Radio frequency fingerprint positioning method based on self-supervised pre-training model

Abstract

The invention relates to the field of radio frequency fingerprint positioning technology and deep learning, in particular to a radio frequency fingerprint positioning method based on a self-supervised pre-training model, which comprises radio frequency image generation, self-supervised pre-training model and downstream positioning tasks. The method is excellent in performance in a deep fingerprint positioning task in a complex dynamic environment, a feature encoder is trained through a large amount of label-free radio frequency data, signal propagation characteristics are represented based on momentum contrast learning, signal features with high quality, strong discrimination and full-scene universality are obtained, and the method is suitable for the deep fingerprint positioning task in the complex dynamic environment. The characteristic significantly improves the positioning precision, robustness and universality of an indoor fingerprint positioning system, the self-supervised pre-training normal form is adopted to learn signal feature representation contained in radio frequency fingerprint data, a flexible dynamic channel dictionary search mechanism is designed, the high cost of constructing a fingerprint database is significantly reduced, and meanwhile, the positioning accuracy, robustness and universality of the indoor fingerprint positioning system are improved. The method is high in robustness and generalization, and the radio frequency fingerprint database does not need to be periodically updated and maintained.

Description

Technical Field

[0001] The present invention relates to the fields of radio frequency fingerprint positioning technology and deep learning, and in particular to a radio frequency fingerprint positioning method based on a self-supervised pre-training model. Background Art

[0002] With the development of 6G networks, RF fingerprint positioning technology based on deep learning has become a key technology for providing high-quality intelligent location services (LBS), and is widely used in indoor positioning, industrial Internet, positioning navigation and other fields. At present, RF fingerprint positioning technology based on deep learning is one of the key technologies with the greatest potential for large-scale application. By designing deep neural networks (DNN) to learn the intrinsic relationship between signal features and location information, the representation ability and positioning accuracy of fingerprint features in complex environments are effectively improved. However, wireless signal propagation in complex and changeable positioning environments is easily affected by external factors, which has seriously hindered the efficient extraction of universal, robust and accurate fingerprint features in RF fingerprint positioning.

[0003] Fingerprint positioning methods based on deep learning are not yet mature in commercial applications. The complex and changeable positioning environment places extremely high demands on the quality of RF fingerprint features. In order to achieve high-precision positioning, it is necessary to collect a large number of wireless fingerprints with precise location information in actual application scenarios and train deep neural networks to learn applicable deep fingerprint features. The traditional method of building a fingerprint database is time-consuming and labor-intensive, resulting in high system deployment costs. In fact, because the existing algorithms are sensitive to changes in spatiotemporal features, slight changes in time and environment may cause the fingerprint database to become invalid, requiring regular manual re-collection of fingerprints in the new environment, further increasing the maintenance cost of the fingerprint database. These problems severely limit its application in actual systems.

[0004] In response to the above challenges, especially the insufficient fingerprint feature representation capabilities in commercial applications, as well as the difficulties of large-scale data collection and efficient training, a radio frequency fingerprint positioning method based on a self-supervised pre-training model is designed. This method has high accuracy, high robustness and strong generalization ability, and can better cope with complex and changing positioning environments. Summary of the invention

[0005] In order to overcome the insufficient fingerprint-based feature representation capabilities in commercial applications, as well as the challenges of large-scale data collection and efficient training, new solutions can be designed.

[0006] The technical solution of the present invention is: a radio frequency fingerprint positioning method based on a self-supervised pre-training model, the steps are as follows:

[0007] RF image generation:

[0008] First, physical layer channel state information (CSI) is collected at each sampling position to collect unlabeled RF data. Then, a large number of multi-antenna and multi-subcarrier signal amplitudes are used to design data frames to generate RF images.

[0009] Self-supervised pre-trained models:

[0010] Based on the self-supervised pre-training paradigm, the Momentum Contrastive Learning algorithm is used to design a twin convolutional network to extract signal features. A flexible dynamic channel dictionary search mechanism is proposed to learn to represent the wireless propagation characteristics in complex dynamic environments. Momentum is used to update the network model parameters, thereby significantly improving the quality of signal feature representation.

[0011] Downstream positioning tasks:

[0012] In specific scenarios, the encoder network in the pre-trained model is directly migrated, and the location prediction network is trained using small sample RF fingerprints to obtain accurate location estimation.

