An indoor positioning method and device based on multi-scale wireless channel feature fusion

Abstract

The application provides an indoor positioning method and device based on multi-scale wireless channel feature fusion, and relates to the technical field of indoor positioning. The method comprises the following steps: collecting an RSSI vector and a CSI matrix of a terminal to be positioned through a Wi-Fi access point, rearranging to obtain a complex CSI matrix, performing wavelet decomposition, standardization and wavelet multi-scale feature extraction to obtain an enhanced complex CSI matrix amplitude, constructing an access point level local feature and a global wireless fingerprint feature, calculating an access point confidence weight and obtaining a main positioning feature through a local-global gating fusion mechanism, fusing a Token level relationship feature into the main positioning feature and inputting the main positioning feature into a deep network regression head, and combining a deep feature tree regressor and an original fingerprint tree regressor to obtain a two-dimensional positioning coordinate. The application can improve the feature expression ability, fusion adaptive ability and coordinate output stability in a complex indoor multipath environment, and realizes high-precision and high-robustness indoor wireless positioning.

Description

Technical Field

[0001] This invention relates to the field of indoor positioning technology, and in particular to an indoor positioning method and apparatus based on multi-scale wireless channel feature fusion. Background Technology

[0002] With the development of applications such as smart manufacturing, warehousing and logistics, mobile robots, indoor navigation, and location services, achieving low-cost, high-precision, and continuously operating target positioning in indoor environments with weak or no satellite signals has become a significant technical challenge in the fields of wireless sensing and the Industrial Internet. Existing indoor positioning solutions typically employ technologies such as ultra-wideband, Bluetooth, vision, LiDAR, or Wi-Fi wireless signals. Among these, Wi-Fi infrastructure is widely deployed and can generate an indoor wireless fingerprint and estimate target location by collecting Received Signal Strength Indicator (RSSI) and Channel State Information (CSI) data at access points without deploying expensive dedicated hardware.

[0003] Existing Wi-Fi-based positioning methods can be broadly categorized into geometrically based ranging / angle measurement methods, fingerprint-based positioning methods using traditional machine learning, and end-to-end positioning methods based on deep learning. These different approaches have developed distinct implementation schemes around key aspects such as wireless propagation parameter extraction, fingerprint feature construction, location matching, coordinate regression, and wireless signal representation learning.

[0004] In geometrically based ranging / angle-measuring positioning methods, existing solutions typically utilize Wi-Fi signal propagation models to convert information such as received signal strength, time of arrival, angle of arrival, phase difference, or frequency domain response from CSI into spatial distance or angle parameters. These parameters are then combined with the positional relationships of multiple access points to estimate the location of the terminal to be located using triangulation, ranging, angle-measuring, or graph search methods. For example, ZL202210797947.4 discloses an indoor positioning scheme combining CSI and sensor information. It obtains the multipath channel angle of arrival (AOA) based on CSI and fuses sensor information, historical location, real access points (APs), and virtual APs to estimate the location of the device to be located using a fusion optimization algorithm.

[0005] In fingerprint localization methods based on traditional machine learning, existing schemes typically collect RSSI, CSI amplitude, CSI phase, or other wireless signal features at different reference points in the offline phase to establish a location fingerprint database. In the online phase, they collect the wireless signal of the terminal to be located and perform location matching or coordinate estimation using methods such as K-nearest neighbors, support vector machines, random forests, clustering algorithms, or multilayer perceptrons. For example, application number CN202511662694.X discloses a localization scheme based on a location fingerprint database and an adaptive weighted K-nearest neighbor algorithm. It establishes an offline database by collecting location fingerprints of reference points and adaptively determines the K value and modifies the distance metric based on fingerprint similarity differences in the online phase, thereby estimating the location of the point to be located.

[0006] In terms of end-to-end positioning methods based on deep learning, existing solutions mainly revolve around wireless fingerprint feature construction, deep feature extraction, attention modeling, sequence modeling, and coordinate regression. For example, application number ZL201810300922.2 discloses a Wi-Fi indoor positioning scheme combining wavelet transform and convolutional neural networks. It constructs a location feature map from Wi-Fi signals using wavelet transform and uses an improved convolutional neural network for classification and positioning, demonstrating the technical approach of combining wavelet features with deep models. Patent CN202310165399.8 discloses a neural network positioning scheme combining data augmentation, data filtering, noise reduction autoencoder, floor classification model, and coordinate regression model to learn relative coordinates from wireless fingerprints such as RSSI, demonstrating the technical approach of combining wireless fingerprint preprocessing, deep feature compression, and coordinate regression. Patent ZL202310206508.6 discloses a scheme for RSSI data augmentation using a generator and discriminator. It expands the training samples by generating new RSSI fingerprint data, demonstrating the technical approach of enhancing the distribution of wireless fingerprint data through sample generation. Patent ZL202410324559.3 discloses a localization scheme based on an attention mechanism to reconstruct RSSI fingerprint features. It increases the weight of important RSSI features through an attention matrix, demonstrating an adaptive weighting approach for different wireless fingerprint features. Patent CN202510881057.5 discloses an RSSI sequence localization scheme combining an autoencoder and an attention mechanism. This scheme extracts RSSI sequence features and performs localization estimation, demonstrating a combination of feature compression, sequence feature extraction, and attention weighting. Patent CN202511583367.5 discloses a localization scheme based on a mobile terminal RSSI dataset and a deep learning neural network. It predicts the target location through a trained network, demonstrating a technical approach that utilizes an end-to-end deep network to establish a mapping relationship between RSSI fingerprints and location coordinates.

[0007] Besides RSSI strength fingerprints, CSI channel state information can provide more granular amplitude, phase, and frequency domain channel response information, and is therefore also used in deep learning end-to-end positioning schemes. For example, patent ZL202011228750.6 discloses a scheme for indoor positioning using CSI measurements. It constructs location results by combining CSI amplitude and phase information, PCA dimensionality reduction, wavelet transform, and CSI amplitude and phase images, and integrates them with a convolutional neural network classification model, demonstrating a technical approach that combines CSI amplitude and phase features, wavelet processing, and deep classification models. Patent CN202511840292.4 discloses a CSI positioning scheme for multi-source wireless signals. It extracts CSI amplitude and phase as fingerprint information and combines anomaly detection, filtering and smoothing, sequence coding, Transformer models, and uncertainty estimation to achieve location prediction, demonstrating a technical approach that combines CSI fingerprint preprocessing, sequence relationship modeling, and prediction reliability estimation.

[0008] While existing Wi-Fi-based indoor positioning methods have developed into various implementation schemes, including ranging / angle measurement based on geometric relationships, fingerprint positioning based on traditional machine learning, and end-to-end positioning based on deep learning, certain shortcomings still exist in complex indoor environments. Indoor environments are characterized by walls, equipment, personnel movement, and metal objects, which cause significant multipath propagation, obstruction, and non-line-of-sight interference, resulting in fluctuations in the amplitude, phase, and frequency response of wireless signals with time, location, and environmental conditions. Geometric ranging or angle measurement methods typically rely on relatively stable propagation paths, time of arrival, or angle of arrival estimates. When indoor multipath propagation is severe or access point deployment conditions are limited, distance and angle parameters are prone to deviation, thus affecting the final positioning accuracy. Therefore, an indoor positioning method that does not solely rely on ranging or angle measurement parameters and can fully utilize the fingerprint characteristics of the wireless channel for location estimation is needed.

