distribution of latent variables based on a semi-supervised dataset, i.e., inference of latent variables; This represents the posterior distribution fitted by the neural network based on waveform and latent variable location information; This represents a supervised dataset within a semi-supervised dataset. This represents the unsupervised dataset within the semi-supervised dataset.

[0069] The above process can be represented as follows: Figure 2 As shown. Figure 2 As shown, specifically, unsupervised data is sequentially extracted from the semi-supervised dataset and input into the parameters. In the defined unsupervised encoder, the latent variables of the unsupervised data are obtained, and the supervised data is extracted and input into the parameters. In the defined supervised encoder, the latent variables of the supervised data are obtained, and the two latent variables are aggregated to obtain z. e , will z e and received waveform The input is fed into the regressor defined by parameter θ to obtain the distribution inference results.

[0070] This indicates a supervised encoder; θ represents an unsupervised encoder; θ represents a regressor.

[0071] The location parameter estimation method provided by this invention, after training and deployment, can directly encode the semi-supervised dataset for the new environment during the inference process, and directly initiate the user's location parameter estimation process based on the encoding results. From a communication resource perspective, the location parameter estimation method provided by this invention does not require sending the dataset back to the server for training, decoupling the communication coupling between the model and the server during the generalization process and saving communication resources. From a computational perspective, in the new environment after training, the location parameter estimation method provided by this invention does not need to be trained in the new environment, only the inference process needs to be executed, thus relaxing the requirements for terminal computing power. The location parameter estimation method provided by this invention has the ability to generalize to new scenarios under conditions of limited communication and computing resources.

[0072] In some embodiments of the present invention, the semi-supervised location parameter estimation model is obtained by training a sample dataset using variational inference theory based on a deep neural network.

[0073] Specifically, assume a single-base station ultra-wideband positioning system includes a base station with a P-antenna and user equipment (UE). The center position of the base station array is known, denoted as a = {a x ,a y} T At time k, in the m-th environment, the unknown UE position is defined as p. m,k Consider an environment with L unknown reflective surfaces in the m-th environment. The parameters of the l-th reflective surface relative to the base station are described using the image of the base station relative to the l-th reflective surface (defined as a virtual anchor (VA)). The position of the l-th VA, unknown at time k, in environment m is denoted as... and For the sake of model simplicity, we assume that the 0th VA is the known base station location, i.e., a. m,0,k =a. Since the energy of second-order reflections is very low, only the L first-order multipath components (MPCs) with the strongest energy of all first-order reflections are discussed. This is achieved through a virtual base station a. m,l,k Given the locations of base station a and UE, the location p of the virtual tag can be calculated. m,l,k According to Fresnel's theorem, all MPC signals received by the base station can be considered as virtual tags p. m,l,k The transmitted signal. Therefore, at each time k, p can be estimated by the received waveform. m,0:L,k Dually, the UE position p can be estimated from the waveform. m,k and all virtual base station VA locations a m,0:L,k For the sake of simplicity in the motion model, the estimated UE position p is used. m,kand all virtual base station VA locations a m,1:L,k The proposed scheme connects all the unknown variables at time k and represents them as x. m,k ,Right now:

[0074]

[0075] After receiving the UWB waveform transmitted by the UE, the base station can record each received CIR segment as Y. m,k Its signal model is as follows:

[0076]

[0077] Among them, g l It is non-line-of-sight phase distortion, T NLoS,l It is the non-line-of-sight ranging error (NLoS) of the l-th virtual base station, s l (τ-T l -T NLoS,l ) is the waveform distortion caused by the superposition of non-line-of-sight diffuse reflections, a(θ) l ) represents the guide vector. T represents the Hadama product, τ represents time. l Let represent the minimum time delay of the l-th multipath cluster, and let n(τ) represent a Gaussian noise random process. In the line-of-sight case, g l It is a unit vector, T NLoS =0, waveform distortion does not exist, s l (τ) = s(τ). Assuming the LosS (line-of-sight) scene is the 0th environment, in the LosS scene, one can directly start from Y... 0,k The ToA and AoA parameters of each MPC are estimated, and then Y is obtained. 0,k and x 0,k The likelihood between them. That is, f(Y) m,k |x m,k (m=0) is known.

[0078] However, in non-line-of-sight environments where m≠0, due to g l T NLoS , and s l (τ) are all unknowns related to the environment, therefore the likelihood function f(Y) m,k |x m,k The value of f(Y) is unknown. At this point, a machine learning-based approach is needed to determine the value of f(Y). m,k |x m,k Fitting.