[0013] As a preferred method, the generation process of the radio frequency image is as follows: Figure 2 As shown, the multi-subcarrier CSI amplitude vectors of a single antenna in 10 CSI data packets are spliced ​​row by row to form a RF sub-image, and then the three groups of RF sub-images obtained by integrating three antennas in a commercial WiFi device are spliced ​​row by row to finally obtain a RF image.

[0014] As a preferred embodiment, the present invention adopts a self-supervised pre-training paradigm to learn the intrinsic signal feature representation of radio frequency fingerprint data, proposes a flexible dynamic channel dictionary search mechanism to optimize the network model, and significantly improves the positioning accuracy, robustness and universality of the indoor fingerprint positioning system;

[0015] The feature encoder of the self-supervised pre-training model adopts a convolutional neural network architecture. It uses a large-scale unlabeled RF image to design the feature encoder f and extract its feature vector:

[0016] h=f(x)=CNN(x)

[0017] Where CNN stands for convolutional neural network, x is the input RF image, and h represents the feature representation vector output by the feature encoder f.

[0018] As a preferred method, the mapping network of the self-supervised pre-training model uses a multi-layer perceptron to design a feature mapping network g, which maps the feature vector to the feature representation space, thereby improving the quality of signal feature representation:

[0019] z=g(h)=g(f(X))

[0020] According to the above steps, design the feature mapping network g q , g k , and design a contrastive learning training strategy, where f q and g q represents the feature encoder and mapping network, f k , g k Represents the momentum encoder and momentum mapping networks.

[0021] As a preferred method, based on the momentum contrastive learning framework of the self-supervised pre-training model, this method designs a dynamic channel dictionary search mechanism to learn to represent the intrinsic features of RF data. For any generated RF image, the data from the same position constitutes a positive sample pair, and the data in the channel dictionary constitutes a negative sample pair, whose size is K. For any RF image x q , design feature mapping network g q , g k Map it to the signal representation space:

[0022] z q =g q (h q ) = g q (f q (x q ))

[0023] z k =g k (h k ) = g k (f k (x q ))

[0024] Embedding vector z k Stored in a channel dictionary of size K {z1,z2,z3,…,z K}, and with the model training process, new embedding vectors are continuously stored and old embedding vectors are removed, and the dictionary is dynamically updated, which effectively maintains the consistency of the channel dictionary and thus improves the quality of the signal feature representation vector;

[0025] This method uses contrast loss to optimize the self-supervised pre-training model, which is expressed as follows:

[0026]

[0027]

[0028] Where L is the loss function of the self-supervised pre-training model, τ is the temperature parameter that controls the sensitivity of the similarity measure between different encoding embeddings, and z0 is x q After f q , g qMapping obtains feature embedding, z k represents the embedding vector in the channel dictionary, and K represents the size of the RF dictionary

[0029] As a preferred method, the momentum update mechanism of the self-supervised pre-training model adopts the contrast loss function L to perform self-supervised learning to optimize the network model, and updates f through back propagation. q , g q The network parameters of f are updated through the momentum mechanism k , g k The network parameters of the momentum mechanism are as follows:

[0030] This method uses a momentum mechanism to update model parameters based on a sliding average:

[0031] θ k ←mθ k +(1_m)θ q

[0032] where m∈[0,1) represents the weight updated by the momentum mechanism, θ q ,θ k They are the feature encoder f q and momentum encoder f k The learnable parameters, θ q The model parameters are optimized by back propagation, and θ k It is updated through the momentum mechanism;

[0033] At the same time, the parameters of the mapping network are updated in the same mechanism:

[0034]

[0035] where φ q and φ k Represent the mapping network g q and g k The learnable parameters, φ q will be updated via backpropagation, and Will be updated through the momentum mechanism.