[0009] Traditional machine learning fingerprint localization methods typically rely on wireless fingerprint databases established offline and use models such as K-nearest neighbors, support vector machines, random forests, or multilayer perceptrons for online matching or regression. While these methods are relatively simple to implement, they often depend heavily on low-dimensional signal strength features like RSSI, failing to fully utilize the amplitude, phase, and frequency domain details contained in CSI. Furthermore, the signal reliability of different access points varies with location, occlusion, and environment. Using only fixed feature concatenation or uniform weighting makes it difficult to adequately distinguish the effective contributions of different access points to the current localization sample. Therefore, a localization method is needed that can jointly utilize RSSI, CSI, and multi-scale frequency domain features, and adaptively evaluate the reliability and contribution weights of different access point signals based on their states, thereby effectively fusing local access point features with global wireless fingerprint features.

[0010] Existing deep learning-based end-to-end localization methods can automatically learn nonlinear location representations from wireless fingerprints, improving the expressive power of localization models to some extent. However, some methods mainly focus on single steps such as RSSI fingerprint enhancement, CSI amplitude and phase image construction, convolutional feature extraction, attention weighting, or sequence modeling, lacking unified modeling of RSSI, complex CSI, and wavelet multi-scale features. For CSI signals, using only amplitude images, phase images, or simple feature stitching may not be sufficient to maintain the complex structural relationship between the real and imaginary parts of the CSI, nor can it simultaneously express its multi-scale frequency domain changes. Furthermore, the correlations between different access points and between access points and global wireless features are not adequately modeled. Therefore, a localization method is needed that can maintain the complex structure of CSI, fuse wavelet multi-scale frequency domain information, and further model the correlations between different access points and between access points and global wireless features.

[0011] Furthermore, some existing localization models use single-path neural networks to output location results, making their predictions susceptible to variations in training sample distribution, feature scale, and environmental changes. In complex multipath scenarios, across users, or with local signal anomalies, localization errors may fluctuate, particularly affecting median and tail errors. Therefore, a localization output mechanism is needed that can comprehensively utilize the expressive power of deep models, the original features of wireless fingerprints, and the stability of traditional regression models to improve the stability and reliability of the final two-dimensional coordinate estimation. Summary of the Invention

[0012] To address the technical problems in existing technologies, such as insufficient utilization of fine-grained features of wireless channels, inadequate modeling of reliability differences among multiple access points, insufficient fusion of local and global features, and limited stability of single-path positioning output, this invention provides an indoor positioning method and apparatus based on multi-scale wireless channel feature fusion. The technical solution is as follows: On the one hand, an indoor positioning method based on multi-scale wireless channel feature fusion is provided. This method is implemented by an indoor positioning device based on multi-scale wireless channel feature fusion, and includes: S1. Collect the received signal strength indication vector and channel state information matrix of the terminal to be located through multiple Wi-Fi access points.

[0013] S2. The channel state information matrix is ​​rearranged to obtain a complex channel state information matrix. Discrete wavelet decomposition is performed on the complex channel state subcarrier sequence of each access point to obtain wavelet multi-scale features. The received signal strength indication vector, the complex channel state information matrix, and the wavelet multi-scale features are standardized. The amplitude of the standardized channel state information matrix is ​​enhanced according to the standardized wavelet multi-scale features to obtain the amplitude of the enhanced channel state information matrix. The scaling factor is calculated based on the amplitude of the enhanced channel state information matrix to scale the real and imaginary parts of the standardized channel state information matrix to obtain the enhanced complex channel state information matrix for each access point.

[0014] S3. Based on the standardized received signal strength indication vector and the enhanced complex channel state information matrix, local features and global wireless fingerprint features at the access point level are obtained. Based on the local features at the access point level, the access point confidence weight is calculated and the backbone positioning features are obtained through the local-global gating fusion mechanism and gating multi-scale feature extraction. Token-level relationship features are obtained through the Token relationship modeling module. The Token-level relationship features are fused into the backbone positioning features to obtain the enhanced deep positioning features.

[0015] S4. Construct a joint loss function to train the deep network. Input the enhanced deep positioning features, the standardized received signal strength indication vector, and the enhanced complex channel state information matrix into the trained deep network regression head, deep feature tree regressor, and original fingerprint tree regressor to obtain the deep network prediction results, the deep feature tree regressor prediction results, and the original fingerprint tree regressor prediction results. Weight the results to obtain the final two-dimensional positioning coordinates.

[0016] Optionally, S2 includes: S21. The channel state information matrix is ​​rearranged according to the access point and the subcarrier dimension of the access point to obtain a complex channel state information matrix; wherein, the complex channel state information matrix includes a real part and an imaginary part.

[0017] S22. Perform discrete wavelet decomposition on the complex channel state information subcarrier sequence of each access point to obtain low-frequency approximate components and multi-level high-frequency detail components, and concatenate the amplitudes of the low-frequency approximate components and multi-level high-frequency detail components to obtain wavelet multi-scale features.

[0018] S23. Standardize the received signal strength indication vector, the real part of the complex channel state information matrix, the imaginary part of the complex channel state information matrix, and the wavelet multi-scale features to obtain the standardized received signal strength indication vector, the standardized channel state information matrix, and the standardized wavelet multi-scale features.

[0019] S24. Linearly project the standardized wavelet multi-scale features onto the subcarrier dimension feature space that is the same as the amplitude features of the standardized channel state information matrix. Then, enhance the amplitude of the standardized channel state information matrix according to the projected wavelet multi-scale features to obtain the enhanced channel state information matrix amplitude. Calculate the scaling factor based on the enhanced channel state information matrix amplitude, and scale the real and imaginary parts of the standardized channel state information matrix according to the scaling factor to obtain the enhanced complex channel state information matrix for each access point.

[0020] Optionally, the complex channel state information matrix is ​​as shown in equation (1): (1) In the formula, Indicates the first The access point Complex channel state information matrix on each subcarrier The magnitude of the complex channel state information matrix is ​​represented. Represents the imaginary unit. The phase of the complex channel state information matrix is ​​represented. Represents the real part of the complex channel state information matrix. This represents the imaginary part of the complex channel state information matrix.

[0021] Optionally, S3 includes: S31. Construct a complex local encoder for each access point, calculate the average received signal strength indication vector for each access point, and input the average received signal strength indication vector and the enhanced complex channel state information subcarrier features into the complex linear layer and complex normalization layer of the complex local encoder of the corresponding access point to obtain the access point-level local features.

[0022] S32. Construct a global encoder by concatenating the standardized received signal strength indication vector with the amplitude vector of the enhanced complex channel state information matrix and inputting the result into the global encoder to obtain the global wireless fingerprint feature.

[0023] S33. Calculate the access point confidence weight based on the access point-level local features of each access point, and weight the access point-level local features according to the access point confidence weight to obtain the access point-level local weighted features. Construct local context features using the mean and standard deviation of the access point-level local features.

[0024] S34. By controlling the fusion ratio of local weighted features at the access point level and global wireless fingerprint features through a local-global gating fusion mechanism, fusion features are obtained. The fusion features are then extracted with positioning-related features at different scales through multi-scale dilated convolution branches. The positioning-related features at different scales are fused to obtain gated multi-scale fusion features. The backbone positioning features are obtained based on the fusion features and the gated multi-scale fusion features.

[0025] S35. Construct a Token sequence based on access point-level local features, global wireless fingerprint features, and gated multi-scale fusion features. Input the Token sequence into the Token self-attention module and the lightweight Mamba state space module to obtain relation features and sequence relation features.

[0026] S36. The relation features and sequence relation features are fused into the backbone localization features through learnable gain to obtain enhanced deep localization features.

[0027] Optionally, S33 calculates the access point confidence weight based on the access point-level local features of each access point, including: S331. For the local features at the access point level of each access point, the access point confidence level is calculated using the following formula (2): (2) In the formula, Indicates the first Confidence level of each access point This represents the access point confidence estimation network. Indicates the first Local features at the access point level for each access point The mean statistic represents a local characteristic.

