Position estimation method and device based on multi-task joint learning

Abstract

The invention relates to a position estimation method and device based on multi-task joint learning. The method comprises the following steps: acquiring a channel impulse response sample and an estimated value of a receiving end to a distance; constructing a variational auto-encoder model; the variational auto-encoder model comprises an encoder, a decoder, a channel state clustering device, a mutual information estimator, a channel state discriminator and a geometric information estimator; pre-training and joint training are sequentially performed on the variational auto-encoder model to obtain a trained variational auto-encoder model, in the pre-training stage, an encoder, a decoder and a channel state clustering device are trained, and in the joint training stage, full model parameters are optimized based on preset joint loss; and inputting a channel impulse response to be positioned into the trained variational auto-encoder model, and determining a target position according to the output geometric information estimator. By adopting the method, the accuracy, the stability and the robustness of positioning estimation can be improved, and the high-precision indoor positioning requirement can be met.

Description

Technical Field

[0001] This application relates to the field of wireless positioning technology, and in particular to a location estimation method and apparatus based on multi-task joint learning. Background Technology

[0002] In wireless positioning systems, it is typically necessary to estimate the geometric information related to distance and angle at the receiver and transmitter from the channel impulse response or channel state information, and then use these estimates to calculate the position. However, in complex indoor environments, due to the prevalence of multipath effects and non-line-of-sight propagation, signals are blocked, reflected, or scattered by objects such as walls and furniture, and reach the receiver through complex propagation paths. The received signal is superimposed with a large amount of multipath signal, causing the geometric information estimated by the receiver to deviate or contain errors compared to the true value, ultimately leading to a decrease in the accuracy of position estimation based on this geometric information.

[0003] Currently, by optimizing physical layer signal design, employing multi-antenna arrays, designing advanced signal processing and estimation algorithms, and utilizing fingerprint database matching, multipath interference can be suppressed to some extent, improving positioning performance in complex scenarios. However, these methods have limitations to varying degrees, including: insufficient utilization of channel state information, failure to consider estimation bias, high costs of offline database construction and fingerprint updates, poor model generalization ability in unknown environments, and insufficient adaptability to dynamic scene changes. These limitations make it difficult for existing technologies to achieve a good balance between complexity, cost, accuracy, and robustness, and cannot meet the growing demand for diversified, high-precision, and reliable indoor positioning services. Summary of the Invention

[0004] Therefore, it is necessary to provide a location estimation method and apparatus based on multi-task joint learning to address the aforementioned technical problems.

[0005] A location estimation method based on multi-task joint learning, the method comprising: Obtain channel impulse response samples and receiver's distance estimate; A variational autoencoder model is constructed, comprising an encoder, a decoder, a channel state clusterer, a mutual information estimator, a channel state discriminator, and a geometric information estimator. The encoder extracts the original features from the channel impulse response samples, which are decoupled into channel state-invariant features and channel state-specific features. The decoder reconstructs the channel impulse response based on the original features and the re-fused channel state-invariant and channel state-specific features, respectively, to obtain the reconstructed channel impulse response. The mutual information estimator estimates the mutual information between the channel state-invariant and channel state-specific features. The channel state clusterer clusters the channel state based on the channel state-specific features to obtain channel state categories. The channel state discriminator uses the channel state-specific features and the channel state-invariant features processed by the gradient inversion layer as inputs to discriminate the channel state categories, respectively calculating a first discrimination loss and a second discrimination loss to enhance the separability of the channel state by the specific features and the inseparability of the channel state by the invariant features. The geometric information estimator outputs a geometric information estimate based on the channel state-invariant features and the receiver's distance estimate. The variational autoencoder model is subjected to pre-training and joint training in sequence to obtain a trained variational autoencoder model. In the pre-training stage, the encoder, decoder and channel state clusterer are trained. In the joint training stage, the parameters of the whole model are optimized based on the pre-set joint loss. The channel impulse response to be located is input into the trained variational autoencoder model, and the target position is determined based on the output geometric information estimate.

