Self-adaptive positioning method based on CSI, electronic equipment and storage medium

Abstract

The invention relates to the technical field of wireless network positioning, in particular to an adaptive positioning method based on CSI, electronic equipment and a storage medium. According to the method, high-discrimination fusion fingerprint data is generated through a fusion representation model according to amplitude information data and phase information data, and mapping between the high-discrimination fusion fingerprint data and corresponding physical positions in the environment is completed through a positioning model; the positioning model comprises an adaptive convolutional neural network composed of a source network and a target network, the source network processes data from a source domain, and the target network processes data from a target domain. The positioning model uses a self-adaptive convolutional neural network composed of a source network and a target network, so that the adaptability to the environment is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless network positioning, and in particular to a CSI-based adaptive positioning method, electronic equipment and storage medium. Background Art

[0002] Indoor positioning has become a key technology in the development of smart environments, playing a vital role in applications such as security monitoring, healthcare, human-computer interaction, and asset tracking. Traditional indoor positioning technologies mainly rely on active devices such as smartphones or wearable sensors, which use signals such as Wi-Fi, Bluetooth, or RFID to determine location. Although these methods have proven their effectiveness in many scenarios, they also have some limitations. Issues such as user compliance - requiring individuals to carry specific devices, privacy issues related to tracking personal devices, and the need for additional hardware all hinder the widespread application and practical deployment of these methods.

[0003] To overcome these challenges, device-free passive positioning (DFPL) has emerged as a promising alternative. DFPL systems do not require carrying any electronic devices, but rely on ambient signals present in the environment to detect and locate individuals. An important advancement in DFPL is the development of device-free passive Wi-Fi positioning technology (DFPWL). DFPWL uses channel state information (CSI) extracted from existing Wi-Fi infrastructure to detect and locate individuals without their active participation.

[0004] Although recent research on CSI-based DFPWL techniques has demonstrated its potential, existing methods still have some limitations. The first is the problem of insufficient fingerprint features. Many fingerprint-based positioning methods mainly focus on the amplitude or phase in CSI measurements. This limited focus may result in CSI fingerprints that lack sufficient discrimination ability and cannot accurately distinguish different locations. As a result, these methods usually require a large number of monitoring points (MPs) and access points (APs) to achieve acceptable positioning accuracy, which increases deployment complexity and maintenance costs. In addition, reliance on a large amount of infrastructure may also become an obstacle for implementation in environments where it is not possible to add a large number of MPs and APs. Second, environmental sensitivity is a major challenge faced by existing CSI-based positioning systems. CSI measurements are very sensitive to environmental changes, such as furniture rearrangement, door opening and closing, and the appearance of new obstacles, which may change the signal propagation path and cause multipath effects. These environmental dynamics can significantly degrade positioning accuracy because the CSI fingerprints collected during the training phase may not match the changed environment at runtime. Existing positioning systems usually lack mechanisms to adapt to these changes, so they are less robust and reliable in dynamic environments with changing environmental conditions. Summary of the invention

[0005] Based on this, the present invention provides a CSI-based adaptive positioning method, an electronic device and a storage medium, which at least solve one problem in the prior art.

[0006] In a first aspect, the present invention provides a CSI-based adaptive positioning method, which comprises the following steps: Collect channel state information data of wireless signals in the environment, perform preprocessing, and obtain amplitude information data and phase information data; Constructing a fusion representation model, and using the amplitude information data and the phase information data to train the fusion representation model; wherein the fusion representation model generates high-discrimination fusion fingerprint data according to the amplitude information data and the phase information data; A positioning model is constructed, and the positioning model is trained using the high-resolution fused fingerprint data and its coordinate labels; wherein the positioning model completes the mapping between the high-resolution fused fingerprint data and the corresponding physical location in the environment; the positioning model includes an adaptive convolutional neural network composed of a source network and a target network, the source network processes data from a source domain, and the target network processes data from a target domain; wherein the data in the source domain is the channel state information data of the wireless signal in the environment, and the data in the target domain is the channel state information data of the wireless signal without a label; The channel state information data of the wireless signal to be located is obtained, and preprocessed to obtain amplitude information data and phase information data, and a trained fusion representation model and a trained positioning model are used to generate high-resolution fusion fingerprint data and predict the probability of the target location.

[0007] It should be noted that the environment refers to the area where target positioning is required, such as indoors. The positioning model uses an adaptive convolutional neural network composed of a source network and a target network to improve its adaptability to the environment.

[0008] In some optional embodiments, collecting channel state information data of wireless signals in an environment includes: Place a pair of Wi-Fi signal transceiver devices in the environment; Divide the environment into multiple training points; For each training point, wireless signal data is collected through a Wi-Fi signal transceiver, where the wireless signal data includes channel state information data.

