Cross-scene radio frequency fingerprint positioning method based on self-supervised deep learning
Through the cross-scenario RF fingerprint positioning method based on self-supervised deep learning, robust RF features are extracted using self-supervised comparative learning and deep neural networks, which solves the problems of large positioning errors and high deployment costs of existing wireless positioning methods when the environment changes, and achieves high-precision, low-cost cross-scenario positioning.
Patent Information
- Application Number
- CN202510736746.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-04
- Publication Date
- 2025-09-05
AI Technical Summary
Existing wireless positioning methods based on fingerprint feature recognition have large positioning errors when the environment changes, and have high deployment and maintenance costs, making them difficult to adapt to complex and changing external environments.
A cross-scenario RF fingerprint positioning method based on self-supervised deep learning is proposed. By processing unlabeled RF data into RF images, a general wireless fingerprint feature extraction model is trained using a self-supervised contrastive learning algorithm, and an online location estimation network is trained using a small amount of labeled fingerprint features. An autoencoder and contrastive learning framework based on deep neural networks are designed, and the feature distance loss function is optimized to extract robust location features.
It achieves high-precision positioning in complex scenarios, reduces map construction and maintenance costs, improves the robustness and environmental adaptability of the positioning system, and reduces dependence on large amounts of labeled data.
Smart Images
Figure CN120602945A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of radio frequency fingerprint positioning technology, and in particular to a cross-scenario radio frequency fingerprint positioning method based on self-supervised deep learning. Background Art
[0002] For future 6G networks, deep learning-based wireless positioning technology holds great promise for intelligent services and location awareness. With its excellent nonlinear fitting and feature extraction capabilities, it has brought new breakthroughs to wireless fingerprint positioning. By designing deep neural networks (DNNs) to learn the underlying relationship between signal features and location information, the fingerprint feature representation capability and positioning accuracy have been effectively improved. However, obtaining neural networks that can dynamically change with the external environment and possess generalization capabilities still presents many challenges. Furthermore, the complex and ever-changing external environment can affect the propagation of wireless signals, severely hindering the efficient extraction of universal, robust, and accurate fingerprint features for wireless fingerprint positioning.
[0003] Fingerprint feature recognition-based positioning methods are not mature in commercial applications. This method usually collects radio frequency fingerprints with corresponding location information to build a radio frequency map, and then selects the best matching result to determine the location information. In fact, wireless signals are affected by multipath propagation during the propagation process, making it difficult to establish a positioning model with accurate spectrum signals. When the external environment changes, the positioning error increases, and it is necessary to manually collect radio frequency signals again and annotate the location information to build a new wireless fingerprint. This not only consumes a lot of time and manpower, but is also easily interfered by noise, making deployment impractical, which greatly limits its application in actual systems.
[0004] Therefore, to address the above problems, the present invention proposes a cross-scenario RF fingerprint positioning method based on self-supervised deep learning. Summary of the Invention
[0005] In order to overcome the problem of poor environmental adaptability of existing positioning methods based on fingerprint feature recognition, the present invention proposes a cross-scene radio frequency fingerprint positioning method based on self-supervised deep learning.
[0006] The technical solution of the present invention is: a cross-scenario radio frequency fingerprint positioning method based on self-supervised deep learning, comprising the following steps:
[0007] Processing unlabeled RF data into RF images;
[0008] A general wireless fingerprint feature extraction model is trained using a self-supervised contrastive learning algorithm;
[0009] Use a small amount of labeled fingerprint features to train an online location estimation network and provide location information query services.
[0010] Furthermore, the present invention processes a large amount of unlabeled RF data from different scenarios into RF images, and uses a self-supervised contrastive learning algorithm to train a universal wireless fingerprint feature extraction model to obtain high-quality, scenario-universal fingerprint feature representation. For specific positioning scenarios, the present invention designs an online position estimation network, uses only a small number of labeled fingerprint features for model training, and provides location information query services.