[0006] A location estimation device based on multi-task joint learning, the device comprising: The sample construction module is used to obtain channel impulse response samples and the receiver's estimate of the distance; A variational autoencoder model is constructed, comprising an encoder, a decoder, a channel state clusterer, a mutual information estimator, a channel state discriminator, and a geometric information estimator. The encoder extracts the original features from the channel impulse response samples, which are decoupled into channel state-invariant features and channel state-specific features. The decoder reconstructs the channel impulse response based on the original features and the re-fused channel state-invariant and channel state-specific features, respectively, to obtain the reconstructed channel impulse response. The mutual information estimator estimates the mutual information between the channel state-invariant and channel state-specific features. The channel state clusterer clusters the original features to obtain channel state categories. The channel state discriminator uses the channel state-specific features and the channel state-invariant features processed by the gradient inversion layer as inputs to discriminate the channel state categories, respectively calculating a first discrimination loss and a second discrimination loss to enhance the separability of the channel state by the specific features and the inseparability of the channel state by the invariant features. The geometric information estimator outputs a geometric information estimate based on the channel state-invariant features and the receiver's distance estimate. The model training module is used to perform pre-training and joint training on the variational autoencoder model in sequence to obtain a trained variational autoencoder model. In the pre-training stage, the encoder, decoder and channel state clusterer are trained. In the joint training stage, the parameters of the whole model are optimized based on the pre-set joint loss. The output module is used to input the channel impulse response to be located into the trained variational autoencoder model and determine the target position based on the output geometric information estimate.

[0007] The aforementioned location estimation method and apparatus based on multi-task joint learning, by constructing a variational autoencoder model, extracts the original features from the channel impulse response and decouples them into channel state-invariant features and specific features, achieving accurate separation of geometric information and scene interference information. The mutual information estimator ensures that the two types of features are complementary and independent, so that invariant features are only associated with geometric information and specific features only reflect environmental and channel characteristics, effectively filtering out interference from multipath and non-line-of-sight propagation. By fusing signal reconstruction, channel state clustering identification, and geometric information estimation through pre-training and joint training, it fully utilizes channel state information, avoiding the limitations of single tasks. It eliminates the need for complex offline library construction, reducing system complexity and deployment costs, and enhancing the model's generalization ability and adaptability in unknown environments and dynamic scenarios. Ultimately, it improves the accuracy, stability, and robustness of location estimation, meeting the requirements for high-precision indoor positioning. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a location estimation method based on multi-task joint learning in one embodiment; Figure 2This is a flowchart illustrating a location estimation method based on multi-task joint learning in a specific embodiment. Figure 3 This is a schematic diagram of a variational autoencoder model structure in one embodiment; Figure 4 The waveform diagram shows the channel impulse response in one embodiment. Figure 5 This is a schematic diagram of channel state clustering in one embodiment; Figure 6 This is a schematic diagram of the original signal and the reconstructed signal in one embodiment; Figure 7 This is a comparison chart of the original ranging error and the model-estimated ranging error in one embodiment; Figure 8 This is a graph showing the cumulative distribution function of the positioning error in one embodiment. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0010] In one embodiment, such as Figure 1 As shown, a location estimation method based on multi-task joint learning is provided, including the following steps: Step 102: Obtain the channel impulse response sample and the receiver's estimated distance.

[0011] Channel impulse response samples refer to the response signals captured by the receiver that reflect the change in signal strength over time as the signal travels along the propagation path (including direct, reflected, and scattered signals) to the receiver. These are the fundamental data for extracting channel characteristics. The distance estimate at the receiver is typically the ranging result (including bias) obtained by the receiving device based on its embedded ranging algorithm (such as leading-edge detection).

[0012] Step 104: Construct the variational autoencoder model.

[0013] The variational autoencoder model comprises an encoder, a decoder, a channel state clusterer, a mutual information estimator, a channel state discriminator, and a geometric information estimator. The encoder extracts the original features from the channel impulse response samples, which are decoupled into channel state-invariant features and channel state-specific features. The decoder reconstructs the channel impulse response based on the original features and by re-fusing the channel state-invariant and channel state-specific features. The mutual information estimator estimates the mutual information between the channel state-invariant and channel state-specific features. The channel state clusterer clusters the original features to obtain channel state categories. The channel state discriminator uses both the channel state-specific features and the channel state-invariant features processed by the gradient inversion layer as input to discriminate channel state categories, calculating a first discrimination loss and a second discrimination loss to reinforce the separability of channel states by specific features and the inseparability of channel states by invariant features, respectively. The geometric information estimator outputs a geometric information estimate based on the channel state-invariant features and the receiver's distance estimate.

[0014] The encoder is the model's feature extraction module, mapping the raw channel impulse response to raw features containing key channel information, providing the foundation for subsequent feature decoupling and information extraction. Raw features are feature vectors output by the encoder, containing all key channel information (geometric correlation and scene interference), and are the core input for subsequent feature decoupling and processing by each module. Channel state-invariant features are features decoupled from the raw features that are only related to the geometric position of the receiver-transmitter (e.g., distance, angle) and are unaffected by scene interference such as wall obstruction and multipath scattering; they are the core basis for accurate positioning. Channel state-specific features are features decoupled from the raw features that only reflect the current environment and channel characteristics (e.g., multipath intensity, obstruction type) and do not carry geometric position information; they are used for channel propagation state.