[0009] It should be noted that the Wi-Fi signal transceiver device can be a wireless router and / or a computer with a wireless network card.

[0010] In some optional embodiments, preprocessing the channel state information data of the wireless signal includes: denoising and phase calibrating the channel state information data of the wireless signal, then splicing the obtained amplitude information data and phase information data into feature fingerprint data, and obtaining single target fingerprint data after normalization.

[0011] In some optional embodiments, the denoising is wavelet domain denoising (WDD).

[0012] In some optional embodiments, the fusion representation model includes a feature extractor for amplitude mode and phase mode, a feature fusion layer and a decoding network, wherein the feature extractor extracts high-level feature maps from the amplitude mode and the phase mode respectively. and , the feature fusion layer will and Splicing is performed along the channel dimension, as shown in formula (1); Among them, F represents the fusion feature, Concat(·) represents the concatenation operation; The decoding network is derived from the fusion features Reconstruct the output matrix .

[0013] In some optional embodiments, the ensemble represents the total loss of the model It is a weighted combination of reconstruction loss and triplet loss, as shown in formula (4); in, represents a hyperparameter; represents the reconstruction loss, calculated by formula (2); represents the triplet loss, calculated by formula (3); in, Indicates the number of training samples; Indicates i The reconstructed output of samples is Indicates i The true value of samples; Indicates The fusion features of samples; Represents the fusion features of the positive sample; Represents the fusion features of negative samples; represents a margin hyperparameter; represents the positive partial function; i Represents a positive integer.

[0014] In some optional embodiments, in the positioning model, a weight regularization term is used Minimize the difference between the parameters of the transferable layers in the source network and the target network, as shown in Equation (5); in, Represents a collection of transferable layers; and Represents the source network and the target network respectively. The weight parameters of the layer; and denote the layer-specific scaling and shifting parameters learned during training, respectively; exp(·) denotes an exponential function; represents the weight coefficient, calculated by formula (7); in, , (·) indicates a meta-network; j and k Represent positive integers respectively.

[0015] In some optional embodiments, in the localization model, the maximum mean difference is used to measure and minimize the distance between the feature representation distributions in the source network and the target network, and the domain difference loss based on the maximum mean difference is calculated according to formula (9): ; in, and Represents the number of training set samples in the source domain and the target domain respectively; and Respectively represent Source samples and Feature representation of target samples; and Respectively represent i' Source samples and j' Feature representation of target samples; i' and j' Represent positive integers respectively; use the standard radial basis function kernel to calculate , as shown in formula (10); in, u express or , v express , or ; Represents the width parameter of the kernel function.

[0016] In a second aspect, the present invention provides an electronic device, comprising: at least one processor; and a memory communicatively coupled to the at least one processor; The memory stores instructions, and when the instructions are executed by at least one processor, the CSI-based adaptive positioning method as described above is implemented.

[0017] In a third aspect, the present invention provides a computer-readable storage medium storing instructions, which, when executed by a processor, implement the CSI-based adaptive positioning method as described above.

[0018] Due to the adoption of the above technical solution, the embodiments of the present invention have at least the following beneficial effects: (1) Only one transmitter and one receiver are used, which not only reduces deployment and maintenance costs, but also simplifies the practical implementation of the positioning system in various indoor environments; (2) A fusion representation model is adopted to combine the amplitude and phase information of CSI measurements, capturing the often-ignored spatiotemporal redundancy and channel correlation, thereby enhancing the fingerprint differentiation capability, achieving more accurate positioning, and reducing the number of required monitoring points (MPs) and access points (APs), which not only reduces deployment and maintenance costs, but also simplifies the implementation process in various environments; (3) By using unsupervised domain adaptation and meta-learning dual-stream structured adaptive convolutional neural networks, the positioning model can adapt to changes in the CSI link model and sample features caused by environmental changes. This adaptability ensures that the positioning model remains effective even when the environment changes significantly (such as furniture movement or occupancy pattern changes), and can maintain high positioning accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the architecture of the fusion representation model in an embodiment of the present invention.

[0020] Figure 2 Schematic diagram of the architecture of LC-Net in an embodiment of the present invention.

[0021] Figure 3 Schematic diagram of the architecture of AdaptCNN in an embodiment of the present invention.

[0022] Figure 4 Schematic diagram of the layout of a conference room in an embodiment of the present invention.

[0023] Figure 5 is a cumulative distribution function image of positioning error of different methods in a multi-day environment in the embodiment of the present invention, where: Figure 5(a) is the cumulative distribution function image of the positioning error of the re-collected and trained (Baseline) method. Figure 5 (b) is the cumulative distribution function image of the positioning error of the fine-tuning method. Figure 5 (c) is the cumulative distribution function image of the positioning error of the retraining method. Figure 5 (d) is the cumulative distribution function image of the positioning error without the adaptation method. Figure 5 (e) in the figure is the cumulative distribution function image of the positioning error of the AdaptCNN method.