[0036] As a preferred method, for the specific scenario of downstream positioning tasks, this method directly migrates the feature encoder to obtain a high-quality, highly robust and highly generalized feature representation vector:

[0037] h=f q (x)

[0038] This method further utilizes the high-quality signal feature representation vector generated by the pre-trained model to perform position prediction;

[0039] Specifically, this method first uses a small sample of radio frequency fingerprints with a small number of markers Train the downstream position predictor:

[0040] p i =softmax(MLP(f q (x i )))

[0041]

[0042] Where MLP stands for Multi-layer Perceptron, L p is the loss function of the softmax classifier, L ce represents the cross entropy loss function, y i for y i The corresponding one-hot vector;

[0043] In the positioning stage, the RF image x is generated by the channel state information provided by the user, and the pre-trained feature encoder is used to obtain the high-quality feature vector h = f q (x), then, the posterior probability of the downstream position predictor is calculated:

[0044]

[0045] Then calculate the position estimate as

[0046]

[0047] Where L y is the position coordinate of the reference point, and y is its corresponding position label.

[0048] Beneficial effects of the present invention:

[0049] The present invention proposes a "radio fingerprint positioning method based on a self-supervised pre-training model", which adopts a self-supervised pre-training paradigm to learn the intrinsic feature representation of a large amount of unlabeled radio frequency data, proposes a flexible dynamic channel dictionary search mechanism based on momentum contrast learning, and constructs a wireless pre-training model, which has the following advantages:

[0050] The present invention performs well in deep fingerprint positioning tasks in complex dynamic environments. It trains feature encoders with a large amount of unlabeled radio frequency data, characterizes signal propagation characteristics based on self-supervised comparative learning, and obtains signal features with high quality, strong discrimination, and universal applicability in all scenarios. This feature significantly improves the positioning accuracy, robustness, and universality of indoor fingerprint positioning systems.

[0051] The present invention realizes high-precision deep fingerprint positioning technology based on small sample radio frequency fingerprint, which significantly reduces the high cost of building a fingerprint database. At the same time, the method has strong robustness and generalization, and there is no need to regularly update and maintain the radio frequency fingerprint database, thereby greatly reducing the deployment and maintenance costs of the positioning system. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 What is shown is a model block diagram of the radio frequency fingerprint positioning method of the present invention;

[0053] Figure 2 What is shown is the generation process of the radio frequency image of the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further described below in conjunction with the embodiments.

[0055] See also Figure 1-2 The present invention provides an embodiment: a radio frequency fingerprint positioning method based on a self-supervised pre-training model, the steps are as follows:

[0056] RF image generation:

[0057] First, physical layer channel state information (CSI) is collected at each sampling position to collect unlabeled RF data. Then, a large number of multi-antenna and multi-subcarrier signal amplitudes are used to design data frames to generate RF images.

[0058] Self-supervised pre-trained models:

[0059] Based on the self-supervised pre-training paradigm, the Momentum Contrastive Learning algorithm is used to design a twin convolutional network to extract signal features. A flexible dynamic channel dictionary search mechanism is proposed to learn to represent the wireless propagation characteristics in complex dynamic environments. Momentum is used to update the network model parameters, thereby significantly improving the quality of signal feature representation.

[0060] Downstream positioning tasks:

[0061] In specific scenarios, the encoder network in the pre-trained model is directly migrated, and the location prediction network is trained using small sample RF fingerprints to obtain accurate location estimation.

[0062] The generation process of RF image is as follows Figure 2 As shown, the multi-subcarrier CSI amplitude vectors of a single antenna in 10 CSI data packets are spliced ​​row by row to form a RF sub-image, and then the three groups of RF sub-images obtained by integrating three antennas in a commercial WiFi device are spliced ​​row by row to finally obtain a RF image.

[0063] The present invention adopts a self-supervised pre-training paradigm to learn the intrinsic signal feature representation of radio frequency fingerprint data, proposes a flexible dynamic channel dictionary search mechanism to optimize the network model, and significantly improves the positioning accuracy, robustness and universality of the indoor fingerprint positioning system;

[0064] The feature encoder of the self-supervised pre-training model adopts a convolutional neural network architecture. It uses a large-scale unlabeled RF image to design the feature encoder f and extract its feature vector:

[0065] h=f(x)=CNN(x)

[0066] Where CNN stands for convolutional neural network, x is the input RF image, and h represents the feature representation vector output by the feature encoder f.