[0011] As an example, the specific steps of the local proximity detection are: for a training set of channel state information data D, divide it into where x i Represents the amplitude of the channel state information, and uses the radial basis function (RBF) to evaluate the closeness between each two channel state information samples, as shown in formula (1):
[0012]
[0013] Where σ is the neighborhood width;
[0014] The positive sample pairs and negative sample pairs are determined under the threshold γ constraint, as shown in formula (2):
[0015]
[0016] As a preference, an autoencoder based on a deep neural network is designed to learn robust position features from unlabeled channel state information samples; the high-dimensional channel state information samples are constructed as a normalized feature vector, and the feature vector X is used as the input of the autoencoder in each mini-batch input to train the fully connected network f θ (x) is used to encode the channel state information sample and map it into a low-dimensional feature vector z = f θ (x), where θ is a learnable hyperparameter in the autoencoder, and the low-dimensional representations of positive and negative samples are z p =f θ (x p ) and z n =f θ (x n ).
[0017] Preferably, the training process of the autoencoder includes: mapping channel state information samples into low-dimensional feature vectors using a fully connected network.
[0018] Preferably, the feature distances between positive and negative sample pairs are optimized through a contrastive learning framework.
[0019] Preferably, for the i-th anchor feature in the same batch, as shown in formula (3):
[0020]
[0021] The positive sample set is recorded as
[0022] The negative sample set is denoted as
[0023] The normalized cosine similarity of two vectors is defined as formula (4):
[0024]
[0025] Preferably, the loss function of the contrastive learning framework is Soft Nearest Neighbour Loss, which is used to minimize the feature distance of positive sample pairs and maximize the feature distance of negative sample pairs, as shown in formula (5):
[0026]
[0027] Where τ represents the temperature coefficient.
[0028] Preferably, the overall loss of the contrast loss function is the average loss of all anchor points in a batch. The losses of B anchor points in a batch are averaged to obtain the overall contrast loss as formula (6):
[0029]
[0030] Preferably, the downstream positioning task includes: receiving a channel state information query in a specific scenario When the present invention is Input into the trained deep neural network to obtain features Then, based on the Bayesian estimation, the posterior probability P(l r |v^) can get the corresponding position estimate, as shown in formula (7):
[0031]
[0032] Among them, P(l r ) is the uniform distribution of the reference point position information in the test scene, and its probability distribution can be expressed as formula (8):
[0033]
[0034] in the formula represents the channel state information fingerprint feature of the rth reference point. Finally, the position estimate is determined by the weighted average of the positions of all reference points, as shown in formula (9):
[0035]
[0036] Beneficial effects of the present invention:
[0037] 1. This invention excels in map construction and positioning tasks in complex scenarios. By using a large number of label-free radio frequency fingerprints to train the autoencoder, the generated radio frequency feature representation is high-quality, highly discriminative, and universally applicable to all scenarios. This allows for better cross-scenario positioning, thereby improving the positioning accuracy, robustness, and environmental adaptability of the positioning system.
[0038] 2. The present invention realizes wireless positioning based on a small number of tagged radio frequency fingerprints, significantly reducing the number of samples required to build a fingerprint map, saving collection costs, and reducing the heavy work of on-site surveys. At the same time, the network model trained by this method has strong robustness, and there is no need to repeatedly collect and maintain fingerprint maps due to environmental changes, thereby greatly reducing the deployment and maintenance costs of the positioning system. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 What is shown is a schematic diagram of the workflow of the present invention. DETAILED DESCRIPTION
[0040] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.
[0041] The present invention provides an embodiment: a cross-scenario radio frequency fingerprint positioning method based on self-supervised deep learning, comprising the following steps:
[0042] Processing unlabeled RF data into RF images;
[0043] A general wireless fingerprint feature extraction model is trained using a self-supervised contrastive learning algorithm;
[0044] Use a small amount of labeled fingerprint features to train an online location estimation network and provide location information query services.