[0079] Given a supervised dataset A neural network f defined by parameter ω can be trained.ω (Y k |x k To fit the observation model f(Y) m,k |x m,k (m = m0). For parameter ω in the dataset The process of performing maximum likelihood estimation is as follows:

[0080]

[0081] However, by The defined likelihood function lacks generalization. When the scenario changes, the likelihood function... The accuracy drops significantly and it lacks generalization ability.

[0082] The goal of the position parameter estimation method provided by this invention is to enable the estimation of position parameters in a new environment m after training is completed. new In this case, semi-supervised datasets are used directly. and waveform To infer the distribution

[0083] Based on this, neural processes are introduced. Neural processes (NPs) possess the ability to adapt to new environments. new The ability to directly absorb knowledge from the dataset and complete the inference process. NP believes that there is a latent variable. Capable of representing the environment in m and The mapping relationship between them. Latent variables. prior distribution It is a standard Gaussian distribution. Based on this, the posterior distribution can be... Represented as:

[0084]

[0085] in, This represents the posterior distribution of the position corresponding to the k-th waveform in the new environment under the constraints of a semi-supervised dataset. This indicates the position corresponding to the k-th waveform in the new environment; This represents the k-th waveform in the new environment; This represents a supervised dataset in a new environment; This represents an unsupervised dataset in a new environment; The mean of the posterior distribution of the new environment's location information at time k under the constraints of latent variables inferred from the waveform and semi-supervised dataset is represented. Represents the posterior distribution of latent variables inferred from waveforms and semi-supervised datasets; This represents the posterior distribution mean of the location information at time k in the new environment under the latent variable constraints inferred from the semi-supervised dataset. This represents the posterior distribution of latent variables inferred from a semi-supervised dataset; This represents the posterior distribution of the location information at time k in the new environment under latent variable constraints. This represents the latent variables in the new environment.

[0086] Furthermore, the posterior distribution of the latent variables is:

[0087]

[0088] in, This represents the posterior distribution of latent variables inferred from a supervised dataset; This represents the posterior distribution of latent variables inferred from an unsupervised dataset; This represents the prior distribution of the latent variable.

[0089] Therefore, as Figure 3 The inference of position parameter estimation is to be completed through steps S101-S103. The training concept for the semi-supervised position parameter estimation model is as follows:

[0090] S101: Supervised and unsupervised encoders jointly infer mapped latent variables and obtain estimates of the current environment's mapped latent variables through an aggregator. Specifically, in any environment m, the parameters are used... and Defined neural network distribution and Variational inference of distribution and Furthermore estimate

[0091] S102: Based on the mapping relationship between the latent variables and the input waveform, the neural process regressor outputs the estimated location parameters. Specifically, in any environment m, a neural network distribution f defined by parameters θ is used. θ (x m,k |z m ,Y m,k To fit f(x) m,k |z m ,Y m,k ).

[0092] S103: Calculate the cost function and perform gradient descent on the estimation results of the latent variables and location parameters of the mapping relationship.

[0093] To train modules S101 and S102 in step S103, this invention derives a semi-supervised multi-scene sample dataset. and The defined evidence lower bound (ELBO) for neural processes. By applying gradient descent to this evidence lower bound, neural networks can achieve the performance described in S101 and S102.

[0094] Based on this, the semi-supervised location parameter estimation model is obtained by training a deep neural network using a sample dataset and variational inference theory, specifically including:

[0095] S210: Construct the sample dataset. The sample dataset is a pre-constructed semi-supervised multi-scene dataset, which includes at least a large number of waveform sample data pairs. Each waveform sample data pair includes waveform signal samples and location labels. In some embodiments, the sample dataset may also include multiple unlabeled waveform sample data and a large number of waveform sample data pairs.

[0096] Preferably, constructing a sample dataset includes the following steps:

[0097] Collect massive amounts of waveform signal samples. First, acquire massive amounts of unlabeled waveform sample data Y. m,n It should be noted that the unlabeled waveform sample data Y... m,n It can collect any amount of data.

[0098] S1: Extract a waveform signal sample and locate the preset coordinates of the waveform signal sample in the current coordinate system to obtain the position label of the waveform signal sample. Next, obtain the unlabeled waveform sample data Y. m,n Location labels. It's important to note that location labels are used to annotate waveform sample data. The labels include location-related information and can be unlabeled waveform sample data Y. m,n The preset coordinates. In some embodiments, the location label may include unlabeled waveform sample data Y in the current coordinate system. m,n The location labels are two-dimensional Cartesian coordinates; that is, the preset coordinates can be two-dimensional Cartesian coordinates. The location labels can be obtained through manual on-site surveying during dataset collection, or through other positioning technologies. This invention does not limit the scope of the invention.

[0099] S2: The location label and the waveform signal sample form a waveform sample data pair. Then, the unlabeled waveform sample data Y... m,n The waveform sample data pairs consist of the waveform sample data pairs and their location labels.