[0067] The mapping network of the self-supervised pre-training model uses a multi-layer perceptron to design a feature mapping network g, which maps the feature vector to the feature representation space, thereby improving the quality of signal feature representation:

[0068] z=g(h)=g(f(X))

[0069] According to the above steps, design the feature mapping network g q , g k , and design a contrastive learning training strategy, where f q and g q represents the feature encoder and mapping network, f k , g k Represents the momentum encoder and momentum mapping networks.

[0070] Based on the momentum contrastive learning framework of the self-supervised pre-training model, this method designs a dynamic channel dictionary search mechanism to learn to represent the intrinsic features of RF data. For any generated RF image, the data from the same position constitutes a positive sample pair, and the data in the channel dictionary constitutes a negative sample pair, whose size is K. For any RF image x q , design feature mapping network g q , g k Map it to the signal representation space:

[0071] z q =g q (h q ) = g q (f q (x q ))

[0072] z k =g k (h k ) = g k (f k (x q ))

[0073] Embedding vector z k Stored in a channel dictionary of size K {z1,z2,z3,…,z K}, and with the model training process, new embedding vectors are continuously stored and old embedding vectors are removed, and the dictionary is dynamically updated, which effectively maintains the consistency of the channel dictionary and thus improves the quality of the signal feature representation vector;

[0074] This method uses contrast loss to optimize the self-supervised pre-training model, which is expressed as follows:

[0075]

[0076]

[0077] Where L is the loss function of the self-supervised pre-training model, τ is the temperature parameter that controls the sensitivity of the similarity measure between different encoding embeddings, and z0 is x q After f q , g q Mapping obtains feature embedding, z k represents the embedding vector in the channel dictionary, and K represents the size of the RF dictionary.

[0078] The momentum update mechanism of the self-supervised pre-training model uses the contrast loss function L to perform self-supervised learning to optimize the network model and update f through back propagation. q , g q The network parameters of f are updated through the momentum mechanism k , g k The network parameters of the momentum mechanism are as follows:

[0079] This method uses a momentum mechanism to update model parameters based on a sliding average:

[0080] θ k ←mθ k +(1_m)θ q

[0081] where m∈[0,1) represents the weight updated by the momentum mechanism, θ q ,θ k They are the feature encoder f q and momentum encoder f k The learnable parameters, θ q The model parameters are optimized by back propagation, and θ k It is updated through the momentum mechanism;

[0082] At the same time, the parameters of the mapping network are updated in the same mechanism:

[0083]

[0084] where φ q and φ k Represent the mapping network g qand g k The learnable parameters, φ q will be updated via backpropagation, and Will be updated through the momentum mechanism.

[0085] For the specific scenario of downstream positioning tasks, this method directly migrates the feature encoder to obtain a high-quality, highly robust and highly generalized feature representation vector:

[0086] h=f q (x)

[0087] This method further utilizes the high-quality signal feature representation vector generated by the pre-trained model to perform position prediction;

[0088] Specifically, this method first uses a small sample of radio frequency fingerprints with a small number of markers Train the downstream position predictor:

[0089] p i =softmax(MLP(f q (x i )))

[0090]

[0091] Where MLP stands for Multi-layer Perceptron, L p is the loss function of the softmax classifier, L ce represents the cross entropy loss function, y i for y i The corresponding one-hot vector;

[0092] In the positioning stage, the RF image x is generated by the channel state information provided by the user, and the pre-trained feature encoder is used to obtain the high-quality feature vector h = f q (x), then, the posterior probability of the downstream position predictor is calculated:

[0093]

[0094] Then calculate the position estimate as

[0095]

[0096] Where L y is the position coordinate of the reference point, and y is its corresponding position label.

[0097] Through Example 1, the present invention adopts a self-supervised pre-training paradigm to learn the intrinsic feature representation of radio frequency fingerprint data, and proposes a wireless pre-training model, which aims to characterize the signal propagation characteristics in complex environments, improve the quality of radio frequency data signal feature representation, and provide a theoretical basis for the development of high-precision, high-robustness and strong generalization indoor positioning systems;

[0098] A feature-rich dynamic channel dictionary is constructed and the momentum encoder f is updated through the momentum mechanism k and mapping network g k , the embedding vector z of the two outputs k Continuously update the channel dictionary. This update strategy not only ensures the consistency of dictionary lookup, but also ensures that the dictionary is always valid, which helps to achieve high-quality fingerprint feature representation and high-precision positioning.