[0045] Further, the operation process of the present invention is described, specifically:
[0046] First, for local proximity detection:
[0047] For a training set of channel state information data D, it is divided into where x iRepresents the amplitude of the channel state information. A radial basis function (RBF) is used to evaluate the closeness between each two channel state information samples. The specific formula is as follows:
[0048]
[0049] Where σ is the neighborhood width, and the corresponding logical label is determined under the threshold γ constraint, that is:
[0050]
[0051] Among them I ij =1 indicates a positive sample pair, otherwise it is a negative sample pair.
[0052] Second, let’s explain the autoencoder:
[0053] The present invention designs an autoencoder based on a deep neural network to learn position features from unlabeled channel state information samples. The learned features are highly robust and can adapt to changes in the environment. The high-dimensional channel state information samples are constructed into a normalized feature vector by using the channel state information amplitude data containing multiple antennas and multiple carriers. The feature vector X is used as the input of the autoencoder in each small batch input to train the fully connected network f θ (x) is used to encode the channel state information sample and map it into a low-dimensional feature vector z = f θ (x), where θ is a learnable hyperparameter in the autoencoder, and the low-dimensional representations of positive and negative samples are z p =f θ (x p ) and z n =f θ (x n ).
[0054] Third, explain the contrast loss:
[0055] The present invention adopts a contrastive learning framework, aiming to enable the model to learn essential features that are resistant to dynamic interference. These features can be effectively generalized to new scene positioning to achieve higher robustness. The feature learning model based on deep neural networks is optimized by the principle that positive sample pairs composed of adjacent anchor points are close to each other and negative sample pairs composed of distant anchor points are far away from each other, thereby learning the local proximity of signal features in signal state information data and maintaining robustness in dynamic environments. Therefore, the design of the contrastive loss function needs to satisfy the requirements of minimizing the feature distance of positive sample pairs while maximizing the feature distance of negative sample pairs.
[0056] For the i-th anchor feature in the same batch, as shown in formula (3):
[0057]
[0058] The positive sample set is recorded as
[0059] The negative sample set is denoted as
[0060] The normalized cosine similarity of two vectors is defined as formula (4)
[0061]
[0062] For each anchor point, the Soft Nearest Neighbour Loss function is defined as formula (5):
[0063]
[0064] where τ represents the temperature coefficient, which is used to adjust the sensitivity of the similarity between different representations.
[0065] The average loss of B anchor points in a batch is taken to get the overall contrast loss as formula (6):
[0066]
[0067] Third, the downstream positioning tasks are explained:
[0068] Use the learned signal features to estimate the position of different scenes and receive channel state information query in specific scenes When the present invention is Input into the trained deep neural network to obtain features Then, based on the Bayesian estimation, the posterior probability is calculated The corresponding position estimate can be obtained, as shown in formula (7):
[0069]
[0070] Among them, P(l r ) is the uniform distribution of the reference point position information in the test scene, and its probability distribution can be expressed as formula (8):
[0071]
[0072] in the formula represents the channel state information fingerprint feature of the rth reference point. Finally, the position estimate is determined by the weighted average of the positions of all reference points, as shown in formula (9):
[0073]
[0074] See also Figure 1, further, the workflow of the present invention is described:
[0075] First, a large amount of unlabeled RF signal data is collected, including multi-antenna and multi-carrier channel state information (CSI), and converted into RF images or normalized feature vectors for subsequent model processing.
[0076] The preprocessed channel state information data is divided, and the radial basis function (RBF) is used to calculate the proximity between samples. The positive sample pairs (similar samples) and negative sample pairs (dissimilar samples) are determined according to the set threshold, providing a training basis for self-supervised learning.
[0077] An autoencoder based on a deep neural network is constructed. High-dimensional channel state information samples are input into the model, which is mapped into low-dimensional feature vectors through the encoder part to learn wireless fingerprint feature representation with robustness and generalization ability.
[0078] A self-supervised contrastive learning framework is adopted and the Soft Nearest Neighbour Loss function is used to optimize the model, so that the features of positive sample pairs are close to each other in low-dimensional space and the features of negative sample pairs are far away from each other, thereby enhancing the feature discrimination ability and environmental adaptability.