[0100] Repeat steps S1-S2 to obtain a large number of waveform sample data pairs, and use the large number of waveform sample data pairs to construct a sample dataset.

[0101] Furthermore, in some embodiments, the sample dataset can be divided into a first sample dataset and a second sample dataset; the first sample dataset includes only waveform signal samples, and the second sample dataset includes waveform sample data pairs. For ease of description, the first sample dataset can also be called a supervised sample dataset, and the second sample dataset can also be called an unsupervised sample dataset. It should be noted that the number of waveform signal samples in the supervised sample dataset is not necessarily the same as the number of waveform signal samples in the unsupervised sample dataset. For example, the first sample dataset... This represents a supervised dataset with N data points across M environments; the second sample dataset. This represents an unsupervised dataset with a data volume of S across M environments. Further, in some embodiments, the supervised sample dataset includes waveform sample data pairs with location labels from the sample dataset; the unsupervised sample dataset includes waveform signal samples without location labels from the sample dataset.

[0102] S220: Obtain waveform sample data pairs based on the sample dataset.

[0103] Waveform sample data pairs are extracted from the sample dataset for subsequent training. More specifically, this can be understood as extracting a waveform sample data pair (Y) from the sample dataset. m,n ,x m,n Then extract waveform samples Y from them. m,n , using Y m,n With (Y) m,n ,x m,n Simultaneously, follow-up training will be conducted.

[0104] S230: The waveform sample data pairs are jointly inferred using an unsupervised encoder and a supervised encoder to obtain a first latent variable distribution and a second latent variable distribution; wherein, the first latent variable distribution is obtained by using the unsupervised encoder on the unsupervised sample dataset. The second latent variable distribution is obtained by dimensionality reduction of the input waveform in the supervised sample dataset using a supervised encoder. The waveform sample data is obtained by dimensionality reduction; the input waveform is the waveform signal sample included in the waveform sample data pair.

[0105] Specifically, such as Figure 4 , Figure 4 This illustrates the process of training using only a supervised sample dataset. From the supervised sample dataset... Obtain waveform sample data pairs (Y) m,n ,x m,n Then, an unsupervised encoder is used to process the waveform sample signal Y. m,n Dimensionality reduction yields the distribution of the first latent variable. Then, aggregate all the distributions of the first latent variable in the supervised dataset to obtain... Among them, the unsupervised encoder is a parameter-based encoder. Defined encoder (i.e. Figure 4 Winning bid The trapezoidal structure is a standard variational autoencoder (VAE). Simultaneously, a supervised encoder is used to process waveform sample data pairs (Y...). m,n ,x m,n Dimensionality reduction yields the distribution of the second latent variable. Then, aggregate all the distributions of the second latent variable in the supervised dataset to obtain... Among them, the supervised encoder is a function consisting of parameters. Defined encoder (i.e. Figure 4 Winning bid (A trapezoidal structure). There is a supervised encoder (i.e.,...) Figure 4 Winning bid (trapezoidal structure) and regressor (i.e. Figure 4 The square structure marked with θ together constitutes the NP part.

[0106] For unsupervised datasets The sample data in the sample signal is then repeatedly processed using an unsupervised encoder to analyze the waveform sample signal Y. m,n Dimensionality reduction yields the distribution of the first latent variable. Aggregate all the distributions of the first latent variable in the unsupervised dataset to obtain Then, aggregate all the distributions of the first latent variables in the supervised and unsupervised datasets to obtain... The above process of training using unsupervised sample datasets and supervised sample datasets is carried out simultaneously.

[0107] S240: The first error, the second error, and the third error are calculated based on the first latent variable distribution and the second latent variable distribution.

[0108] Specifically, the waveform distribution is first reconstructed based on the distribution of the first latent variable to obtain the reconstructed waveform. And compare the waveform sample signal Y m,n With reconstructed waveform The error is calculated by summing the errors of all reconstructed waveforms, thus obtaining the first error: in, This represents the result of variational inference of the posterior distribution of latent variables based on waveforms from supervised and unsupervised sample datasets, i.e., the result of variational inference of the posterior distribution of latent variables using waveform signal samples from the sample dataset. z represents the likelihood function of supervised and unsupervised waveforms under a semi-supervised position parameter estimation model, i.e., the likelihood function of the reconstructed waveform; mRepresents hidden variables; This represents the waveform in the supervised sample dataset in environment m; This represents the waveform in the unsupervised sample dataset in environment m. It's important to note that reconstructing the waveform distribution based on the first latent variable distribution involves: reconstructing the waveform distribution using an encoder defined by θ based on the first latent variable distribution, and using another neural network (i.e.,...) Figure 4 The waveform is reconstructed by taking the latent variable as input and the waveform as output (a trapezoidal structure labeled θ).