[0024] Figure 6 is a cumulative distribution function image of positioning errors of different systems under a multi-day environment in an embodiment of the present invention, where: Figure 6 (a) is the cumulative distribution function image of the positioning error of different systems under the first day environment. Figure 6 (b) is the cumulative distribution function image of the positioning error of different systems in the second day environment. Figure 6 (c) is the cumulative distribution function image of the positioning error of different systems under the third day environment. Figure 6 (d) is the cumulative distribution function image of the positioning error of different systems under the fourth day environment. Figure 6 (e) is the cumulative distribution function image of the positioning error of different systems under the fifth day environment. Figure 6 (f) in the figure is the cumulative distribution function image of positioning errors of different systems in the sixth day environment. DETAILED DESCRIPTION

[0025] The following is a clear and complete description of the concept of the present invention and the technical effects produced, so as to fully explain the purpose, scheme and effect of the present invention.

[0026] According to an embodiment of the present invention, a CSI-based adaptive positioning method is provided, in which only one wireless signal transmitter and one wireless signal receiver are used to collect channel state information (CSI) amplitude and phase samples of wireless signals in an environment to construct a fingerprint database.

[0027] The CSI data is preprocessed to improve the quality and reliability of the input features. To mitigate the noise and anomalies in the amplitude component of the CSI measurement, wavelet domain denoising (WDD) is used for denoising. Wavelet threshold denoising effectively distinguishes between real signals and noise by decomposing the signal into different frequency components and thresholding the wavelet coefficients. This denoising step retains the essential features of the signal while significantly reducing high-frequency noise, thereby maintaining the integrity of the amplitude information that is critical for positioning.

[0028] For the phase component, phase calibration is performed to obtain stable and reliable phase information, which can be effectively used for further analysis. The calibrated phase information data enhances the robustness of the system by providing additional discriminative features and complements the amplitude information data.

[0029] Two different feature matrices are generated from the pre-processed amplitude and phase information, one for amplitude information data and the other for phase information data. These matrices capture different aspects of the signal characteristics and are subsequently fused using a fusion representation model. This fusion process combines the complementary information of amplitude and phase measurements to generate highly discriminative fused fingerprint features, called high-discriminative fused fingerprints (HDFFs). HDFFs encapsulate the temporal and spatial redundancy and channel correlation inherent in CSI data, providing a rich representation that enhances the system's ability to distinguish between different locations. These HDFFs are then used to train a convolutional neural network (CNN) localization model. CNNs learn complex mappings between fused fingerprints and corresponding physical locations in the environment, laying the foundation for accurate positioning during the operational phase.

[0030] like Figure 1 As shown in the figure, the fused representation model in the offline training phase includes a feature extractor for each modality, a feature fusion layer, a decoding network, and a dedicated loss function designed to enhance the discriminative ability of learned features. In the initial stage, for each modality (amplitude and phase), a convolutional neural network (CNN) is used as a feature extractor to obtain high-level feature representations from their respective input matrices, denoted as A (amplitude) and B (phase), respectively. The CNN architecture of the two feature extractors is the same, consisting of multiple convolutional layers and maximum pooling layers with ReLU activation functions, a pooling window size of 2×2, and a stride of 2. The convolutional layer is responsible for learning local patterns and feature hierarchies in the data, while the pooling layer reduces the computational complexity by reducing the spatial dimension and mitigates the risk of overfitting. The output of each modality feature extractor is a high-level feature map and , capturing the salient features of amplitude and phase information data respectively. In order to effectively combine these features, a feature fusion layer is used to and Splicing is performed along the channel dimension, as shown in formula (1); Among them, Concat(·) represents the concatenation operation.

[0031] The obtained fusion features The dimension of is 7 × 7 × 64, encapsulating the comprehensive information of amplitude and phase modes. This fusion enhances the richness of representation by integrating complementary features and can improve the discriminative ability of the localization model.

[0032] After fusion, a decoder network is introduced, which extracts the fused features Reconstruct the output matrix The goal of the decoder is to generate an output matrix of size 30×30×3, corresponding to the dimensions of the original input matrix. The decoder network consists of convolutional layers with ReLU activation functions and upsampling layers, gradually increasing the spatial dimensions. The last layer is a convolutional layer with a Sigmoid activation function, with an output range of [0, 1], which is suitable for reconstruction tasks. Padding or cropping operations are applied as needed to adjust the output size to match the required dimensions.

[0033] In order to effectively train the fused representation model, a combination of reconstruction loss and triplet loss is adopted.