[0079] During the training process, the contrast loss is calculated for each batch of anchor samples, and the model parameters are adjusted through back-propagation, ultimately enabling the model to extract high-quality, scenario-universal RF fingerprint features.
[0080] For specific positioning scenarios, only a small amount of labeled RF fingerprint data is needed to fine-tune the pre-trained model to adapt it to the location estimation needs of specific environments and reduce dependence on large amounts of labeled data.
[0081] In the actual positioning stage, the channel state information to be queried is input into the trained model, its feature representation is extracted, and the posterior probability is calculated based on Bayesian estimation. Combined with the fingerprint feature distribution of the reference point, the target position coordinates are determined by weighted average.
[0082] Because the model extracts highly universal feature representations through self-supervised learning, only minor adjustments are required to maintain high-precision positioning when the environment changes or new scenarios are deployed, significantly reducing the cost of repeated data collection and retraining.
[0083] Furthermore, the present invention provides an embodiment for positioning in an indoor shopping mall:
[0084] Multiple Wi-Fi access points are deployed in large shopping malls to collect unlabeled CSI data and preprocess it into RF images. A universal fingerprint feature model is trained through self-supervised comparative learning. In actual application, only a small number of location points need to be annotated in key areas (such as entrances, exits, and elevators). The model can use the learned robust features to achieve meter-level positioning accuracy. It can still work stably even when the mall layout is adjusted or the flow of people changes, thus significantly reducing the intensive sampling and maintenance costs required for traditional fingerprint positioning.
[0085] Furthermore, the present invention provides an embodiment for tracking equipment in smart factories:
[0086] In industrial plants, existing 5G base station signals are used to collect radio frequency data with obvious multipath effects. RBF functions are used to construct positive and negative sample pairs for training autoencoders. After the model learns the feature representation that is resistant to metal reflection interference, it only needs to mark 20% of the reference points on the AGV's route to achieve centimeter-level equipment positioning. When a new production line is installed or equipment is relocated, there is no need to re-build the map; the pre-trained model can be reused and fine-tuned to adapt.
[0087] Furthermore, the present invention provides an embodiment of a hospital emergency navigation system:
[0088] In a hospital building with a complex structure, Bluetooth beacon signals are used to build a cross-floor fingerprint database. Through comparative learning, the model can distinguish subtle signal differences in similar spaces (such as wards with the same number on different floors). When a disaster occurs, rescue workers enter an unfamiliar area with terminal devices. The system can quickly match real-time signals with sparsely labeled emergency exit reference points based on pre-trained features, thereby generating the optimal escape route within 10 seconds, breaking through the failure limitation of traditional fingerprint positioning in unfamiliar scenarios.
[0089] Furthermore, the present invention provides an embodiment of underground parking lot navigation:
[0090] In underground parking lots where GPS fails, the vehicle-mounted OBU is used to continuously collect unlabeled 4G / 5G signal fluctuation data, and dynamic fingerprint features including the Doppler effect are extracted through self-supervised learning. When the car owner enters for the first time, he only needs to mark 3-5 parking space coordinates. The model can then combine the motion trajectory to infer the fingerprint map of the entire site, thereby realizing the "reverse car search" function. The positioning error is still less than 1.5 meters when the vehicle density changes at different time periods.
[0091] Furthermore, the present invention provides an embodiment of cooperative positioning of a group of drones:
[0092] In the wild without base station coverage, the drone swarm receives each other's radio frequency signals as training data and uses the method of the present invention to learn universal feature representations that are not affected by terrain undulations. When new drones join the formation, they only need to share the pre-trained model and exchange a small number of real-time location tags to achieve sub-meter relative positioning within the formation, reducing communication overhead by 80% compared to traditional SLAM technology.