[0015] The decoder is a module that reconstructs the channel impulse response by inversely reconstructing the original features and by re-fusing channel state-invariant and channel state-specific features. By comparing the reconstructed result with the original signal, it verifies whether the effective channel information is completely preserved during feature extraction and decoupling. The reconstructed channel impulse response is the reconstructed signal output by the decoder, used to evaluate the completeness of the features extracted by the encoder and indirectly verify the reliability of subsequent localization of relevant features.

[0016] The mutual information estimator is a module used to calculate the mutual information (measure the correlation between features) between channel state invariant features and specific features. By constraining the mutual information to approach zero, it ensures that there is no redundant information overlap between the two types of features and avoids the mixing of geometric information by scene interference.

[0017] The channel state clusterer is a module that obtains channel state categories based on original features through unsupervised clustering (such as classifying line-of-sight, non-line-of-sight, and multipath scattering types). It is used to mine the environmental and characteristic patterns of the channel. The channel state category is the channel scene classification result (such as line-of-sight, non-line-of-sight, strong multipath, etc.) output by the channel state clusterer, which is used to identify the interference type of the current signal and assist in localization correction.

[0018] The channel state discriminator is a module that uses channel state categories as labels to identify the scene type to which input features belong, enhancing the ability of specific features to identify channel states and the inability of invariant features to identify channel states. The gradient inversion layer performs an identity transformation during forward propagation and inverts the gradient during backward propagation. Channel state-invariant features are input into the channel state discriminator via the gradient inversion layer, forcing the discriminator to fail to identify channel states through invariant features, thereby strengthening the scene independence of invariant features.

[0019] The geometric information estimator (GEE) is a module that outputs geometric information estimates (such as the distance and angle between the receiver and transmitter) based on the channel state-invariant characteristic, directly providing core geometric parameters for positioning calculation. The geometric information estimates, such as distance and angle, output by the GEE are the direct basis for calculating the target position using algorithms such as polygonal positioning and triangulation.

[0020] Understandably, the model achieves precise feature separation, solving the core problem of geometric information being confused with scene interference in existing technologies. Invariant features filter out multipath interference, while specific features focus on environmental and channel characteristics. The collaboration of various modules ensures feature purity, providing a reliable feature foundation for subsequent accurate localization.

[0021] Step 106: Perform pre-training and joint training on the variational autoencoder model in sequence to obtain the trained variational autoencoder model. In the pre-training stage, the encoder, decoder and channel state clusterer are trained. In the joint training stage, the parameters of the whole model are optimized based on the pre-set joint loss.

[0022] Joint loss is the overall objective function that integrates reconstruction loss, clustering loss, mutual information loss, and other multi-task losses, and is used to simultaneously optimize the various modules of the model; pre-training is to first train the encoder, decoder, and clusterer to enable them to initially master the capabilities of feature extraction, signal reconstruction, and channel clustering.

[0023] Step 108: Input the channel impulse response to be located into the trained variational autoencoder model, and determine the target position based on the output geometric information estimate.

[0024] The channel impulse response to be located is the channel signal of the receiver at an unknown location, and the target location is the receiver coordinates obtained by solving geometric information.

[0025] This step inputs the signal to be located into the trained model and uses the output geometric information estimate to calculate the position. The geometric information estimate is the core basis for positioning. In addition, the reconstructed channel impulse response can be used to verify the effectiveness of signal features, and the channel state category can be used to adapt the scene correction strategy. This solves the problems of low positioning accuracy and large susceptibility to multipath interference in the existing technology. Feature separation ensures the reliability of geometric estimation, and combined with dynamic correction by scene category, it ultimately improves the accuracy and stability of positioning in complex environments.

[0026] The aforementioned location estimation method based on multi-task joint learning constructs a variational autoencoder model to extract the original features from the channel impulse response and decouple them into channel state-invariant features and specific features. This achieves accurate separation of geometric information and scene interference information. The mutual information estimator ensures that the two types of features are independent, so that invariant features are only associated with geometric information and specific features only reflect environmental and channel characteristics. This effectively filters out interference caused by multipath and non-line-of-sight propagation. By fusing signal reconstruction, channel state clustering identification, and geometric information estimation through pre-training and joint training, the method fully utilizes channel state information, avoids the limitations of a single task, eliminates the need for complex offline library construction, reduces system complexity and deployment costs, and enhances the model's generalization ability and adaptability in unknown environments and dynamic scenarios. Ultimately, this improves the accuracy, stability, and robustness of location estimation, meeting the requirements for high-precision indoor positioning.