[0028] S332. Based on the set temperature and access point confidence level, the normalized confidence weight of the access point is obtained as shown in the following formula (3): (3) In the formula, Indicates the first Normalized confidence weights for each access point This represents the softmax temperature parameter. Indicates the number of access points.

[0029] Optionally, S34 includes: S341. Calculate the gating vector by concatenating the local weighted features at the access point level and the global wireless fingerprint features.

[0030] S342. Based on the gating vector, the access point-level local weighted features and the global wireless fingerprint features are fused to obtain the fused features.

[0031] S343. Multi-scale extraction is performed on the fused features using multiple convolutional branches with different dilation rates to obtain the multi-scale features extracted by each convolutional branch.

[0032] S344. The multi-scale features of each convolutional branch are fused using learnable scale weights to obtain gated multi-scale fused features.

[0033] S345. Based on the fusion features and gated multi-scale fusion features, obtain the backbone localization features.

[0034] Optionally, S4 includes: S41. Construct a joint loss function to train a deep network. Based on the trained deep network regression head, perform initial coordinate estimation of the enhanced depth localization features to obtain the deep network prediction results.

[0035] S42. Train a deep feature tree regressor using the enhanced deep localization features and the prediction results of the deep network as input, and obtain the prediction results of the deep feature tree regressor based on the trained deep feature tree regressor.

[0036] S43. The original wireless fingerprint features are used to train the original fingerprint tree regressor, which is composed of the normalized received signal strength indication vector, the amplitude of the enhanced complex channel state information matrix, the phase of the enhanced complex channel state information matrix and wavelet statistics. The prediction result of the original fingerprint tree regressor is obtained based on the trained original fingerprint tree regressor.

[0037] S44. The prediction results of the deep network, the deep feature tree regressor, and the original fingerprint tree regressor are weighted according to the fusion weights automatically selected by the internal validation set of the training fold to obtain the final two-dimensional positioning coordinates.

[0038] On the other hand, an indoor positioning device based on multi-scale wireless channel feature fusion is provided. This device is applied to an indoor positioning method based on multi-scale wireless channel feature fusion, and the device includes: The multi-access point wireless signal acquisition module is used to acquire the received signal strength indication vector and channel state information matrix of the terminal to be located through multiple Wi-Fi access points.

[0039] The signal preprocessing and feature construction module is used to rearrange the channel state information matrix to obtain a complex channel state information matrix, perform discrete wavelet decomposition on the complex channel state subcarrier sequence of each access point to obtain wavelet multi-scale features, standardize the received signal strength indication vector, the complex channel state information matrix, and the wavelet multi-scale features, and enhance the amplitude of the standardized channel state information matrix according to the standardized wavelet multi-scale features to obtain the enhanced channel state information matrix amplitude. Based on the enhanced channel state information matrix amplitude, a scaling factor is calculated to scale the real and imaginary parts of the standardized channel state information matrix to obtain the enhanced complex channel state information matrix for each access point.

[0040] The adaptive positioning model module is used to obtain access point-level local features and global wireless fingerprint features based on the standardized received signal strength indication vector and the enhanced complex channel state information matrix. It calculates the access point confidence weight based on the access point-level local features and obtains the backbone positioning features through a local-global gating fusion mechanism and gating multi-scale feature extraction. It obtains token-level relationship features through the token relationship modeling module and integrates the token-level relationship features into the backbone positioning features to obtain enhanced deep positioning features.

[0041] The hybrid regression output module is used to construct a joint loss function to train the deep network. It inputs the enhanced deep positioning features, the standardized received signal strength indication vector, and the enhanced complex channel state information matrix into the trained deep network regression head, the deep feature tree regressor, and the original fingerprint tree regressor to obtain the deep network prediction results, the deep feature tree regressor prediction results, and the original fingerprint tree regressor prediction results. The weighted results are then used to obtain the final two-dimensional positioning coordinates.

[0042] On the other hand, an indoor positioning device based on multi-scale wireless channel feature fusion is provided. The indoor positioning device based on multi-scale wireless channel feature fusion includes: a processor; a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, any one of the above-described indoor positioning methods based on multi-scale wireless channel feature fusion is implemented.

[0043] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described indoor positioning methods based on multi-scale wireless channel feature fusion.

[0044] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention proposes an indoor positioning method and apparatus based on multi-scale wireless channel feature fusion, targeting indoor wireless positioning scenarios with weak or unavailable satellite positioning signals. It provides a technical solution for indoor location awareness in applications such as intelligent manufacturing, warehousing and logistics, mobile robots, indoor navigation, and location services. The method utilizes existing Wi-Fi access points to collect RSSI and CSI signals from the terminal to be positioned. Without deploying expensive dedicated positioning hardware, it achieves two-dimensional position coordinate estimation of the terminal to be positioned through multi-access point wireless signal preprocessing, complex CSI modeling, wavelet multi-scale frequency domain enhancement, AP adaptive fusion, local-global gated multi-scale fusion, token relationship modeling, and hybrid regression output. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of an indoor positioning method based on multi-scale wireless channel feature fusion provided by an embodiment of the present invention; Figure 2 This is a flowchart illustrating the overall structure and implementation process of the indoor positioning method based on multi-scale wireless channel feature fusion provided in this embodiment of the invention. Figure 3 This is a complex CSI and wavelet multi-scale feature construction diagram provided in an embodiment of the present invention; Figure 4 This is an AP confidence and local context fusion graph provided in an embodiment of the present invention; Figure 5 This is a diagram of the gated multi-scale feature fusion module provided in an embodiment of the present invention; Figure 6 This is a hybrid graph of AP Token self-attention and lightweight Mamba provided in an embodiment of the present invention; Figure 7 This is the three-way regression and adaptive fusion weight graph provided in the embodiments of the present invention; Figure 8 This is a block diagram of an indoor positioning device based on multi-scale wireless channel feature fusion provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an indoor positioning device based on multi-scale wireless channel feature fusion provided in an embodiment of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0048] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0049] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0050] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0051] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0052] Existing deep learning Wi-Fi positioning technologies have covered areas such as RSSI fingerprint modeling, CSI amplitude and phase feature extraction, wavelet frequency domain processing, attention weighting, sequence modeling, and coordinate regression. These technologies provide a technical foundation for this invention to further construct joint inputs of RSSI, complex CSI, and wavelet multi-scale features, and to introduce AP adaptive fusion, local-global gated fusion, token relationship modeling, and hybrid regression output.

[0053] This invention provides an indoor positioning method based on multi-scale wireless channel feature fusion. This method can be implemented by an indoor positioning device based on multi-scale wireless channel feature fusion, which can be a terminal or a server. Figure 1 The flowchart shown is for an indoor positioning method based on multi-scale wireless channel feature fusion. The processing flow of this method may include the following steps: S1. Collect the received signal strength indication vector and channel state information matrix of the terminal to be located through multiple Wi-Fi access points.

[0054] One feasible implementation method is, for example Figure 2As shown, the present invention can be divided into a multi-access point wireless signal acquisition layer, a signal preprocessing and feature construction layer, an adaptive positioning model layer, a hybrid regression output layer, and a positioning result evaluation layer. The layers are connected sequentially according to the data flow. Furthermore, after model training is completed, the hybrid regression output weights are determined using the internal validation set within the training layer, enabling the final positioning result to integrate the feature representation capabilities of the deep network and the stability of the wireless fingerprint regression.

[0055] Specifically, the multi-access point wireless signal acquisition layer is used to acquire the wireless signals received by the terminal to be located at its current indoor location. For each positioning sample, the data is acquired from... RSSI vector of each wireless access point and CSI matrix Taking four access points and 90 CSI subcarriers per access point as an example, CSI can be represented as follows: ,in , RSSI reflects changes in received signal strength, while CSI contains more granular information on channel amplitude, phase, and frequency response. Together, they constitute the wireless fingerprint input for subsequent positioning models.