[0034] Reconstruction loss Reconstruction output and the true value The mean square error (MSE) between them is used to calculate, as shown in formula (2); in, Indicates the number of training samples; Indicates i The reconstructed output of samples is Indicates i Since reconstruction does not have a fixed true output, the input of any modality is used or As a reference The reconstruction loss encourages the model to generate realistic outputs that are consistent with the original input data, thereby ensuring that the fused features retain key information from the input.

[0035] Triplet loss Calculate the fusion feature The triplet loss encourages the model to generate feature embeddings that make the distances between samples of the same class (i.e., similar positions) closer, and the distances between samples of different classes (i.e., different positions) farther. The triplet loss is defined as: in, Indicates The fusion features of samples (anchor points); Represents the fused features of positive samples (i.e., samples from the same location); Represents the fused features of negative samples (i.e., samples from different locations); represents a margin hyperparameter that defines the minimum expected separation between positive and negative sample pairs (set to ); represents the positive part function, ensuring that only positive values ​​contribute to the loss.

[0036] The total loss used to train the ensemble representation model It is a weighted combination of reconstruction loss and triplet loss, as shown in formula (4); Among them, the hyperparameters It is used to balance the influence of the two loss terms. Since the main goal is to enhance the feature discrimination ability to improve positioning accuracy, in this embodiment, , thus emphasizing more on triplet loss during training.

[0037] By integrating feature extraction, feature fusion, and specialized loss functions in this deep learning framework, the fused representation model is able to effectively combine data from both amplitude and phase modalities. The introduction of triplet loss ensures that the learned features are highly discriminative, allowing the model to more accurately distinguish between different locations. The fusion of amplitude and phase information captures the full spectrum of CSI features and fully exploits the complementarity of the two modalities. Amplitude information data provides information about signal strength and attenuation, while phase information data reveals signal propagation paths and multipath effects. By combining these two aspects, the model is able to fully understand the conditions of the wireless channel, resulting in more robust and accurate positioning results.

[0038] A convolutional neural network (CNN) is trained using the high-discriminative fused fingerprint (HDFF) from the fused representation model and its corresponding coordinate labels, which is called the localization model (LC-Net). This approach enables the network to learn the complex spatial relationships and nonlinear mappings inherent in the fused fingerprint, thereby improving localization accuracy. Figure 2 As shown in the figure, LC-Net consists of three convolutional blocks, each of which contains a convolutional layer followed by a layer with a LeakyReLU activation function. A key design choice is that no pooling layers are used in the network architecture. Although pooling layers are effective in reducing computational complexity and capturing abstract features in image classification tasks, they may lead to a loss of spatial resolution. In localization tasks, it is crucial to preserve detailed spatial information because small differences in the fused fingerprint correspond to different physical locations. After the convolutional block, LC-Net contains two fully connected layers that act as regression layers to estimate the precise location coordinates of the target. The output of the final fully connected layer provides the predicted Coordinates, locate the precise position of the target in the indoor environment and achieve accurate positioning.

[0039] During the training of LC-Net, the network weights are initialized using a Gaussian distribution with zero mean and a standard deviation suitable for the network size. Stochastic gradient descent (SGD) is used as the optimization algorithm, and the learning rate is set to 0.0001 to ensure stable convergence. The batch size is set to 25 to balance the training efficiency and the accuracy of the gradient estimation. The mean square error (MSE) is used as the loss function to quantify the difference between the predicted coordinates and the true label. This training configuration effectively maps complex fused fingerprint features to precise location coordinates, improving the overall positioning performance. By excluding the pooling layer to retain detailed spatial information and adopting a regression-based CNN model, high positioning accuracy suitable for dynamic indoor environments (such as industrial environments where environmental conditions may change frequently) is achieved.

[0040] After training the localization model LC-Net with the initial fingerprint database, a robust mapping between the fused fingerprint features and the corresponding coordinates is established. In order to further maintain high localization accuracy in dynamic environments, an adaptive convolutional neural network (AdaptCNN) is constructed, aiming to utilize the knowledge retained by the intermediate layers of LC-Net and adapt to new fingerprint distributions through unsupervised domain adaptation methods. Figure 3 As shown in the figure, the AdaptCNN framework consists of a two-stream architecture, including a source stream and a target stream. The source stream processes data from the original fingerprint database (source domain) and uses an existing positioning model trained in the offline stage; the target stream processes data from the target domain training set, which causes changes in the CSI distribution due to environmental changes. By jointly training these two streams, AdaptCNN uses a meta-learning method to automatically identify and transfer key layers and parameters from the source network to the target model. This process minimizes the differences between the corresponding layers while retaining the key knowledge obtained from the source domain. The meta-learning mechanism enables the model to adapt to new domains with almost no labeled data or even relying entirely on unlabeled data.