[0109] Secondly, for each first latent variable distribution Constraints are applied to all of them, the KL divergence between them and the standard Gaussian distribution is measured, and these KL divergences are summed to obtain the second error:

[0110] Then, based on the obtained latent variable distribution That is, the second latent variable distribution, which is the dimensionality reduction distribution of the unlabeled latent variables for each set of sample data (the first latent variable distribution). The KL divergence is passed over and measured between the latent variables of labeled and unlabeled data, and then summed. In other words, the KL divergence between the first and second latent variable distributions is measured and summed to obtain the third error. in, This represents the variational inference results of the posterior distribution of latent variables using a supervised sample dataset; This represents the result of variational inference of the posterior distribution of the latent variables using waveforms from a supervised sample dataset, which is related to the distribution of the first latent variable. The relationship is a merging relationship, i.e., the distribution of the first latent variable. yes and The aggregation result; z m Represents hidden variables; The waveform represents a supervised sample dataset; This refers to the labeled data of the supervised sample dataset. No unsupervised data needs to be added in this step.

[0111] It is important to note that and and and and as well as With All relationships can be understood as merge relationships. First latent variable distribution. It is a single waveform sample Y m,n The latent variable distributions are obtained by aggregating the latent variable distributions of all waveform samples in the sample dataset. That is, the distribution of the first latent variable of all waveform signal samples in the sample dataset. yes and The aggregation result. A single waveform sample Y m,n The reconstructed waveform is The reconstructed waveforms are obtained by aggregating all waveform samples in the sample dataset. Second latent variable distribution A single data-label pair (Y) in a supervised sample dataset m,n ,x m,n The latent variable distribution of a waveform is obtained by aggregating the latent variable distributions of all waveform samples (N) in the supervised dataset.

[0112] It's important to understand that this training step brings the supervised latent variable distribution closer to the target distribution. and unsupervised latent variable distribution The distance between them allows the latent variables inferred by VAE to represent, to some extent, the latent variables z in the environment. m This information gives neural networks the ability to perform semi-supervised learning.

[0113] S250: Use an aggregator to aggregate the first latent variable distribution and the second latent variable distribution to obtain the mapping latent variable estimate of the current environment.

[0114] Specifically, the aggregator is used to aggregate all the generated latent variables (all distributions of the second latent variables) from the supervised data. ) and all latent variables generated from unsupervised data (all first latent variable distributions, Obtain the mapping latent variable estimate Complete the inference of latent variables.

[0115] S260: Input the mapped latent variable estimate and the input waveform to the regressor to obtain the position parameter estimation result; compare the position parameter estimation result with the position label of the input waveform to obtain the fourth error; wherein, the input waveform is a waveform signal sample included in the waveform sample data pair.

[0116] Specifically, because aggregation of large-scale datasets optimizes the estimation of mapping latent variables. Until the variance is very small, therefore only the expectation is considered. Perform a Monte Carlo estimation, that is

[0117]

[0118]

[0119] Among them, the samples obtained As latent variables representing the mapping relationship of environment m, they are input into the regressor defined by θ. Therefore, the regressor (i.e., Figure 4 The square structure marked with θ has completed the mapping relationship inference in the m environment. Based on which can be given Output position parameters The estimation results are used as the location parameter estimation results. Based on the supervised data, the fourth error is calculated:

[0120] S270: Calculate the lower bound of evidence for the location parameter estimation result; wherein the lower bound of evidence is calculated based on a first preset formula using the first error, the second error, the third error, and the fourth error.

[0121] In summary, the marginal distribution of a sample dataset in the form of a semi-supervised dataset described by a neural network can be expressed by a first pre-defined formula:

[0122]