[0056] S2. The channel state information matrix is ​​rearranged to obtain a complex channel state information matrix. Discrete wavelet decomposition is performed on the complex channel state subcarrier sequence of each access point to obtain wavelet multi-scale features. The received signal strength indication vector, the complex channel state information matrix, and the wavelet multi-scale features are standardized. The amplitude of the standardized channel state information matrix is ​​enhanced according to the standardized wavelet multi-scale features to obtain the amplitude of the enhanced channel state information matrix. The scaling factor is calculated based on the amplitude of the enhanced channel state information matrix to scale the real and imaginary parts of the standardized channel state information matrix to obtain the enhanced complex channel state information matrix for each access point.

[0057] In one feasible implementation, the signal preprocessing and feature construction layer is used to convert the raw wireless signal into multiple types of features suitable for model processing.

[0058] Specifically, the process of constructing complex CSI and wavelet multi-scale features is as follows: Figure 3 As shown, for each access point's acquired CSI, this invention first retains both the real and imaginary parts in a two-channel format to avoid the loss of phase structure and complex correlation caused by using only the amplitude. Then, discrete wavelet decomposition is performed on the CSI sequence to extract low-frequency approximation components and multi-level high-frequency detail components. The wavelet features are then projected onto the CSI subcarrier dimension to enhance the CSI amplitude. Through this processing method, the model can utilize the complex representation capability of CSI while also introducing multi-scale frequency domain perturbation information, thereby providing more comprehensive wireless channel characteristics for subsequent AP-level local coding.

[0059] Optionally, step S2 above may include the following steps S21-S24: S21. The channel state information matrix is ​​rearranged according to the access point and the subcarrier dimension of the access point to obtain a complex channel state information matrix; wherein, the complex channel state information matrix includes a real part and an imaginary part.

[0060] In one feasible implementation, the present invention first rearranges the CSI according to the access point and subcarrier dimensions, and represents it as a two-channel form with real and imaginary parts. To preserve the complex structure of CSI.

[0061] Specifically, no. The access point Complex CSI on each subcarrier The calculation is as follows:

[0062] in, , These represent the corresponding CSI amplitude and phase, respectively. Indicates the access point index. , Number of access points; Indicates subcarrier index, , This indicates the number of CSI subcarriers for each access point.

[0063] Furthermore, to preserve the complex structure of CSI, this invention uses the real and imaginary parts as two separate input channels and employs a complex linear transformation in the local encoder. For complex inputs... Output the real part of the complex feature. and the virtual part The calculation is as follows:

[0064]

[0065] in and These are the real and imaginary parts of the complex input feature, respectively; and These are the weight matrices corresponding to the real and imaginary parts of the complex linear layer, respectively; and The real and imaginary parts are biased respectively.

[0066] S22. Perform discrete wavelet decomposition on the complex channel state information subcarrier sequence of each access point to obtain low-frequency approximate components and multi-level high-frequency detail components, and concatenate the amplitudes of the low-frequency approximate components and multi-level high-frequency detail components to obtain wavelet multi-scale features.

[0067] In one feasible implementation, the RSSI, CSI real part, CSI imaginary part and wavelet features are standardized respectively to reduce the scale differences introduced by different acquisition batches and terminal devices.

[0068] Specifically, the CSI subcarrier sequence is obtained by arranging the complex channel responses of the access point on multiple OFDM subcarriers according to the subcarrier index.

[0069] S23. Standardize the received signal strength indication vector, the real part of the complex channel state information matrix, the imaginary part of the complex channel state information matrix, and the wavelet multi-scale features to obtain the standardized received signal strength indication vector, the standardized channel state information matrix, and the standardized wavelet multi-scale features.

[0070] In one feasible implementation, while performing standardization, discrete wavelet decomposition is performed on the CSI sequence of each access point to extract low-frequency approximate components and multi-level high-frequency detail components, and their amplitudes are spliced ​​together to form wavelet multi-scale features.

[0071] CSI sequence for each access point implement Discrete wavelet decomposition:

[0072] in Indicates the first Low-frequency approximation coefficients of the layer; Indicates the first High-frequency detail factor of the layer; Indicate execution Layered discrete wavelet decomposition; Indicates the wavelet decomposition level; This indicates the decomposition of the hierarchical index. .

[0073] The amplitude values ​​of each level of coefficients are extracted and concatenated to obtain the wavelet multi-scale eigenvector. :

[0074] in, This indicates a feature splicing operation.

[0075] S24. Linearly project the standardized wavelet multi-scale features onto the subcarrier dimension feature space that is the same as the amplitude features of the standardized channel state information matrix. Then, enhance the amplitude of the standardized channel state information matrix according to the projected wavelet multi-scale features to obtain the enhanced channel state information matrix amplitude. Calculate the scaling factor based on the enhanced channel state information matrix amplitude, and scale the real and imaginary parts of the standardized channel state information matrix according to the scaling factor to obtain the enhanced complex channel state information matrix for each access point.

[0076] In one feasible implementation, wavelet features are linearly projected and mapped to the CSI subcarrier dimension to enhance the CSI amplitude. The same scaling factor is applied synchronously to the real and imaginary parts of the CSI, thereby enhancing the multi-scale information in the frequency domain while maintaining the original phase structure.

[0077] Specifically, the CSI amplitude after wavelet feature enhancement for:

[0078] in, Indicates the first CSI amplitude vector of each access point; This represents a linear projection layer used to map wavelet features to the CSI subcarrier dimension.

[0079] Furthermore, the corresponding scaling factor is calculated based on the enhanced amplitude. :

[0080] in, It is a very small positive number, used to prevent division by zero.

[0081] Then, the real and imaginary parts of CSI are scaled synchronously to obtain the first... Complex CSI with enhanced access points :

[0082] in, Indicates the first The original complex CSI vector of each access point.

[0083] By applying the same scaling factor to both the real and imaginary parts, the original phase structure can be maintained while enhancing the CSI amplitude.

[0084] S3. Based on the standardized received signal strength indication vector and the enhanced complex channel state information matrix, local features and global wireless fingerprint features at the access point level are obtained. Based on the local features at the access point level, the access point confidence weight is calculated and the backbone positioning features are obtained through the local-global gating fusion mechanism and gating multi-scale feature extraction. Token-level relationship features are obtained through the Token relationship modeling module. The Token-level relationship features are fused into the backbone positioning features to obtain the enhanced deep positioning features.

[0085] In one feasible implementation, the adaptive positioning model layer is the core computational part of the invention, used to learn location-related representations from multi-access point RSSI, enhanced complex CSI, and wavelet multi-scale features.

[0086] Specifically, step S3 above may include the following steps S31-S36: S31. Construct a complex local encoder for each access point, calculate the average received signal strength indication vector for each access point, and input the average received signal strength indication vector and the enhanced complex channel state information subcarrier features into the complex linear layer and complex normalization layer of the complex local encoder of the corresponding access point to obtain the access point-level local features.

[0087] In one feasible implementation, a complex local encoder is first constructed for each access point. The average RSSI of the access point and the enhanced CSI complex sequence are then input into a complex linear layer and a complex normalization layer to obtain AP-level local features.

[0088] S32. Construct a global encoder by concatenating the standardized received signal strength indication vector with the amplitude vector of the enhanced complex channel state information matrix and inputting the result into the global encoder to obtain the global wireless fingerprint feature.

[0089] In one feasible implementation, the RSSI vector and CSI amplitude vector are concatenated and then input into a global encoder to obtain global wireless fingerprint features. :

[0090] in, Indicates a global encoder; Represents the RSSI vector; This indicates the amplitude characteristics of CSI.