[0093] Through the above steps, by using a large number of label-free RF fingerprints to train the autoencoder, the generated RF feature representation is made to have high quality, strong discrimination and universal applicability in all scenarios, which can better achieve cross-scenario positioning, thereby improving the positioning accuracy, robustness and environmental adaptability of the positioning system, and solving the problem of poor environmental adaptability of existing fingerprint feature recognition-based positioning methods.
Claims
1. A self-supervised deep learning cross-scenario radio frequency fingerprint positioning method, characterized in that: The following steps are included: Processing unlabeled RF data into RF images; A general wireless fingerprint feature extraction model is trained using a self-supervised contrastive learning algorithm; Use a small amount of labeled fingerprint features to train an online location estimation network and provide location information query services.
2. The cross-scenario RF fingerprint positioning method based on self-supervised deep learning according to claim 1 is characterized in that: The self-supervised contrastive learning algorithm includes: performing local proximity detection on channel state information data, and then dividing positive sample pairs and negative sample pairs; evaluating the proximity between samples through radial basis functions, and determining logical labels based on thresholds.
3. The cross-scenario RF fingerprint positioning method based on self-supervised deep learning according to claim 2 is characterized in that: The specific steps of the local proximity detection are: for a training set of channel state information data D, divide it into where x i Represents the amplitude of the channel state information, and uses the radial basis function (RBF) to evaluate the closeness between each two channel state information samples, as shown in formula (1): Where σ is the neighborhood width; The positive sample pairs and negative sample pairs are determined under the threshold γ constraint, as shown in formula (2):
4. The cross-scenario RF fingerprint positioning method based on self-supervised deep learning according to claim 1 is characterized by: Design a deep neural network-based autoencoder to learn robust position features from unlabeled channel state information samples; The high-dimensional channel state information sample is constructed as a normalized feature vector, and the feature vector X is used as the input of the autoencoder in each small batch input to train the fully connected network f θ (x) is used to encode the channel state information sample and map it into a low-dimensional feature vector z = f θ (x), where θ is a learnable hyperparameter in the autoencoder, and the low-dimensional representations of positive and negative samples are z p =f θ (x p ) and z n =f θ (x n ).
5. The cross-scenario RF fingerprint positioning method based on self-supervised deep learning according to claim 4 is characterized in that: The training process of the autoencoder includes: using a fully connected network to map channel state information samples into low-dimensional feature vectors.
6. The cross-scenario RF fingerprint positioning method based on self-supervised deep learning according to claim 4 is characterized by: The feature distance between positive and negative sample pairs is optimized through a contrastive learning framework.
7. The cross-scenario RF fingerprint positioning method based on self-supervised deep learning according to claim 6 is characterized by: For the i-th anchor feature in the same batch, as shown in formula (3): The positive sample set is recorded as The negative sample set is denoted as The normalized cosine similarity of two vectors is defined as formula (4):
8. The cross-scenario RF fingerprint positioning method based on self-supervised deep learning according to claim 7 is characterized by: The loss function of the contrastive learning framework is Soft Nearest Neighbour Loss, which is used to minimize the feature distance of positive sample pairs and maximize the feature distance of negative sample pairs, as shown in formula (5): Where τ represents the temperature coefficient.
9. The cross-scenario RF fingerprint positioning method based on self-supervised deep learning according to claim 8 is characterized by: The overall loss of the contrast loss function is the average loss of all anchor points in a batch. The average loss of B anchor points in a batch is taken to get the overall contrast loss as formula (6):
10. The cross-scenario RF fingerprint positioning method based on self-supervised deep learning according to claim 9 is characterized in that: The downstream positioning task includes: receiving a channel state information query in a specific scenario When the present invention is Input into the trained deep neural network to obtain features Then, based on the Bayesian estimation, the posterior probability P(l r |v^) can get the corresponding position estimate, as shown in formula (7): Among them, P(l r ) is the uniform distribution of the reference point position information in the test scene, and its probability distribution can be expressed as formula (8): in the formula represents the channel state information fingerprint feature of the rth reference point. Finally, the position estimate is determined by the weighted average of the positions of all reference points, as shown in formula (9):