[0123] in, This represents the intrinsic likelihood of the sample dataset that the semi-supervised position parameter estimation model neural network can describe. The larger the likelihood, the better the neural network performs on the corresponding sample dataset. This represents the location labels in the sample dataset, specifically the N location labels collected for each of environments 1 to M, for a total of M environments, with N location labels for each environment. This represents the waveform signal sample corresponding to the location label in the sample dataset, that is, the supervised waveform signal sample collected in each environment from environment 1 to environment M, for a total of M environments, with N waveform signal samples in each environment; This indicates that the sample data is concentrated in M ​​environments, with S unsupervised waveform signal samples collected in each environment; ELBO NPR,m This represents the lower bound of evidence related to NPR in the m-th environment. ELBO VAE,m This represents the lower bound of evidence related to VAE in the m-th environment. Indicates the third error; This represents the aggregated result of the distributions of the N second latent variables in the m-th environment, which is the variational inference of the posterior distribution of the latent variables under supervised sample data. z represents the variational inference of the posterior distribution of latent variables in the m-th environment using unsupervised sample data; m This represents the hidden variable in the m-th environment; This represents the N location labels collected in the m-th environment of the sample dataset; This represents the waveform signal sample corresponding to the N location labels collected in the m-th environment of the sample dataset, i.e., the supervised waveform; Indicates the fourth error; This indicates that the expectation of latent variables is calculated using a sample dataset in a semi-supervised manner. This represents the estimation of latent variables in mapping; Let f(x) represent the likelihood of N waveform signal sample data with location labels in the m-th environment, assuming the presence of latent variables. This represents the S unlabeled waveform signal samples in the m-th environment of the sample dataset; Indicates the second error; f represents the aggregated result of all first latent variable distributions in the m-th environment, i.e., the posterior distribution of latent variables in the unsupervised case; θ (z m ) represents the prior of the hidden variable; Indicates the first error; This represents the aggregated result of all reconstructed waveforms in the m-th environment, i.e., the likelihood of the waveform in the unsupervised case.

[0124] S280: Perform gradient descent based on the aforementioned lower bound of evidence to obtain a semi-supervised location parameter estimation model. In one embodiment, the following can be derived from the above lower bound of evidence: Figure 4 The network structure, that is to say Figure 4 This is a schematic diagram of a network structure for a semi-supervised location parameter estimation model. In some embodiments, the neural network structure derived above is named semi-NP (semi-supervised neural process). After training, semi-NP can directly estimate the new environment m in a semi-supervised manner during inference through the design of the aggregator. new Latent variables in mapping relationships It also helps the regressor complete the parameter estimation task in the new environment.

[0125] The location parameter estimation method provided by this invention performs dimensionality reduction encoding on semi-supervised datasets based on both supervised and unsupervised encoders, maintaining the generalization ability of neural process algorithms while relaxing the tight coupling relationship with labeled data. This invention mainly includes: a latent variable inference and aggregation framework based on semi-supervised learning, enabling automatic learning and inference on the dataset and automatic encoding of signal-location mapping relationships; a semi-supervised neural process regressor based on the semi-supervised encoding results; and a multi-scenario joint semi-supervised training scheme based on semi-supervised neural processes, enabling the neural network to flexibly utilize semi-supervised datasets for joint training and inference.

[0126] Specifically, this invention derives a lower bound of evidence for location parameter inference using a neural network based on a semi-supervised dataset under conditions of limited terminal communication and computing resources, based on variational inference. Guided by this, a neural process-based network framework and corresponding semi-supervised training and inference algorithms are designed. The semi-supervised neural process proposed in this invention has the ability to actively learn during inference and can effectively utilize semi-supervised datasets for parameter inference, significantly reducing the communication, computing, and dataset acquisition costs associated with deploying machine learning algorithms in positioning systems.

[0127] The location parameter estimation method provided by this invention can improve the generalization performance of traditional neural networks through joint learning of multi-scene datasets. After training, the neural process can learn common knowledge in multiple environments and has flexible transfer capabilities. It can collect low-density datasets in any new scene and make direct inferences, making it suitable for generalized localization problems in new scenes under conditions of limited communication and computing resources.

[0128] This invention can be used in the field of machine learning-based positioning and navigation, and features semi-supervised sustainable learning. It should be noted that the signal systems described in the embodiments of this invention include, but are not limited to, 5G NR, WIFI, Bluetooth, geomagnetic sensing, and other signal sensing systems. It can also be used in application scenarios requiring positioning and navigation, such as smart factories and emergency search and rescue, where satellite rejection or weak signals are present.

[0129] This invention provides a location parameter estimation method that acquires a semi-supervised dataset and the received waveform at the target time. The semi-supervised dataset and the received waveform are then input into a pre-constructed semi-supervised location parameter estimation model to obtain a distribution inference result. The semi-supervised location parameter estimation model is trained using variational inference theory on a sample dataset based on a deep neural network, and includes at least an encoder and a regressor. The semi-supervised location parameter estimation model provided by this invention has the ability to actively learn during inference and can effectively utilize semi-supervised datasets for parameter inference. This significantly reduces the communication, computational, and dataset acquisition costs associated with deploying machine learning algorithms in positioning systems, achieving higher accuracy and generalization in location parameter estimation.

[0130] The position parameter estimation device provided by the present invention is described below. The position parameter estimation device described below can be referred to in correspondence with the position parameter estimation method described above. Figure 5 A schematic diagram of the position parameter estimation device provided by the present invention is shown below. Figure 5 As shown, it includes an acquisition unit 510 and an inference unit 520. Wherein,

[0131] The acquisition unit 510 is used to acquire the semi-supervised dataset and the received waveform at the target time.