[0091] S33. Calculate the access point confidence weight based on the access point-level local features of each access point, and weight the access point-level local features according to the access point confidence weight to obtain the access point-level local weighted features. Construct local context features using the mean and standard deviation of the access point-level local features.

[0092] In one feasible implementation, the model calculates the AP confidence weight based on the local features of each access point, and constructs local context features using the mean and standard deviation of the access point features to characterize the reliability differences and overall distribution of different access points in the current sample.

[0093] Specifically, the process of fusing AP confidence with local context is as follows: Figure 4 As shown, different access points are affected to varying degrees by occlusion, multipath reflection, and terminal attitude changes in indoor environments. Fixedly splicing or averaging multiple AP features can easily introduce noise from low-quality access points. Therefore, this invention calculates confidence weights for the local features of each access point and adaptively adjusts the contribution ratio of each access point based on the AP feature distribution of the current sample. Simultaneously, to describe the dispersion of the overall features of multiple access points, this invention further utilizes the mean and standard deviation of AP features to construct local contextual features, which assist the subsequent fusion module in suppressing the influence of abnormal APs.

[0094] Optionally, calculating the access point confidence weight based on the access point-level local features of each access point in step S33 may include the following steps S331-S332: S331. For the access point-level local features of each access point, the access point confidence level is calculated using the following formula:

[0095] in, This represents the AP confidence estimation network; Indicates the first AP-level local features obtained by complex local coding of each access point; This represents the mean statistic of local features; 0.35 is the local feature mean correction coefficient.

[0096] S332. Adaptively set temperature based on access point characteristic fluctuations. , obtained the Normalized weights of each access point The expression, and satisfying :

[0097] in, This represents the softmax temperature parameter; This represents the summation index of the access point.

[0098] Furthermore, local fusion features are obtained by weighting based on AP confidence scores. :

[0099] Simultaneously, local context features are constructed based on the mean and standard deviation of the local features of all access points. The calculation formula is as follows:

[0100] in, This represents a local context coding network; and These represent the mean and standard deviation of the local features of all access points along the access point dimension, respectively, and are used to describe the overall distribution of features of multiple access points in the current sample.

[0101] Local context can characterize the overall dispersion among multiple access points in the current sample, and is used to suppress the impact of accidental abnormal access points on the localization results.

[0102] S34. By controlling the fusion ratio of local weighted features at the access point level and global wireless fingerprint features through a local-global gating fusion mechanism, fusion features are obtained. The fusion features are then extracted with positioning-related features at different scales through multi-scale dilated convolution branches. The positioning-related features at different scales are fused to obtain gated multi-scale fusion features. The backbone positioning features are obtained based on the fusion features and the gated multi-scale fusion features.

[0103] In one feasible implementation, the local-global gating fusion and multi-scale feature extraction process is as follows: Figure 5 As shown, after AP confidence weighting, local features can reflect the fine-grained contributions of different access points, while global features obtained by concatenating RSSI and CSI amplitudes can describe the overall wireless fingerprint distribution. To utilize the complementary information of both, this invention adaptively controls the fusion ratio of local and global features through gating vectors; subsequently, multiple convolutional branches with different dilation rates are introduced to enhance the fused features at multiple scales, enabling the model to capture wireless channel variations at different ranges and scales.

[0104] Optionally, step S34 above may include the following steps S341-S345: S341. Calculate the gating vector by concatenating the local weighted features at the access point level and the global wireless fingerprint features.

[0105] In one feasible implementation, local weighted features are used. With global features Calculate the gating vector after concatenation :

[0106] in, This represents the sigmoid activation function; This indicates a gated network.

[0107] S342. Based on the gating vector, the access point-level local weighted features and the global wireless fingerprint features are fused to obtain the fused features.

[0108] In one feasible implementation, the feature obtained by fusing local features and global features :

[0109] in, Represents a fusion mapping network; This represents element-wise multiplication; For local context residual fusion coefficients.

[0110] S343. Multi-scale extraction is performed on the fused features using multiple convolutional branches with different dilation rates to obtain the multi-scale features extracted by each convolutional branch.

[0111] In one feasible implementation, the fused features are then extracted at multiple scales using convolutional branches with different dilation rates to obtain the first... Multi-scale features extracted by each dilated convolution branch The calculation formula is as follows:

[0112] in, Indicates the expansion rate Convolution operations; This represents the total number of branches across multiple scales.

[0113] S344. The multi-scale features of each convolutional branch are fused using learnable scale weights to obtain gated multi-scale fused features.

[0114] In one feasible implementation, each scale feature is weighted by a learnable scale weight. By fusing the features, multi-scale enhanced features are obtained. :

[0115] in, Indicates the projection mapping layer; Indicates the multi-scale residual fusion coefficient; Indicates the first Learnable fusion weights for each scale branch.

[0116] S345. Based on the fusion features and gated multi-scale fusion features, obtain the backbone localization features.

[0117] In one feasible implementation, the final backbone localization features obtained after gated multi-scale fusion The calculation is as follows:

[0118] S35. Construct a Token sequence based on access point-level local features, global wireless fingerprint features, and gated multi-scale fusion features. Input the Token sequence into the Token self-attention module and the lightweight Mamba state space module to obtain relation features and sequence relation features.

[0119] In one feasible implementation, the token relationship mixing and initial depth coordinate estimation process is as follows: Figure 6 As shown. To avoid prematurely compressing all access point information into a single vector, this invention constructs local features, global features, and fused features of each access point as token sequences, enabling the model to explicitly model the relationships between access points (APs) and between APs and the global wireless fingerprint at the token level. Furthermore, this invention combines a token self-attention module and a lightweight Mamba state space module: the former captures the correlation between different tokens, and the latter supplements the expression of sequence relationships. The fused enhanced deep features are then input into the regression head to obtain an initial estimate of the neural network coordinates.

[0120] Specifically, to avoid compressing all features into a single vector, this invention constructs an access point token, a global token, and a fused token to form a token sequence. :

[0121] in, express Local features of each access point, and used as input for the AP local token; It represents the global wireless fingerprint feature and serves as the input for the global token; This represents the backbone localization features after gating multi-scale fusion and serves as the fusion token input.

[0122] Relationship features are obtained from the token sequence through a self-attention hybrid module. :

[0123] in, This represents the Token self-attention relationship modeling module.

[0124] Meanwhile, the token sequence is input into a lightweight Mamba state space module for supplementary modeling, and the sequence relationship features are obtained through mean pooling. :

[0125] in, This indicates a mean pooling operation performed along the token dimension. This represents a lightweight state-space relation modeling module.

[0126] S36. The relation features and sequence relation features are fused into the backbone localization features through learnable gain to obtain enhanced deep localization features.

[0127] Token self-attention features and lightweight Mamba features are fused into the backbone features through learnable gain to obtain enhanced deep features. :

[0128] in, and represents the adaptive fusion coefficients of the Token self-attention feature and the lightweight Mamba sequence feature, respectively.

[0129] The adaptive localization model layer controls the fusion ratio of AP-level local weighted features and global wireless fingerprint features through a local-global gated fusion mechanism, and extracts localization-related features at different scales through multi-scale dilated convolution branches. To avoid prematurely compressing multi-access point information into a single vector, this invention also constructs AP local features, global wireless fingerprint features, and gated multi-scale fusion features into a token sequence, which is input into the token self-attention module and the lightweight Mamba state space module to model the relationships between different access points and between access points and global features. Through the above processing, the model can simultaneously utilize local AP features, global wireless fingerprint features, multi-scale frequency domain features, and token-level relationship features to form enhanced deep localization features.

[0130] S4. Construct a joint loss function to train the deep network. Input the enhanced deep positioning features, the standardized received signal strength indication vector, and the enhanced complex channel state information matrix into the trained deep network regression head, deep feature tree regressor, and original fingerprint tree regressor to obtain the deep network prediction results, the deep feature tree regressor prediction results, and the original fingerprint tree regressor prediction results. Weight the results to obtain the final two-dimensional positioning coordinates.