[0027] In one embodiment, the pre-training phase for training the encoder, decoder, and channel state clusterer includes: inputting channel impulse response samples into the encoder, which maps the channel impulse response samples to low-dimensional latent variables to obtain original features; reconstructing the channel impulse response based on the original features using the decoder to obtain the reconstructed channel impulse response, and calculating the reconstruction loss between the channel impulse response samples and the reconstructed channel impulse response; calculating the KL divergence loss between the encoder output posterior distribution and a preset standard normal distribution to constrain the regularity of the latent variable distribution; obtaining the initial channel state categories and cluster centers, using the t-distribution as the kernel function, calculating the soft assignment probability of the channel impulse response samples belonging to different channel categories using the channel state clusterer, and then calculating the KL divergence between the soft assignment probability and the target distribution as the clustering loss; iteratively optimizing the parameters of the encoder, decoder, and channel state clusterer with the goal of minimizing the pre-training loss composed of the reconstruction loss, KL divergence loss, and clustering loss until the iteration stopping condition is met, thus obtaining the pre-trained variational autoencoder model.

[0028] Specifically, the model uses an encoder to process the original channel impulse response. We obtain latent variables (original features) mapped to a low-dimensional space. The encoder is made by Reconstructing the channel impulse response Loss due to Kullback-Leibler divergence (KL divergence) The constrained latent space distribution approximates a standard normal distribution, determined by the reconstruction loss. The difference between the original input and the reconstructed output is measured by the following expression: KL divergence loss: ,in The posterior distribution of the encoder output. For the preset standard normal distribution distributed.

[0029] Reconstruction loss: , For the sample size, .

[0030] Given the channel state category is Given initial cluster centers ( Using the t-distribution as the kernel to measure the sample Category soft assignment probability The expression is: .

[0031] Define matching soft allocation and target distribution The KL divergence loss between them is: ,in , .

[0032] Pre-training optimizes the parameters of the encoder, channel state clusterer, and decoder, with the optimization objective being to minimize the pre-training loss. ,in and This is a hyperparameter. In one embodiment, the joint training phase, based on a pre-set joint loss, optimizes the full model parameters, including: sharing the encoder from the channel state clustering phase and retaining the channel state clusterer and decoder; inputting channel impulse response samples into the pre-trained encoder to obtain the original features and decoupling them into channel state-invariant features and channel state-specific features; inputting the original features into the channel state clusterer and decoder to calculate the clustering loss and reconstruction loss; inputting the channel state-specific features into the channel state clusterer and channel state discriminator, and inputting the channel state-invariant features into the channel state discriminator via a gradient inversion layer to calculate the first discrimination loss and the second discrimination loss, respectively; and then inputting the channel state-invariant features into the channel state clusterer and channel state discriminator. After the channel state-specific features are re-fused, they are input into the encoder to reconstruct the channel impulse response and calculate the reconstruction loss. The channel state-invariant features and channel state-specific features are input into the mutual information estimator to calculate the mutual information loss. The geometric information estimator outputs the geometric information estimate based on the channel state-invariant features and the receiver's distance estimate, and calculates the geometric estimation loss. Based on the outputs of each module, the loss values ​​of each item in the joint loss are calculated. With the goal of minimizing the joint loss, the parameters of the encoder, decoder, mutual information estimator, channel state clusterer, channel state discriminator, and geometric information estimator are optimized in reverse until the iteration stopping condition is met, and the trained variational autoencoder model is obtained.

[0033] Specifically, the mutual information estimator uses Jensen-Shannon divergence (JS divergence) to estimate the channel state-invariant features. and channel state specific characteristics The mutual information, and the mutual information loss are: ,in and Channel state invariant features and channel state specific characteristics marginal distribution, Channel state invariant feature and channel state specific characteristics The joint distribution The output of the mutual information estimator, Depend on The result is obtained by random rearrangement. In addition to JS divergence, other methods such as KL (Kullback-Leibler) divergence can also be used to calculate mutual information loss.

[0034] Cross-entropy loss is used to measure the discrimination capability of the model channel state discriminator, taking into account channel state invariant features. The identification loss is: Channel state invariant characteristics The identification loss is: ,in, The channel states are obtained by unsupervised clustering of the channel state clusterer. and These are the prediction results from the channel state discriminator.