[0132] The inference unit 520 is used to input the semi-supervised dataset and the received waveform into a pre-built semi-supervised position parameter estimation model to obtain the distribution inference result;

[0133] The semi-supervised position parameter estimation model is based on a deep neural network trained using variational inference theory on a sample dataset, and includes at least an encoder and a regressor.

[0134] Based on the above embodiments, in this device, the semi-supervised dataset and the received waveform are input into a pre-built semi-supervised position parameter estimation model to obtain distribution inference results, specifically including:

[0135] The semi-supervised dataset is dimensionality reduced and aggregated using an encoder to obtain inferred latent variables; wherein the encoder includes a supervised encoder and an unsupervised encoder;

[0136] The distribution inference results are obtained using a regressor based on the inferred latent variables.

[0137] Based on the above embodiments, in this device, the semi-supervised location parameter estimation model is trained using variational inference theory on a sample dataset based on a deep neural network, specifically including:

[0138] Construct a sample dataset;

[0139] Based on the aforementioned sample dataset, waveform sample data pairs are obtained; wherein, the sample dataset includes at least a large number of waveform sample data pairs, and each waveform sample data pair includes waveform signal samples and location labels;

[0140] The waveform sample data pairs are inferred by jointly using an unsupervised encoder and a supervised encoder to obtain a first latent variable distribution and a second latent variable distribution; wherein, the first latent variable distribution is obtained by dimensionality reduction of the input waveform using an unsupervised encoder, and the second latent variable distribution is obtained by dimensionality reduction of the waveform sample data pairs using a supervised encoder; the input waveform is the waveform signal sample included in the waveform sample data pairs;

[0141] The first error, the second error, and the third error are calculated based on the first latent variable distribution and the second latent variable distribution.

[0142] The first and second latent variable distributions are aggregated using an aggregator to obtain the mapping latent variable estimate of the current environment;

[0143] The mapped latent variable estimate and the input waveform are input to a regressor to obtain the position parameter estimation result; the position parameter estimation result and the position label of the input waveform are compared to obtain the fourth error; wherein, the input waveform is a waveform signal sample included in the waveform sample data pair;

[0144] Calculate the lower bound of evidence for the location parameter estimation result; wherein the lower bound of evidence is calculated based on a first preset formula using a first error, a second error, a third error, and a fourth error;

[0145] Gradient descent is performed based on the lower bound of the evidence to obtain a semi-supervised location parameter estimation model.

[0146] Based on the above embodiments, in this device, the first error, the second error, and the third error are calculated based on the first latent variable distribution and the second latent variable distribution, specifically including:

[0147] The waveform distribution is reconstructed based on the first latent variable distribution to obtain the reconstructed waveform;

[0148] A first error is calculated based on the reconstructed waveform and the input waveform, and a second error is calculated based on the distribution of the first latent variable.

[0149] The KL divergence between the second latent variable distribution and the first latent variable distribution is measured and summed to obtain the third error.

[0150] Based on the above embodiments, in this device, the first preset formula includes:

[0151]

[0152] in, This represents the inherent likelihood relationship of the sample dataset that the semi-supervised location parameter estimation model can describe; Indicates the location label in the sample dataset; This represents the waveform signal sample corresponding to the position label in the sample dataset; This indicates that the sample data is concentrated in M ​​environments, with S unsupervised waveform signal samples collected in each environment; ELBO NPR,m This represents the lower bound of evidence related to NPR in the m-th environment. ELBO VAE,m This represents the lower bound of evidence related to VAE in the m-th environment. Indicates the third error; This represents the aggregated result of the distributions of the N second latent variables in the m-th environment; z represents the variational inference of the posterior distribution of latent variables in the m-th environment using unsupervised sample data;m This represents the hidden variable in the m-th environment; This represents the N location labels collected in the m-th environment of the sample dataset; This represents the waveform signal sample corresponding to the N location labels collected in the m-th environment of the sample dataset; Indicates the fourth error; This represents the estimation of latent variables in mapping; Let f(x) represent the likelihood of N waveform signal sample data with location labels in the m-th environment, assuming the presence of latent variables. This represents the S unlabeled waveform signal samples in the m-th environment of the sample dataset; Indicates the second error; f represents the aggregated result of all distributions of the first latent variable in the m-th environment; θ (z m ) represents the prior of the hidden variable; Indicates the first error; This represents the aggregated result of all reconstructed waveforms in the m-th environment.