[0131] In one feasible implementation, the hybrid regression output layer is used to output the final two-dimensional position coordinates.

[0132] Specifically, the three-way candidate prediction and adaptive weight fusion process is as follows: Figure 7As shown, when a single deep network directly regresses coordinates, it is easily affected by cross-user distribution differences and local anomalies in fingerprints; while tree models have a strong nonlinear fitting ability for local fingerprint neighborhoods. Based on this, this invention constructs three candidate prediction results: initial deep network coordinate estimation, deep feature tree regression, and original fingerprint tree regression, and automatically selects fusion weights on the validation set within the training fold. This hybrid output method can combine the expressive power of deep features and the local stability of traditional fingerprint regression, thereby improving the robustness of the final localization result.

[0133] Optionally, step S4 above may include the following steps S41-S44: S41. Construct a joint loss function to train a deep network. Based on the trained deep network regression head, perform initial coordinate estimation of the enhanced depth localization features to obtain the deep network prediction results.

[0134] In one feasible implementation, the present invention first uses a deep network regression head to perform initial coordinate estimation on the enhanced depth localization features to obtain the deep network prediction results.

[0135] Specifically, during the training phase, this invention employs a joint loss function consisting of location smoothing loss, Euclidean distance loss, input reconstruction loss, and scale-gated regularization to constrain the output coordinates of the deep network, maintain the recoverability of wireless signal feature representation, and avoid excessive concentration of multi-scale gating weights. The total loss during the deep network training phase is calculated as follows:

[0136] in, This represents the location smoothing loss, used to constrain the difference between predicted and true coordinates; This represents the Euclidean distance loss, used to directly constrain positioning errors; This represents the input reconstruction loss, used to maintain the recoverability of the wireless signal feature representation; This represents the scale-gating regularization term, used to avoid excessive concentration of multi-scale gating weights; represents the weight coefficients of the Euclidean distance loss, reconstruction loss, and gating regularization term, respectively.

[0137] After completing the deep network training, this invention further constructs a three-way hybrid regression output: the first path is the initial coordinate estimation of the deep network. The second path enhances deep features. Preliminary estimation of network coordinates Input the deep feature tree regressor to train it, and obtain The third approach uses RSSI, CSI amplitude, CSI phase, and wavelet statistics to construct the original fingerprint features, and trains the original fingerprint tree regressor to obtain... .

[0138] The enhancement features are first input into the deep network regression head to obtain an initial estimate of the neural network coordinates. :

[0139] in, Represents a coordinate regression network; These represent the two-dimensional horizontal and vertical coordinates predicted by the neural network, respectively.

[0140] S42. Train a deep feature tree regressor using the enhanced deep localization features and the prediction results of the deep network as input, and obtain the prediction results of the deep feature tree regressor based on the trained deep feature tree regressor.

[0141] In one feasible implementation, a deep feature tree regressor is trained using enhanced deep features and preliminary estimates of deep network coordinates as input.

[0142] S43. The original wireless fingerprint features are used to train the original fingerprint tree regressor, which is composed of the normalized received signal strength indication vector, the amplitude of the enhanced complex channel state information matrix, the phase of the enhanced complex channel state information matrix and wavelet statistics. The prediction result of the original fingerprint tree regressor is obtained based on the trained original fingerprint tree regressor.

[0143] In one feasible implementation, the original fingerprint tree regressor is trained using RSSI, CSI amplitude, CSI phase, and wavelet statistics to form the original wireless fingerprint features.

[0144] Specifically, wavelet statistics are obtained by statistically calculating the wavelet multi-scale features along the subcarrier dimension, including the mean, standard deviation, and maximum value.

[0145] S44. The prediction results of the deep network, the deep feature tree regressor, and the original fingerprint tree regressor are weighted according to the fusion weights automatically selected by the internal validation set of the training fold to obtain the final two-dimensional positioning coordinates.

[0146] In one feasible implementation, the three candidate prediction results are weighted according to the fusion weights automatically selected by the validation set within the training fold to obtain the final two-dimensional localization coordinates. This hybrid output method can combine the feature representation ability of deep networks and the fitting stability of tree models to local fingerprint neighborhoods, thereby reducing the error fluctuations caused by single-path regression output.

[0147] Final output two-dimensional positioning coordinates Obtained by adaptive weighting of three-way prediction:

[0148] in, Indicates to enhance deep features Preliminary estimation of depth coordinates The input is the obtained deep feature tree regression prediction result; This indicates the initial estimation results of the depth network coordinates; This represents the original fingerprint tree regression prediction result composed of RSSI, CSI amplitude, CSI phase, and wavelet statistics; Let each represent the fusion weight of the three candidate predictions, satisfying the following conditions: And it is automatically selected through the internal validation set of the training fold.

[0149] Furthermore, the present invention also includes a positioning result evaluation layer, used to statistically analyze the positioning error based on the Euclidean distance between the final predicted coordinates and the actual coordinates.

[0150] Preferably, the present invention focuses on statistically analyzing Median error and Tail error to evaluate the model's median localization error and tail error performance, respectively.

[0151] In summary, the overall implementation process of this invention can be summarized as follows: acquiring RSSI and CSI signals from multiple access points; performing complex reconstruction, standardization, and wavelet multi-scale frequency domain feature extraction on the CSI; enhancing the CSI amplitude while maintaining the complex phase structure using wavelet features; extracting AP-level local features and global wireless fingerprint features respectively; obtaining backbone positioning features through AP confidence, local context, local-global gated fusion, and gated multi-scale feature extraction; further learning the association relationships between multiple access points using the Token relationship modeling module; and finally obtaining the final positioning result through the hybrid output of deep network coordinate initial estimation, deep feature tree regression, and original fingerprint tree regression.

[0152] Specifically, this invention proposes a wavelet-enhanced RSSI / CSI joint fingerprint construction method that preserves the complex structure of the CSI. This method represents the multi-access point CSI as a two-channel form with real and imaginary parts, performs AP-level local coding in conjunction with the corresponding RSSI, and simultaneously performs wavelet multi-scale decomposition on the CSI sequence of each access point, mapping the wavelet features to the CSI subcarrier dimension to enhance the CSI amplitude while maintaining the consistency of the phase structure jointly characterized by the real and imaginary parts. This results in a wireless fingerprint feature that fuses RSSI, complex CSI, and wavelet frequency domain information.

[0153] This invention proposes a multi-scale RSSI / CSI localization model that combines AP confidence, local-global gating, and token relationship modeling. The model calculates confidence weights based on AP local features and constructs a local context. It then fuses AP-level local features with RSSI / CSI global features through a gating mechanism, and further combines this with multi-scale feature extraction to obtain enhanced localization features. Simultaneously, it constructs a token sequence from AP local features, global features, and fused features to model the relationships between different access points and between access points and global features.

[0154] This invention proposes a three-path wireless fingerprint hybrid regression method based on training fold validation weight selection. The strategy first uses a deep network to obtain initial coordinate estimates, then constructs a deep feature tree regressor and an original wireless fingerprint tree regressor, and adaptively selects the fusion weights of the three candidate predictions on the validation set within the training fold, finally outputting two-dimensional positioning coordinates.

[0155] It should be clarified here that the above application scenarios are only used to explain the principles and steps of the indoor positioning method in this invention, to facilitate the reader's understanding, and do not represent a limitation on the scope of protection of this patent.