[0041] In order to prevent excessive deviation between the weights of the corresponding layers in the source network and the target network, a weight regularization term is introduced to constrain the parameters of the target network. This constraint helps to reduce overfitting in the target domain, especially when there are few samples, and encourages the target network to retain useful knowledge from the source network. represents the sample set in the source domain (original fingerprint database), Represents the number of samples in the set, and the corresponding label is ,in Representation and Sample Related location information. Similarly, represents the set of samples from the target domain (new fingerprint distribution), represents the number of samples in the set, which are unlabeled. The goal is to learn a target model that can accurately predict the location information of samples in the target domain despite changes in the data distribution. and Represents all layers in the source network and the target network respectively Parameters (weights and biases).

[0042] Weight regularization term Defined as minimizing the difference between the parameters of the transferable layers in the source and target networks, the regularization term is expressed as: in, Represents a collection of transferable layers; and Represents the source network and the target network respectively. The weight parameters of the layer; and denotes layer-specific scaling and shifting parameters (hyperparameters) learned during training, which allow flexible transformation of the parameters of the source network to better match the target domain; represents a layer-specific weight coefficient that controls the importance of each layer in the regularization term, which prioritizes certain layers based on their relevance to the target task; exp(·) represents an exponential function that ensures that the regularization penalty increases rapidly as the difference increases, thereby encouraging the parameters of the target network to remain close to those of the source network.

[0043] Weight coefficient is determined dynamically through a compact meta-network whose parameters are The meta-network The source network The layer's attributes are taken as input and output , as shown in formula (6); In order to ensure that the coefficients form an effective probability distribution between layers, a softmax function is applied for normalization, as shown in formula (7); in, This normalization ensures that .

[0044] In order to align the feature representations between the source and target domains, the maximum mean difference (MMD) is used to measure and minimize the distance between the feature representation distributions in the two domains. By reducing this difference, the knowledge transfer from the source domain to the target domain is promoted. The MMD measurement is based on the probability distribution distance between the samples drawn from each distribution, which is defined as the squared distance between the mean embeddings of the two distributions in the Reproducing Kernel Hilbert Space (RKHS). Domain difference loss based on MMD As shown in formula (8); in, and Respectively represent Source samples and The feature representation of a target sample (e.g. the output of a certain layer); The kernel function The resulting RKHS (Reproducing Kernel Hilbert Space) mapping.

[0045] By expanding the squared norm, the domain difference loss can be calculated as: in, and Respectively represent i' Source samples and j' Feature representation of target samples; i' and j' represent positive integers, i' and j' They are used to distinguish the features of different input samples in the same layer of the same network; they are calculated using the standard radial basis function (RBF) kernel. , as shown in formula (10); in, u express or , v express , or ; is the width parameter of the kernel function, which is set to .

[0046] By minimizing , encouraging the feature representations of the source and target domains to be similar, thereby facilitating the transfer of knowledge.

[0047] The overall loss function for training the source and target networks is: in, represents the loss associated with the source flow, which in this embodiment is the mean square error (MSE) between the predicted coordinates and the true coordinates of the source domain samples; represents the weight regularization term; Represents domain difference loss; and are the hyperparameters that represent the contribution of the balanced regularization term and the domain-differentiation term, respectively.

[0048] parameter , and By minimizing the loss function To jointly optimize, the optimization process uses back propagation. In this embodiment, the Adam optimizer is used, and the learning rate is set to The weights of the target network Use the source network weights from the offline training phase Initialize. Each layer Meta-network parameters and Initialize to and , thereby achieving identity mapping at the beginning.

[0049] The training process involves the following steps of alternating optimization of the source network, target network, and meta-network parameters: Input source domain dataset , target domain dataset , set the learning rate ; initialization , , and and meta-network parameters; Repeat the following steps until the training is complete: a. Randomly select small batches of data and , each batch size is ; b. Calculate the source loss ; c. Use and exist The weight regularization loss is calculated on the two mini-batch datasets ; d. Use and exist Computational domain difference loss function on two mini-batch datasets ; e. Calculate total loss ; f. Minimize using backpropagation ,renew , and .

[0050] In order to evaluate the positioning performance of the CSI-based adaptive positioning method (the method of the embodiment of the present invention), a series of controlled experiments were conducted. The evaluation focused on the robustness to environmental changes, the effectiveness in adapting to new environmental conditions, and the overall efficiency of the system in terms of deployment and maintenance overhead.

[0051] like Figure 4 As shown, experiments were conducted in a conference room of approximately 120 square meters, which is equipped with a large conference table, multiple stools and cabinets, providing a typical cluttered indoor environment; such an environment poses challenges for positioning due to multipath propagation and signal attenuation caused by obstacles.

[0052] For the experimental hardware, a TP-LINK WR841N router configured as a transmitter was used. The router operates in the 2.4 GHz frequency band, representing a standard Wi-Fi infrastructure commonly found in indoor environments. The receiver is a laptop equipped with an Intel 5300 network interface card (NIC) and running Ubuntu 14.04.