[0153] Based on the above embodiments, the construction of the sample dataset in this device specifically includes:

[0154] Collect massive amounts of waveform signal samples;

[0155] S1: Extract a waveform signal sample, locate the preset coordinates of the waveform signal sample in the current coordinate system, and obtain the position label of the waveform signal sample;

[0156] S2: The location tag and the waveform signal sample form a waveform sample data pair;

[0157] Repeat steps S1-S2 to obtain a large number of waveform sample data pairs, and use the large number of waveform sample data pairs to construct a sample dataset.

[0158] This invention provides a location parameter estimation device that acquires a semi-supervised dataset and a received waveform at a target time. The semi-supervised dataset and the received waveform are then input into a pre-constructed semi-supervised location parameter estimation model to obtain a distribution inference result. The semi-supervised location parameter estimation model is trained using variational inference theory on a sample dataset based on a deep neural network, and includes at least an encoder and a regressor. The semi-supervised location parameter estimation model provided by this invention has the ability to actively learn during inference and can effectively utilize semi-supervised datasets for parameter inference. This significantly reduces the communication, computational, and dataset acquisition costs associated with deploying machine learning algorithms in positioning systems, achieving higher accuracy and generalization in location parameter estimation.

[0159] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a position parameter estimation method, which includes: acquiring a semi-supervised dataset and the received waveform at the target time; inputting the semi-supervised dataset and the received waveform into a pre-built semi-supervised position parameter estimation model to obtain a distribution inference result; wherein the semi-supervised position parameter estimation model is trained based on a deep neural network using variational inference theory on a sample dataset, and the semi-supervised position parameter estimation model includes at least an encoder and a regressor.

[0160] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the 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 to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0161] On the other hand, the present invention also provides a computer program product, which 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 position parameter estimation method provided by the above methods. The method includes: acquiring a semi-supervised dataset and a received waveform at a target time; inputting the semi-supervised dataset and the received waveform into a pre-constructed semi-supervised position parameter estimation model to obtain a distribution inference result; wherein the semi-supervised position parameter estimation model is trained based on a deep neural network using variational inference theory on a sample dataset, and the semi-supervised position parameter estimation model includes at least an encoder and a regressor.

[0162] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the position parameter estimation method provided by the methods described above. The method includes: acquiring a semi-supervised dataset and a received waveform at a target time; inputting the semi-supervised dataset and the received waveform into a pre-constructed semi-supervised position parameter estimation model to obtain a distribution inference result; wherein the semi-supervised position parameter estimation model is trained based on a deep neural network using variational inference theory on a sample dataset, and the semi-supervised position parameter estimation model includes at least an encoder and a regressor.

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

[0164] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This 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 to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments 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 not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for estimating location parameters, characterized in that, include: Obtain the received waveforms at the semi-supervised dataset and the target time. The semi-supervised dataset and the received waveform are input into a pre-built semi-supervised position parameter estimation model to obtain the distribution inference results; The semi-supervised position parameter estimation model is based on a deep neural network trained using variational inference theory on a sample dataset. The semi-supervised position parameter estimation model includes at least an encoder and a regressor. The semi-supervised location parameter estimation model is obtained by training a sample dataset using a deep neural network and variational inference theory. Specifically, it includes: Construct a sample dataset; Based on the aforementioned sample dataset, waveform sample data pairs are obtained; wherein, the sample dataset includes at least a large number of waveform sample data pairs, and each waveform sample data pair includes waveform signal samples and location labels; The waveform sample data pairs are inferred by jointly using an unsupervised encoder and a supervised encoder to obtain a first latent variable distribution and a second latent variable distribution; wherein, the first latent variable distribution is obtained by dimensionality reduction of the input waveform using an unsupervised encoder, and the second latent variable distribution is obtained by dimensionality reduction of the waveform sample data pairs using a supervised encoder; the input waveform is the waveform signal sample included in the waveform sample data pairs; The first error, the second error, and the third error are calculated based on the first latent variable distribution and the second latent variable distribution. The first and second latent variable distributions are aggregated using an aggregator to obtain the mapping latent variable estimate of the current environment; The mapped latent variable estimate and the input waveform are input to a regressor to obtain the position parameter estimation result; the position parameter estimation result and the position label of the input waveform are compared to obtain the fourth error; wherein, the input waveform is a waveform signal sample included in the waveform sample data pair; Calculate the lower bound of evidence for the location parameter estimation result; wherein the lower bound of evidence is calculated based on a first preset formula using a first error, a second error, a third error, and a fourth error; Gradient descent is performed based on the lower bound of the evidence to obtain a semi-supervised location parameter estimation model.

2. The location parameter estimation method according to claim 1, characterized in that, The semi-supervised dataset and the received waveform are input into a pre-built semi-supervised position parameter estimation model to obtain distribution inference results, specifically including: The semi-supervised dataset is dimensionality reduced and aggregated using an encoder to obtain inferred latent variables; wherein the encoder includes a supervised encoder and an unsupervised encoder; The distribution inference results are obtained using a regressor based on the inferred latent variables.