[0156] This invention proposes an indoor positioning method and apparatus based on wireless channel state information and gated multi-scale adaptive fusion. This method addresses multi-access point RSSI / CSI indoor positioning tasks by jointly modeling RSSI, complex CSI, and wavelet multi-scale frequency domain features. It solves problems in existing technologies such as insufficient utilization of fine-grained wireless channel features, inadequate modeling of multi-access point reliability differences, insufficient fusion of local and global features, and limited stability of single-path positioning output. This method combines complex CSI local coding, wavelet feature enhancement, AP confidence and local context modeling, local-global gated multi-scale fusion, token relationship modeling, and a three-path hybrid regression output strategy to form an end-to-end wireless fingerprint positioning process. This improves feature representation capability, fusion adaptability, and coordinate output stability in complex indoor multipath environments, ultimately achieving high-precision and robust indoor wireless positioning.

[0157] Figure 8 This is a block diagram illustrating an indoor positioning device based on multi-scale wireless channel feature fusion according to an exemplary embodiment. The device is used for an indoor positioning method based on multi-scale wireless channel feature fusion. (Refer to...) Figure 8 The device includes a multi-access point wireless signal acquisition module 310, a signal preprocessing and feature construction module 320, an adaptive positioning model module 330, and a hybrid regression output module 340. Among them: The multi-access point wireless signal acquisition module 310 is used to acquire the received signal strength indication vector and channel state information matrix of the terminal to be located through multiple Wi-Fi access points.

[0158] The signal preprocessing and feature construction module 320 is used to rearrange the channel state information matrix to obtain a complex channel state information matrix, perform discrete wavelet decomposition on the complex channel state subcarrier sequence of each access point to obtain wavelet multi-scale features, standardize the received signal strength indication vector, the complex channel state information matrix and the wavelet multi-scale features, and enhance the amplitude of the standardized channel state information matrix according to the standardized wavelet multi-scale features to obtain the enhanced channel state information matrix amplitude. Based on the enhanced channel state information matrix amplitude, a scaling factor is calculated to scale the real and imaginary parts of the standardized channel state information matrix to obtain the enhanced complex channel state information matrix for each access point.

[0159] The adaptive positioning model module 330 is used to obtain access point-level local features and global wireless fingerprint features based on the standardized received signal strength indication vector and the enhanced complex channel state information matrix. It calculates the access point confidence weight based on the access point-level local features and obtains the backbone positioning features through a local-global gating fusion mechanism and gating multi-scale feature extraction. It obtains token-level relationship features through the token relationship modeling module and integrates the token-level relationship features into the backbone positioning features to obtain enhanced deep positioning features.

[0160] The hybrid regression output module 340 is used to construct a joint loss function to train the deep network. It inputs the enhanced deep positioning features, the standardized received signal strength indication vector, and the enhanced complex channel state information matrix into the trained deep network regression head, the deep feature tree regressor, and the original fingerprint tree regressor to obtain the deep network prediction results, the deep feature tree regressor prediction results, and the original fingerprint tree regressor prediction results. The weighted results are then used to obtain the final two-dimensional positioning coordinates.

[0161] This invention proposes an indoor positioning method and apparatus based on wireless channel state information and gated multi-scale adaptive fusion. This method addresses multi-access point RSSI / CSI indoor positioning tasks by jointly modeling RSSI, complex CSI, and wavelet multi-scale frequency domain features. It solves problems in existing technologies such as insufficient utilization of fine-grained wireless channel features, inadequate modeling of multi-access point reliability differences, insufficient fusion of local and global features, and limited stability of single-path positioning output. This method combines complex CSI local coding, wavelet feature enhancement, AP confidence and local context modeling, local-global gated multi-scale fusion, token relationship modeling, and a three-path hybrid regression output strategy to form an end-to-end wireless fingerprint positioning process. This improves feature representation capability, fusion adaptability, and coordinate output stability in complex indoor multipath environments, ultimately achieving high-precision and robust indoor wireless positioning.

[0162] Figure 9This is a schematic diagram of the structure of an indoor positioning device based on multi-scale wireless channel feature fusion provided in an embodiment of the present invention, as shown below. Figure 9 As shown, an indoor positioning device based on multi-scale wireless channel feature fusion can include the above-mentioned... Figure 8 The indoor positioning device shown is based on multi-scale wireless channel feature fusion. Optionally, the indoor positioning device 410 based on multi-scale wireless channel feature fusion may include a first processor 2001.

[0163] Optionally, the indoor positioning device 410 based on multi-scale wireless channel feature fusion may also include a memory 2002 and a transceiver 2003.

[0164] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0165] The following is combined with Figure 9 The various components of the indoor positioning device 410 based on multi-scale wireless channel feature fusion are described in detail below: The first processor 2001 is the control center of the indoor positioning device 410 based on multi-scale wireless channel feature fusion. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0166] Optionally, the first processor 2001 can perform various functions of the indoor positioning device 410 based on multi-scale wireless channel feature fusion by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0167] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 9 CPU0 and CPU1 are shown in the diagram.

[0168] In a specific implementation, as one example, the indoor positioning device 410 based on multi-scale wireless channel feature fusion may also include multiple processors, for example... Figure 9The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0169] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0170] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently, and may be connected via the interface circuit of the indoor positioning device 410 based on multi-scale wireless channel feature fusion. Figure 9 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0171] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0172] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 9 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0173] Optionally, the transceiver 2003 can be integrated with the first processor 2001 or exist independently, and can be connected to the interface circuit of the indoor positioning device 410 based on multi-scale wireless channel feature fusion. Figure 9 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0174] It should be noted that, Figure 9 The structure of the indoor positioning device 410 based on multi-scale wireless channel feature fusion shown in the figure does not constitute a limitation on the router. The actual knowledge structure identification device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0175] Furthermore, the technical effects of the indoor positioning device 410 based on multi-scale wireless channel feature fusion can be referred to the technical effects of the indoor positioning method based on multi-scale wireless channel feature fusion described in the above method embodiments, and will not be repeated here.

[0176] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0177] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0178] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0179] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0180] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0181] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0182] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, 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 this 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.

[0183] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An indoor positioning method based on multi-scale wireless channel feature fusion, characterized in that, The method includes: S1. Collect the received signal strength indication vector and channel state information matrix of the terminal to be located through multiple Wi-Fi access points; S2. The channel state information matrix is ​​rearranged to obtain a complex channel state information matrix. Discrete wavelet decomposition is performed on the complex channel state subcarrier sequence of each access point to obtain wavelet multi-scale features. The received signal strength indication vector, the complex channel state information matrix, and the wavelet multi-scale features are standardized. The amplitude of the standardized channel state information matrix is ​​enhanced according to the standardized wavelet multi-scale features to obtain the amplitude of the enhanced channel state information matrix. The scaling factor is calculated according to the amplitude of the enhanced channel state information matrix to scale the real and imaginary parts of the standardized channel state information matrix to obtain the enhanced complex channel state information matrix of each access point. S3. Based on the standardized received signal strength indication vector and the enhanced complex channel state information matrix, local features and global wireless fingerprint features at the access point level are obtained. Based on the local features at the access point level, the access point confidence weight is calculated and the backbone positioning features are obtained through the local-global gating fusion mechanism and gating multi-scale feature extraction. Token-level relationship features are obtained through the Token relationship modeling module. The Token-level relationship features are fused into the backbone positioning features to obtain the enhanced deep positioning features. S4. Construct a joint loss function to train the deep network. Input the enhanced deep positioning features, the standardized received signal strength indication vector, and the enhanced complex channel state information matrix into the trained deep network regression head, deep feature tree regressor, and original fingerprint tree regressor to obtain the deep network prediction results, the deep feature tree regressor prediction results, and the original fingerprint tree regressor prediction results. Weight the results to obtain the final two-dimensional positioning coordinates.