[0035] Extracted channel state invariant features and channel state specific characteristics The input decoder is re-fused to reconstruct the signal, and the reconstruction loss is: .

[0036] Joint training optimizes the parameters of the encoder, decoder, mutual information estimator, channel state clusterer, channel state discriminator, and geometric information estimator, with the optimization objective being to minimize the total loss function. ,in , , , , , , This is a hyperparameter.

[0037] In one embodiment, the joint loss includes the KL divergence loss of the encoder, the reconstruction loss of the decoder, the clustering loss of the channel state clusterer, the mutual information loss of the mutual information estimator, the first discrimination loss and the second discrimination loss of the channel state discriminator, and the geometric estimation loss of the geometric information estimator.

[0038] In one embodiment, the step of calculating the mutual information loss of the mutual information estimator includes: obtaining channel state invariant features and channel state specific features; randomly rearranging the channel state specific features to obtain reference features; inputting the channel state invariant features, channel state specific features, and reference features into the mutual information estimator; having the mutual information estimator calculate the joint distribution of the channel state invariant features and the channel state specific features, their respective marginal distributions, and the marginal distribution of the channel state invariant features and the reference features; and calculating the mutual information based on the calculation results to obtain the mutual information loss.

[0039] In one embodiment, the step of calculating the first discrimination loss includes: obtaining the true channel state category label corresponding to the channel state specific features; the true channel state category label is obtained based on the output of a pre-trained channel state clusterer; inputting the channel state specific features into a channel state discriminator and outputting the predicted probability of each channel state category; and using a cross-entropy loss function to calculate the cross-entropy between the predicted probability and the true channel state category label to obtain the first discrimination loss.

[0040] In one embodiment, the step of calculating the second discrimination loss includes: obtaining the true channel state category label corresponding to the channel state invariant feature; the true channel state category label is obtained based on the output of a pre-trained channel state clusterer; the channel state invariant feature is input into the channel state discriminator through a gradient inversion layer, and the predicted probability of each channel state category is output; the cross-entropy loss function is used to calculate the cross-entropy between the predicted probability and the true channel state category label to obtain the second discrimination loss.

[0041] In one embodiment, the step of calculating the geometric estimation loss of the geometric information estimator includes: calculating true geometric information based on the true positions of the receiver and transmitter; obtaining the geometric information estimate output by the geometric information estimator; if the geometric information estimate is a deterministic value, obtaining the geometric estimation loss based on the error between the true geometric information and the geometric information estimate; if the geometric information estimate is a probability distribution, obtaining the geometric estimation loss based on the KL divergence of the true geometric information relative to the geometric information estimate. In this embodiment, the true positions of the receiver and transmitter are the true coordinate values ​​measured by a high-precision device.

[0042] In one embodiment, the geometric information estimator outputs a geometric information estimate based on channel state-invariant features and the receiver's distance estimate, including: if the geometric information estimate is a deterministic value, then the channel state-invariant features and the receiver's distance estimate are used as inputs, and a pre-set neural network is used to estimate the distance value; if the geometric information estimate is a probability distribution, then the channel state-invariant features and the receiver's distance estimate are used as inputs, the distance to be estimated is modeled as a Gaussian distribution, and a natural parameter network is used to estimate the probability density function of the distance distribution.

[0043] In this embodiment, assuming the distance distribution to be estimated is Gaussian, the invariant features and the receiver's estimated distance are used as inputs. Natural-Parameter Networks are used as the geometric information estimator to estimate the probability density function of the distance distribution (the model and probability distribution selection are not limited to this). The loss is: ,in, The dimension of the output of the probabilistic neural network. and These are the probabilistic neural networks of the th The mean and variance of the output. It is a constant. Alternatively, invariant features and the receiver's distance estimate can be used as input, and a neural network (model selection is not limited to these) such as a multilayer perceptron or convolutional neural network can be used as a geometric information estimator to estimate the distance value. The loss is: ,in, and These are the actual and estimated values ​​of the distance between the receiver and the transmitter, respectively.

[0044] In one specific embodiment, such as Figure 2 As shown, a location estimation method based on multi-task joint learning is provided, including the following steps: Step 200: Collect the channel impulse response, the receiver's distance estimate, and the actual locations of the receiver and transmitter as samples for modeling to construct a dataset.

[0045] Step 202, as follows Figure 2 As shown, a variational autoencoder model for multi-task joint learning is constructed. The model includes an encoder (feature extractor), a decoder (signal reconstructor), a mutual information estimator, a channel state clusterer, a channel state discriminator, and a geometric information estimator.