3. The location parameter estimation method according to claim 1, characterized in that, The first error, the second error, and the third error are calculated based on the first latent variable distribution and the second latent variable distribution, specifically including: The waveform distribution is reconstructed based on the first latent variable distribution to obtain the reconstructed waveform; A first error is calculated based on the reconstructed waveform and the input waveform, and a second error is calculated based on the distribution of the first latent variable. The KL divergence between the second latent variable distribution and the first latent variable distribution is measured and summed to obtain the third error.

4. The location parameter estimation method according to claim 3, characterized in that, The first preset formula includes: in, This represents the inherent likelihood relationship of the sample dataset that the semi-supervised location parameter estimation model can describe; Indicates the location label in the sample dataset; This represents the waveform signal sample corresponding to the position label in the sample dataset; This indicates that the sample data is concentrated in M ​​environments, with S unsupervised waveform signal samples collected in each environment; ELBO NPR,m This represents the lower bound of evidence related to NPR in the m-th environment. ELBO VAE,m This represents the lower bound of evidence related to VAE in the m-th environment. Indicates the third error; This represents the aggregated result of the distributions of the N second latent variables in the m-th environment; z represents the variational inference of the posterior distribution of latent variables in the m-th environment using unsupervised sample data; m This represents the hidden variable in the m-th environment; This represents the N location labels collected in the m-th environment of the sample dataset; This represents the waveform signal sample corresponding to the N location labels collected in the m-th environment of the sample dataset; Indicates the fourth error; This represents the estimation of latent variables in mapping; Let f(x) represent the likelihood of N waveform signal sample data with location labels in the m-th environment, assuming the presence of latent variables. This represents the S unlabeled waveform signal samples in the m-th environment of the sample dataset; Indicates the second error; f represents the aggregated result of all distributions of the first latent variable in the m-th environment; θ (z m ) represents the prior of the hidden variable; Indicates the first error; This represents the aggregated result of all reconstructed waveforms in the m-th environment.

5. The location parameter estimation method according to claim 1, characterized in that, Constructing the sample dataset specifically includes: Collect massive amounts of waveform signal samples; S1: Extract a waveform signal sample, locate the preset coordinates of the waveform signal sample in the current coordinate system, and obtain the position label of the waveform signal sample; S2: The location tag and the waveform signal sample form a waveform sample data pair; Repeat steps S1-S2 to obtain a large number of waveform sample data pairs, and use the large number of waveform sample data pairs to construct a sample dataset.

6. A position parameter estimation device, characterized in that, include: The acquisition unit is used to acquire the received waveforms at the semi-supervised dataset and the target time. The inference unit is used to input the semi-supervised dataset and the received waveform into a pre-built semi-supervised position parameter estimation model to obtain the distribution inference result; The semi-supervised position parameter estimation model is based on a deep neural network trained using variational inference theory on a sample dataset. The semi-supervised position parameter estimation model includes at least an encoder and a regressor. The semi-supervised location parameter estimation model is obtained by training a sample dataset using a deep neural network and variational inference theory. Specifically, it includes: Construct a sample dataset; Based on the aforementioned sample dataset, waveform sample data pairs are obtained; wherein, the sample dataset includes at least a large number of waveform sample data pairs, and each waveform sample data pair includes waveform signal samples and location labels; The waveform sample data pairs are inferred by jointly using an unsupervised encoder and a supervised encoder to obtain a first latent variable distribution and a second latent variable distribution; wherein, the first latent variable distribution is obtained by dimensionality reduction of the input waveform using an unsupervised encoder, and the second latent variable distribution is obtained by dimensionality reduction of the waveform sample data pairs using a supervised encoder; the input waveform is the waveform signal sample included in the waveform sample data pairs; The first error, the second error, and the third error are calculated based on the first latent variable distribution and the second latent variable distribution. The first and second latent variable distributions are aggregated using an aggregator to obtain the mapping latent variable estimate of the current environment; The mapped latent variable estimate and the input waveform are input to a regressor to obtain the position parameter estimation result; the position parameter estimation result and the position label of the input waveform are compared to obtain the fourth error; wherein, the input waveform is a waveform signal sample included in the waveform sample data pair; Calculate the lower bound of evidence for the location parameter estimation result; wherein the lower bound of evidence is calculated based on a first preset formula using a first error, a second error, a third error, and a fourth error; Gradient descent is performed based on the lower bound of the evidence to obtain a semi-supervised location parameter estimation model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the position parameter estimation method as described in 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 position parameter estimation method as described in 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 position parameter estimation method as described in any one of claims 1 to 5.

Images