2. The indoor positioning method based on multi-scale wireless channel feature fusion according to claim 1, characterized in that, The S2 includes: S21. The channel state information matrix is ​​rearranged according to the access point and the subcarrier dimension of the access point to obtain a complex channel state information matrix; wherein, the complex channel state information matrix includes a real part and an imaginary part; S22. Perform discrete wavelet decomposition on the complex channel state information subcarrier sequence of each access point to obtain low-frequency approximate components and multi-level high-frequency detail components, and concatenate the amplitudes of the low-frequency approximate components and multi-level high-frequency detail components to obtain wavelet multi-scale features. S23. Standardize the received signal strength indication vector, the real part of the complex channel state information matrix, the imaginary part of the complex channel state information matrix, and the wavelet multi-scale features to obtain the standardized received signal strength indication vector, the standardized channel state information matrix, and the standardized wavelet multi-scale features. S24. Linearly project the standardized wavelet multi-scale features onto the subcarrier dimension feature space with the same amplitude features as the standardized channel state information matrix, and enhance the amplitude of the standardized channel state information matrix according to the projected wavelet multi-scale features to obtain the enhanced channel state information matrix amplitude. Calculate the scaling factor according to the enhanced channel state information matrix amplitude, and scale the real and imaginary parts of the standardized channel state information matrix according to the scaling factor to obtain the enhanced complex channel state information matrix for each access point.

3. The indoor positioning method based on multi-scale wireless channel feature fusion according to claim 2, characterized in that, The complex channel state information matrix is ​​shown in equation (1) below: （1） In the formula, Indicates the first The access point Complex channel state information matrix on each subcarrier The magnitude of the complex channel state information matrix is ​​represented. Represents the imaginary unit. The phase of the complex channel state information matrix is ​​represented. Represents the real part of the complex channel state information matrix. This represents the imaginary part of the complex channel state information matrix.

4. The indoor positioning method based on multi-scale wireless channel feature fusion according to claim 1, characterized in that, The S3 includes: S31. Construct a complex local encoder for each access point, calculate the average received signal strength indication vector for each access point, and input the average received signal strength indication vector and the enhanced complex channel state information subcarrier features into the complex linear layer and complex normalization layer of the complex local encoder of the corresponding access point to obtain the access point-level local features. S32. Construct a global encoder by concatenating the standardized received signal strength indication vector with the amplitude vector of the enhanced complex channel state information matrix and inputting the concatenation into the global encoder to obtain the global wireless fingerprint feature. S33. Calculate the access point confidence weight based on the access point-level local features of each access point, and weight the access point-level local features according to the access point confidence weight to obtain the access point-level local weighted features. Construct local context features using the mean and standard deviation of the access point-level local features. S34. By controlling the fusion ratio of local weighted features at the access point level and global wireless fingerprint features through a local-global gating fusion mechanism, fusion features are obtained. The fusion features are then extracted with positioning-related features at different scales through multi-scale dilated convolution branches. The positioning-related features at different scales are fused to obtain gated multi-scale fusion features. The backbone positioning features are obtained based on the fusion features and the gated multi-scale fusion features. S35. Construct a Token sequence based on access point-level local features, global wireless fingerprint features, and gated multi-scale fusion features. Input the Token sequence into the Token self-attention module and the lightweight Mamba state space module to obtain relation features and sequence relation features. S36. The relation features and sequence relation features are fused into the backbone localization features through learnable gain to obtain enhanced deep localization features.

5. The indoor positioning method based on multi-scale wireless channel feature fusion according to claim 4, characterized in that, The step S33, which calculates the access point confidence weight based on the access point-level local features of each access point, includes: S331. For the local features at the access point level of each access point, the access point confidence level is calculated using the following formula (2): （2） In the formula, Indicates the first Confidence level of each access point This represents the access point confidence estimation network. Indicates the first Local features at the access point level for each access point Mean statistics representing local characteristics; S332. Based on the set temperature and access point confidence level, the normalized confidence weight of the access point is obtained as shown in the following formula (3): （3） In the formula, Indicates the first Normalized confidence weights for each access point This represents the softmax temperature parameter. Indicates the number of access points.

6. The indoor positioning method based on multi-scale wireless channel feature fusion according to claim 4, characterized in that, S34 includes: S341. Calculate the gating vector by concatenating the local weighted features at the access point level and the global wireless fingerprint features; S342. Based on the gating vector, the access point-level local weighted features and the global wireless fingerprint features are fused to obtain the fused features; S343. Multi-scale extraction is performed on the fused features using multiple convolutional branches with different dilation rates to obtain the multi-scale features extracted by each convolutional branch. S344. The multi-scale features of each convolutional branch are fused using learnable scale weights to obtain gated multi-scale fused features; S345. Based on the fusion features and gated multi-scale fusion features, obtain the backbone localization features.

7. The indoor positioning method based on multi-scale wireless channel feature fusion according to claim 1, characterized in that, The S4 includes: S41. Construct a joint loss function to train a deep network. Based on the trained deep network regression head, perform initial coordinate estimation on the enhanced depth localization features to obtain the deep network prediction results. S42. Train a deep feature tree regressor using the enhanced deep localization features and the prediction results of the deep network as input, and obtain the prediction results of the deep feature tree regressor based on the trained deep feature tree regressor. S43. The original wireless fingerprint features are used to train the original fingerprint tree regressor, which is composed of the normalized received signal strength indication vector, the amplitude of the enhanced complex channel state information matrix, the phase of the enhanced complex channel state information matrix and wavelet statistics. The prediction result of the original fingerprint tree regressor is obtained based on the trained original fingerprint tree regressor. S44. The prediction results of the deep network, the deep feature tree regressor, and the original fingerprint tree regressor are weighted according to the fusion weights automatically selected by the internal validation set of the training fold to obtain the final two-dimensional positioning coordinates.

8. An indoor positioning device based on multi-scale wireless channel feature fusion, wherein the indoor positioning device based on multi-scale wireless channel feature fusion is used to implement the indoor positioning method based on multi-scale wireless channel feature fusion as described in any one of claims 1-7, characterized in that, The device includes: The multi-access point wireless signal acquisition module is used to acquire the received signal strength indication vector and channel state information matrix of the terminal to be located through multiple Wi-Fi access points; The signal preprocessing and feature construction module is used to rearrange the channel state information matrix to obtain a complex channel state information matrix, perform discrete wavelet decomposition on the complex channel state subcarrier sequence of each access point to obtain wavelet multi-scale features, standardize the received signal strength indication vector, the complex channel state information matrix and the wavelet multi-scale features, and enhance the amplitude of the standardized channel state information matrix according to the standardized wavelet multi-scale features to obtain the enhanced channel state information matrix amplitude. Based on the enhanced channel state information matrix amplitude, the scaling factor is calculated to scale the real part and imaginary part of the standardized channel state information matrix to obtain the enhanced complex channel state information matrix of each access point. The adaptive positioning model module is used to obtain access point-level local features and global wireless fingerprint features based on the standardized received signal strength indication vector and the enhanced complex channel state information matrix. It calculates the access point confidence weight based on the access point-level local features and obtains the backbone positioning features through a local-global gating fusion mechanism and gating multi-scale feature extraction. It obtains token-level relationship features through the token relationship modeling module and integrates the token-level relationship features into the backbone positioning features to obtain enhanced deep positioning features. The hybrid regression output module is used to construct a joint loss function to train the deep network. It inputs the enhanced deep positioning features, the standardized received signal strength indication vector, and the enhanced complex channel state information matrix into the trained deep network regression head, the deep feature tree regressor, and the original fingerprint tree regressor to obtain the deep network prediction results, the deep feature tree regressor prediction results, and the original fingerprint tree regressor prediction results. The weighted results are then used to obtain the final two-dimensional positioning coordinates.

9. An indoor positioning device based on multi-scale wireless channel feature fusion, characterized in that, The indoor positioning device based on multi-scale wireless channel feature fusion includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

Images