[0046] Step 204: Use the training dataset to pre-train the encoder, channel state clusterer, and decoder so that the variational autoencoder model of multi-task joint learning can learn the feature extraction, signal reconstruction, and channel state clustering strategies of the channel impulse response, minimize the difference between the original signal and the reconstructed signal, maximize the sample similarity within the same channel state, and maximize the difference between different channel states. Step 206: The encoder, decoder, mutual information estimator, channel state clusterer, channel state discriminator, and geometric information estimator are jointly trained using the training dataset. Gradient inversion is applied to the channel state discriminator with channel state invariant features, enabling the variational autoencoder model of multi-task joint learning to learn the channel state invariant features, channel state specific features, signal reconstruction, and channel state discrimination strategies of channel impulse response or channel state information. This allows the encoder to minimize the difference between the original signal and the reconstructed signal, minimizes the residual between the geometric information estimate obtained by the geometric information estimator and the true value, and enables the channel state discriminator to distinguish different channel states through channel state specific features, but not through channel state invariant features. The wireless channel state is represented by the extracted channel invariant features and channel state specific features. Step 208: Input the geometric information of the sample to be estimated into the trained multi-task joint learning variational autoencoder model, and output the reconstructed channel impulse response, channel state and distance estimation distribution. Step 210: Use geometric information to estimate the distribution and estimate the location coordinates.

[0047] like Figure 3As shown, a schematic diagram of a variational autoencoder model structure is provided. The dashed line (step 104: model construction and pre-training) completes the initial training of the encoder, decoder, and channel state clusterer. At this point, the encoder can extract the original features, the decoder can reconstruct the signal, and the clusterer can unsupervisedly partition the channel scene, laying the foundation for subsequent joint training. The dashed and solid lines (step 106: joint training) introduce a mutual information estimator, a channel state discriminator, and a geometric information estimator based on the pre-training. All modules are simultaneously optimized through multi-task joint loss: the mutual information estimator constrains the complementarity and independence of the two types of features; the channel state discriminator enhances the purity of feature separation; the geometric information estimator learns accurate geometric parameters; and the decoder ensures feature integrity. By decoupling interference features from geometric features and utilizing supervision and constraints between multiple modules, the problem of decreased positioning accuracy due to multipath interference in complex environments is fundamentally solved, ultimately achieving high-precision and robust indoor positioning.

[0048] like Figure 4 The waveform diagram of the channel impulse response is shown below. Figure 5 The diagram shows a clustering representation of channel states. After pre-training, the channel impulse responses can be clearly divided into four categories, demonstrating good separability and clustering effectiveness for each channel state. Figure 6 The diagram shows the original and reconstructed signals. The encoder reconstructs the signal using channel state-invariant and channel state-specific features; for example... Figure 7 The image shows a comparison between the original ranging error and the model-estimated ranging error. The model can correct distance measurements. Figure 8 The cumulative distribution function of the positioning error shown in the figure illustrates that through design such as feature decoupling and multi-task collaborative training, the impact of multipath / non-line-of-sight interference on positioning is significantly suppressed, achieving high positioning accuracy. In most scenarios, the error can be controlled within a small range, meeting the requirements of high-precision indoor positioning.

[0049] This invention utilizes channel state-invariant and specific features extracted from channel impulse response or channel state information to achieve signal reconstruction, channel state identification, and geometric information estimation. The geometric information is then used to infer the target location. This method significantly reduces system complexity and deployment costs, effectively integrates multi-task information, enhances the model's adaptability and robustness in different scenarios, and further improves the accuracy and stability of positioning estimation.

[0050] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0051] In one embodiment, a location estimation apparatus based on multi-task joint learning is provided, comprising: The sample construction module is used to obtain channel impulse response samples and the receiver's estimate of the distance; The model building module is used to construct a variational autoencoder model. The variational autoencoder model includes an encoder, a decoder, a channel state clusterer, a mutual information estimator, a channel state discriminator, and a geometric information estimator. The encoder extracts the original features from the channel impulse response samples, which are decoupled into channel state-invariant features and channel state-specific features. The decoder reconstructs the channel impulse response based on the original features and by re-fusing the channel state-invariant and channel state-specific features, respectively, to obtain the reconstructed channel impulse response. The mutual information estimator estimates the mutual information between the channel state-invariant and channel state-specific features. The channel state clusterer clusters the original features to obtain channel state categories. The channel state discriminator uses the channel state-specific features and the channel state-invariant features processed by the gradient inversion layer as inputs to discriminate the channel state categories, respectively calculating the first and second discrimination losses to enhance the separability of the channel state by the specific features and the inseparability of the channel state by the invariant features. The geometric information estimator outputs a geometric information estimate based on the channel state-invariant features and the receiver's distance estimate. The model training module is used to perform pre-training and joint training on the variational autoencoder model in sequence to obtain a trained variational autoencoder model. In the pre-training stage, the encoder, decoder and channel state clusterer are trained. In the joint training stage, the parameters of the whole model are optimized based on the pre-set joint loss. The output module is used to input the channel impulse response to be located into the trained variational autoencoder model and determine the target position based on the output geometric information estimate.