[0053] In the conference room, 56 reference points (RPs) were selected and marked for training and 20 test points (TPs) for evaluation; these points were evenly distributed in the open area of ​​the conference room, and the locations of these points covered the entire area, ensuring that the collected data can represent various signal propagation conditions in the conference room.

[0054] To collect CSI data, the CSI Tools software suite was used, which can extract CSI measurements from an Intel 5300 NIC. During data collection, the transmitter (router) used a single antenna configuration to continuously send packets at a rate of 100 packets per second; the receiver (laptop) was equipped with three antennas to receive and record CSI data from multiple spatial streams.

[0055] At each training point, a human target is placed in a stationary position and CSI data is collected for about 15 seconds, with approximately 1,500 CSI samples collected at each point. Each CSI sample contains the amplitude and phase information of 30 subcarriers and three receiving antennas, and the total number of CSI amplitude samples for each data packet is 30×3. The collected CSI samples are then processed and converted into a two-dimensional matrix that captures the spatial and frequency domain characteristics of the wireless channel. Based on these matrices, 50 fingerprint matrices are generated for each training point as input to the fusion model in the subsequent data preprocessing stage.

[0056] In order to evaluate the adaptability of the method of the embodiment of the present invention under changing environmental conditions, experiments were conducted over six consecutive days. During this period, common changes in real indoor environments, such as moving furniture (conference tables, stools, and cabinets), opening and closing doors, etc., were simulated to simulate the impact of actual changes on the distribution of CSI fingerprints, which challenges the adaptability of the positioning model. Importantly, unlabeled CSI samples are collected only when the environment changes. Specifically, the target is randomly placed at one of the 56 reference points, and 150 CSI samples are collected each time without marking the reference point coordinates. Compared with recalibrating all training points, this method greatly reduces the workload and requires almost no manual intervention.

[0057] In the experiment, the meta-network in AdaptCNN is configured as a single-layer fully connected neural network for each transmission channel between the source network (original model) and the target model (adapted model). Specifically, for each layer in the source network , the meta-network takes the feature map as input and outputs a set of layer-specific weights To ensure that these weights are properly normalized and contribute positively to the adaptation process, a softmax function is applied to the outputs to enforce the constraint ,in Indexing Layer This normalization ensures that the weights can be interpreted as a probability distribution over features, facilitating the adaptation process. In addition, the ReLU6 activation function is used to ensure that the weights are strictly positive. The ReLU6 function limits the output value to a maximum value of 6, thus providing an upper bound for the activation value. Therefore, the meta-network learns to adjust the feature representation from the source network to better align with the target domain, allowing the positioning system to maintain high accuracy even when the CSI fingerprint changes due to environmental changes.

[0058] On the first day, CSI data of all 56 training points were collected and the localization model was trained using this comprehensive dataset. The localization model trained on the first day is called the baseline localization model, which serves as a reference point for the localization model performance in the following days. From the second to the sixth day, environmental changes were introduced to simulate common indoor environmental dynamic conditions. Additional unlabeled CSI fingerprint data was collected after each environmental change, and the unlabeled CSI fingerprint data was used to adapt the existing model to the updated environment.

[0059] In order to evaluate the adaptability and effectiveness of AdaptCNN in the embodiment of the present invention, it is compared with the performance of the following alternative methods: No-adaptation: Directly use the positioning model trained on the first day without any adaptation to the changes in CSI fingerprints, aiming to illustrate the impact of environmental changes on positioning accuracy; Re-training: Every day, the localization model is retrained from initialization using the target domain training set collected after the environment changes. Fine-tuning: Use the target domain training set collected every day to fine-tune the existing model; Recollect and train (Baseline): Recollect CSI fingerprints from all training points and retrain the model from initialization every day; this method represents an ideal situation where the model is fully updated with comprehensive new data as a performance baseline.

[0060] Table 1 shows the average positioning error of different methods during the six-day experiment. It can be seen that the average positioning error of the baseline method is about 55 cm; this performance is due to the rich fingerprint features obtained by two-dimensional processing of CSI data and the model is retrained every day with complete and updated data. In contrast, AdaptCNN achieves an average positioning error of about 77 cm when only a small number of unlabeled CSI samples are used for adaptation; this result demonstrates the strong robustness and effectiveness of AdaptCNN by minimizing the recalibration workload while maintaining positioning accuracy; the error is slightly increased compared to the baseline method, but this is acceptable considering the significant reduction in data collection and annotation workload. Other methods, such as fine-tuning and retraining, show higher positioning errors and large performance fluctuations; these methods rely on smaller training datasets and are prone to overfitting and accuracy fluctuations, especially when the CSI fingerprint changes significantly. The method without adaptation performs the worst, highlighting the negative impact of environmental changes on positioning models and emphasizing the importance of using adaptive techniques in dynamic environments.

[0061] Table 1 Average positioning error of different methods in multi-day environment (cm) Figure 5 The cumulative distribution function (CDF) of the positioning error of different methods over six days is shown. It can be seen that although AdaptCNN significantly reduces the required human intervention, its average positioning error only increases by about 22 cm compared to the baseline method. About 80% of the AdaptCNN positioning errors fall within 1 meter, exceeding the performance of the retraining and fine-tuning methods, which have lower probabilities due to overfitting and instability of small data sets. These results highlight the necessity and effectiveness of adapting the model to changes in the CSI fingerprint distribution in order to maintain acceptable positioning accuracy.

[0062] To further evaluate the performance of the method (ADCLoc) of the embodiment of the present invention, it is compared with existing positioning methods (CrossSense, TDLLoc and MFFALoc). These comparisons are performed during the same six-day experiment using the same CSI dataset and in consistent indoor environmental conditions to ensure a fair evaluation.

[0063] As shown in Table 2, on the first day, when the CSI fingerprint remained unchanged, ADCLoc achieved the lowest average positioning error of 53.8 cm, significantly better than CrossSense (152 cm), TDLLoc (111.8 cm), and MFFALoc (77.2 cm); this result shows that ADCLoc has excellent positioning accuracy even in static environments. From the second to the sixth day, after introducing the CSI fingerprint changes caused by environmental changes, ADCLoc maintained an average positioning error of about 77 cm; in contrast, the positioning errors of other methods increased significantly, with CrossSense's error averaging about 190 cm, TDLLoc about 145 cm, and MFFALoc about 116 cm; these increases show that other methods are less effective in adapting to environmental changes.

[0064] Table 2 Average positioning error of CrossSense, TDLLoc, MFFALoc and ADCLoc (cm) Figure 6 The cumulative distribution functions (CDFs) of the positioning errors of different methods during the multi-day experiment are shown. It can be seen that ADCLoc consistently outperforms CrossSense, TDLLoc, and MFFALoc in terms of positioning performance and stability. The impact of CSI fingerprint changes is clearly reflected in the positioning models of other methods, resulting in performance degradation. In contrast, ADCLoc effectively adapts to these changes and maintains high-precision positioning.

[0065] These results demonstrate the robustness and adaptability of ADCLoc in dynamic environments. By leveraging 2D feature processing, using triplet loss for feature learning, and implementing unsupervised domain adaptation through the AdaptCNN method, ADCLoc effectively addresses the challenges posed by environmental changes and CSI fingerprint changes. ADCLoc maintains high positioning accuracy while requiring minimal additional data collection and no full retraining, which demonstrates the feasibility of ADCLoc in real-world applications.

[0066] In order to study the influence of model structure on positioning accuracy, five convolutional neural networks (CNNs) with different architectures are designed, marked as structure A to structure E, as shown in Table 3. Starting from the basic LeNet-5 architecture (structure A), the network is gradually modified by replacing larger convolution kernels with smaller ones, increasing the number of channels in each layer, and removing the pooling layer to maintain spatial information. On the basis of LeNet-5, all convolution layers are modified to smaller convolution kernels (3×3) and the number of channels in the second layer is increased to 32, thus obtaining structure B; the first layer maintains a 5×5 convolution kernel, the next two convolution layers are changed to smaller convolution kernels (3×3), and the number of neurons in the last fully connected layer is further increased to 128, thus obtaining structure C; all convolution layers are modified to smaller convolution kernels (3×3) and the number of channels in the third convolution layer is modified to 64, and all pooling layers are removed to obtain structure D. On the basis of structure D, the number of channels in the first convolution layer is increased to 16, and the number of neurons in the last fully connected layer is modified to 256, thus obtaining structure E. The performance of ADCLoc was evaluated over a six-day period using each network structure, and the average localization error is shown in Table 4.

[0067] Table 3 Different positioning model structures Table 4 Average positioning error under different model structures (cm) The analysis results show that the positioning accuracy gradually improves from structure A to structure E. Structure A is based on the standard LeNet-5 architecture and has the highest positioning error, indicating that this architecture may not be suitable for CSI-based positioning tasks. As the network is modified to include smaller convolution kernels (e.g., 3×3), increase the number of channels, and remove the pooling layer, the positioning accuracy improves. It is worth noting that structure E achieves the best and most stable performance, with the average positioning error always lower than other structures; this structure uses small convolution kernels, a large number of channels ([16, 32, 64, 256]), and no pooling layer; the absence of a pooling layer helps maintain the spatial resolution of the feature map, which is crucial for capturing small differences in CSI fingerprints corresponding to different locations. This analysis shows that the choice of positioning model architecture has a significant impact on the performance of ADCLoc. By optimizing the network structure, especially adopting a design similar to structure E, the positioning accuracy and robustness of the system in dynamic environments are improved.