[0052] Specific limitations regarding the location estimation device based on multi-task joint learning can be found in the limitations of the location estimation method based on multi-task joint learning described above, and will not be repeated here. Each module in the aforementioned location estimation device based on multi-task joint learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for position estimation based on multi-task joint learning, characterized in that, The method comprises: obtaining channel impulse response samples and an estimated value of distance at a receiving end; constructing a variational autoencoder model; the variational autoencoder model comprises an encoder, a decoder, a channel state clusterer, a mutual information estimator, a channel state discriminator and a geometric information estimator; the encoder extracts original features of the channel impulse response samples, the original features are decoupled into channel state invariant features and channel state specific features; the decoder reconstructs the channel impulse response based on the original features and the channel state invariant features and the channel state specific features after re-fusion respectively; the mutual information estimator estimates mutual information of the channel state invariant features and the channel state specific features; the channel state clusterer clusters to obtain channel state categories based on the original features; the channel state discriminator discriminates the channel state categories with the channel state specific features and the channel state invariant features processed by a gradient inversion layer as inputs, and calculates a first discrimination loss and a second discrimination loss respectively to respectively strengthen the separability of the specific features to the channel state and the inseparability of the invariant features to the channel state; and the geometric information estimator outputs a geometric information estimator based on the channel state invariant features and the estimated value of distance at the receiving end; sequentially performing pre-training and joint training on the variational autoencoder model to obtain a trained variational autoencoder model, wherein the pre-training stage trains the encoder, the decoder and the channel state clusterer, and the joint training stage optimizes the whole model parameters based on a pre-set joint loss; inputting channel impulse response to be positioned into the trained variational autoencoder model, and determining a target position according to an output geometric information estimator.

2. The method of claim 1, wherein, The pre-training stage training the encoder, the decoder and the channel state clusterer comprises: inputting the channel impulse response samples into the encoder to map the channel impulse response samples into low-dimensional latent variables to obtain original features by the encoder; reconstructing the channel impulse response based on the original features by the decoder, and calculating a reconstruction loss of the channel impulse response samples and the reconstructed channel impulse response; calculating a KL divergence loss of a posterior distribution output by the encoder and a pre-set standard normal distribution to constrain the regularity of the latent variable distribution; obtaining initial channel state categories and cluster centers, adopting a t-distribution as a kernel function, calculating soft assignment probabilities of the channel impulse response samples belonging to different channel categories by the channel state clusterer, and then calculating a KL divergence of the soft assignment probabilities and a target distribution as a clustering loss; iteratively optimizing parameters of the encoder, the decoder and the channel state clusterer to minimize a pre-training loss composed of the reconstruction loss, the KL divergence loss and the clustering loss, until a condition for stopping iteration is met, and a pre-trained variational autoencoder model is obtained.

3. The method of claim 1, wherein, The joint training stage optimizing the whole model parameters based on the pre-set joint loss comprises: sharing the encoder in the channel state clustering stage and retaining the channel state clusterer and the decoder; inputting the channel impulse response samples into the pre-trained encoder to obtain the original features and decouple them into the channel state invariant features and the channel state specific features; inputting the original features into a channel state clusterer and a decoder, calculating a clustering loss and a reconstruction loss; inputting channel state specific features into a channel state clusterer and a channel state discriminator, and inputting channel state invariant features into the channel state discriminator through a gradient reversal layer, to respectively calculate a first discrimination loss and a second discrimination loss; inputting the channel state invariant features and the channel state specific features into an encoder after re-fusion, to reconstruct channel impulse responses and calculate a reconstruction loss; inputting the channel state invariant features and the channel state specific features into a mutual information estimator, to calculate a mutual information loss; inputting the channel state invariant features and the channel state specific features into a mutual information estimator, to calculate a mutual information loss; outputting a geometric information estimator based on the channel state invariant features and an estimated distance between the receiver and the transmitter, to calculate a geometric estimation loss; 4. The method of claim 1, wherein, calculating loss values in the joint loss according to outputs of the modules, and optimizing parameters of the encoder, the decoder, the mutual information estimator, the channel state clusterer, the channel state discriminator and the geometric information estimator in a reverse direction until a condition for stopping iteration is met, to obtain a trained variational autoencoder model.

5. The method of claim 4, wherein, The joint loss includes a KL divergence loss of the encoder, a reconstruction loss of the decoder, a clustering loss of the channel state clusterer, a mutual information loss of the mutual information estimator, a first discrimination loss and a second discrimination loss of the channel state discriminator, and a geometric estimation loss of the geometric information estimator. The step of calculating the mutual information loss of the mutual information estimator includes: obtaining the channel state invariant features and the channel state specific features, and randomly re-arranging the channel state specific features to obtain reference features; 6. The method of claim 1, wherein, inputting the channel state invariant features, the channel state specific features and the reference features into the mutual information estimator, to calculate a joint distribution of the channel state invariant features and the channel state specific features, marginal distributions of the channel state invariant features and the channel state specific features, and a marginal distribution of the channel state invariant features and the reference features, and to calculate mutual information based on the calculation results to obtain the mutual information loss. The step of calculating the first discrimination loss includes: obtaining real channel state category labels corresponding to the channel state specific features; the real channel state category labels are obtained according to outputs of the pre-trained channel state clusterer; inputting the channel state specific features into the channel state discriminator to output prediction probabilities of channel state categories; 7. The method of claim 1, wherein, using a cross-entropy loss function to calculate cross-entropy of the prediction probabilities and the real channel state category labels, to obtain the first discrimination loss. The step of calculating the second discrimination loss includes: obtaining real channel state category labels corresponding to the channel state invariant features; the real channel state category labels are obtained according to outputs of the pre-trained channel state clusterer; inputting the channel state invariant features into the channel state discriminator through the gradient reversal layer to output prediction probabilities of channel state categories; 8. The method of claim 4, wherein, using a cross-entropy loss function to calculate cross-entropy of the prediction probabilities and the real channel state category labels, to obtain the second discrimination loss. The step of calculating the geometric estimation loss of the geometric information estimator includes: calculating real geometric information according to real positions of the receiver and the transmitter; The geometry information estimator outputs a geometry information estimate based on the channel state invariant feature and the estimated distance of the receiver.

9. The method of claim 1, wherein, The geometry information estimator outputs a geometry information estimate based on the channel state invariant feature and the estimated distance of the receiver. If the geometry information estimate is a deterministic value, the channel state invariant feature and the estimated distance of the receiver are taken as inputs to estimate the distance value using a pre-set neural network. If the geometry information estimate is a probability distribution, the channel state invariant feature and the estimated distance of the receiver are taken as inputs, the distance to be estimated is modeled as a Gaussian distribution, and a natural parameter network is used to estimate the probability density function of the distance distribution. 10.A position estimation apparatus based on multi-task joint learning, characterized in that, The device comprises: A sample construction module for obtaining channel impulse response samples and an estimated distance of the receiver. A model construction module for constructing a variational autoencoder model; the variational autoencoder model comprises an encoder, a decoder, a channel state clusterer, a mutual information estimator, a channel state discriminator, and a geometry information estimator; the encoder extracts original features of the channel impulse response samples, which are decoupled into channel state invariant features and channel state specific features; the decoder reconstructs the channel impulse response based on the original features and the re-fused channel state invariant features and channel state specific features, respectively, to obtain reconstructed channel impulse responses; the mutual information estimator estimates the mutual information of the channel state invariant features and the channel state specific features; the channel state clusterer clusters based on the original features to obtain channel state categories, and the channel state discriminator discriminates the channel state categories based on the channel state specific features and the channel state invariant features processed by a gradient reversal layer as inputs, respectively, to calculate a first discrimination loss and a second discrimination loss to respectively strengthen the separability of the specific features for the channel state and the inseparability of the invariant features for the channel state; and the geometry information estimator outputs a geometry information estimate based on the channel state invariant feature and the estimated distance of the receiver. A model training module for sequentially performing pre-training and joint training on the variational autoencoder model to obtain a trained variational autoencoder model, wherein the encoder, the decoder, and the channel state clusterer are trained in the pre-training stage, and the whole model is parameter-optimized based on a pre-set joint loss in the joint training stage. A result output module for inputting channel impulse responses to be positioned into the trained variational autoencoder model and determining a target position based on the output geometry information estimate.

Images