[0068] The above is only a preferred embodiment of the present invention. The present invention is not limited to the above implementation. As long as the technical effect of the present invention is achieved by the same or equivalent means, it should belong to the protection scope of the present invention. Within the protection scope of the present invention, its technical scheme and / or implementation method can have various modifications and changes.

Claims

1. A CSI-based adaptive positioning method, characterized in that: The following steps are involved: Collect channel state information data of wireless signals in the environment, perform preprocessing, and obtain amplitude information data and phase information data; Constructing a fusion representation model, and using the amplitude information data and the phase information data to train the fusion representation model; wherein the fusion representation model generates high-discrimination fusion fingerprint data according to the amplitude information data and the phase information data; A positioning model is constructed, and the positioning model is trained using the high-resolution fused fingerprint data and its coordinate labels; wherein the positioning model completes the mapping between the high-resolution fused fingerprint data and the corresponding physical location in the environment; the positioning model includes an adaptive convolutional neural network composed of a source network and a target network, the source network processes data from a source domain, and the target network processes data from a target domain; wherein the data in the source domain is the channel state information data of the wireless signal in the environment, and the data in the target domain is the channel state information data of the wireless signal without a label; The channel state information data of the wireless signal to be located is obtained, and preprocessed to obtain amplitude information data and phase information data, and a trained fusion representation model and a trained positioning model are used to generate high-resolution fusion fingerprint data and predict the probability of the target location.

2. The method according to claim 1, characterized in that The channel state information data of the wireless signal in the collection environment includes: Place a pair of Wi-Fi signal transceiver devices in the environment; Divide the environment into multiple training points; For each training point, wireless signal data is collected through a Wi-Fi signal transceiver, where the wireless signal data includes channel state information data.

3. The method according to claim 2, characterized in that Preprocessing the channel state information data of the wireless signal includes: denoising and phase calibration of the channel state information data of the wireless signal, then splicing the obtained amplitude information data and phase information data into feature fingerprint data, and obtaining single target fingerprint data after normalization.

4. The method according to claim 3, characterized in that Denoising is done in wavelet domain.

5. The method according to claim 1, characterized in that: The fusion representation model includes feature extractors for amplitude mode and phase mode, feature fusion layer and decoding network. The feature extractor extracts high-level feature maps from amplitude mode and phase mode respectively. and , the feature fusion layer will and Splicing is performed along the channel dimension, as shown in formula (1); in, F represents fusion features, Concat(·) represents concatenation operation; The decoding network is derived from the fusion features Reconstruct the output matrix .

6. The method according to claim 5, characterized in that The ensemble represents the total loss of the model It is a weighted combination of reconstruction loss and triplet loss, as shown in formula (4); in, represents a hyperparameter; represents the reconstruction loss, calculated by formula (2); represents the triplet loss, calculated by formula (3); in, Indicates the number of training samples; Indicates i The reconstructed output of samples is Indicates i The true value of samples; Indicates The fusion features of samples; Represents the fusion features of the positive sample; Represents the fusion features of negative samples; represents a margin hyperparameter; represents the positive partial function; i Represents a positive integer.

7. The method according to claim 6, characterized in that In the positioning model, use the weight regularization term Minimize the difference between the parameters of the transferable layers in the source network and the target network, as shown in Equation (5); in, Represents a collection of transferable layers; and Represents the source network and the target network respectively. The weight parameters of the layer; and denote the layer-specific scaling and shifting parameters learned during training, respectively; exp(·) denotes an exponential function; represents the weight coefficient, calculated by formula (7); in, , (·) indicates a meta-network; j and k Represent positive integers respectively.

8. The method according to claim 7, characterized in that In the localization model, the maximum mean difference is used to measure and minimize the distance between the feature representation distributions in the source network and the target network. The domain difference loss based on the maximum mean difference is calculated according to formula (9): ; in, and Represents the number of training set samples in the source domain and the target domain respectively; and Respectively represent Source samples and Feature representation of target samples; and Respectively represent i' Source samples and j' Feature representation of target samples; i' and j' Represent positive integers respectively; use the standard radial basis function kernel to calculate , as shown in formula (10); in, u express or , v express , or ; Represents the width parameter of the kernel function.

9. An electronic device, characterized in that: include: at least one processor; and a memory communicatively coupled to the at least one processor; The memory stores instructions, which, when executed by at least one processor, implement the CSI-based adaptive positioning method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: Instructions are stored, and when the instructions are executed by the processor, the CSI-based adaptive positioning method as described in any one of claims 1-8 is implemented.

Images