[0099] The present invention also proposes a RF feature representation technology based on a self-supervised pre-training model, which generates RF images from a large amount of unlabeled RF data from different scenarios, and uses self-supervised comparative learning to characterize the signal propagation law in complex dynamic environments, thereby obtaining high-quality, highly robust, and highly generalized fingerprint feature representation. For specific positioning scenarios, the present invention designs a location prediction network, which only uses a small sample of RF fingerprints for location prediction network training to accurately estimate the user's location information.

Claims

1. The radio frequency fingerprint positioning method based on the self-supervised pre-training model is characterized by: Here are the steps: RF image generation: First, physical layer channel state information (CSI) is collected at each sampling position to collect unlabeled RF data. Then, a large number of multi-antenna and multi-subcarrier signal amplitudes are used to design data frames to generate RF images. Self-supervised pre-trained models: Based on the self-supervised pre-training paradigm, the Momentum Contrastive Learning algorithm is used to design a twin convolutional network to extract signal features. A flexible dynamic channel dictionary search mechanism is proposed to learn to represent the wireless propagation characteristics in complex dynamic environments. Momentum is used to update the network model parameters, thereby significantly improving the quality of signal feature representation. Downstream positioning tasks: In specific scenarios, the encoder network in the pre-trained model is directly migrated, and the location prediction network is trained using small sample RF fingerprints to obtain accurate location estimation.

2. The radio frequency fingerprint positioning method based on the self-supervised pre-training model according to claim 1 is characterized in that: The multi-subcarrier CSI amplitude vectors of a single antenna in 10 CSI data packets are spliced ​​row by row to form an RF sub-image. Then, the three groups of RF sub-images obtained by integrating three antennas in a commercial WiFi device are spliced ​​row by row to finally obtain an RF image.

3. The radio frequency fingerprint positioning method based on the self-supervised pre-training model according to claim 1 is characterized in that: The present invention adopts a self-supervised pre-training paradigm to learn the intrinsic signal feature representation of radio frequency fingerprint data, proposes a flexible dynamic channel dictionary search mechanism to optimize the network model, and significantly improves the positioning accuracy, robustness and universality of the indoor fingerprint positioning system; The feature encoder of the self-supervised pre-training model adopts a convolutional neural network architecture. It uses a large-scale unlabeled RF image to design the feature encoder f and extract its feature vector: h=f(x)=CNN(x) Where CNN stands for convolutional neural network, x is the input RF image, and h represents the feature representation vector output by the feature encoder f.

4. The radio frequency fingerprint positioning method based on the self-supervised pre-training model according to claim 3 is characterized in that: The mapping network of the self-supervised pre-training model uses a multi-layer perceptron to design a feature mapping network g, which maps the feature vector to the feature representation space, thereby improving the quality of signal feature representation: z=g(h)=g(f(x)) According to the above steps, design the feature mapping network g q , g k , and design a contrastive learning training strategy, where f q and g q represents the feature encoder and mapping network, f k , g k Represents the momentum encoder and momentum mapping networks.

5. The radio frequency fingerprint positioning method based on the self-supervised pre-training model according to claim 3 is characterized in that: Based on the momentum contrastive learning framework of the self-supervised pre-training model, this method designs a dynamic channel dictionary search mechanism to learn to represent the intrinsic features of RF data. For any generated RF image, the data from the same position constitutes a positive sample pair, and the data in the channel dictionary constitutes a negative sample pair, whose size is K. For any RF image x q , design feature mapping network g q , g k Map it to the signal representation space: z q =g q (h q )=g q (f q (x q )) z k =g k (h k )=g k (f k (x q )) Embedding vector z k Stored in a channel dictionary of size K {z1,z2,z3,…,z K }, and with the model training process, new embedding vectors are continuously stored and old embedding vectors are removed, and the dictionary is dynamically updated, which effectively maintains the consistency of the channel dictionary and thus improves the quality of the signal feature representation vector; This method uses contrast loss to optimize the self-supervised pre-training model, which is expressed as follows: Where L is the loss function of the self-supervised pre-training model, τ is the temperature parameter that controls the sensitivity of the similarity measure between different encoding embeddings, and z0 is x q After f q , g q The mapping obtains feature embedding, zk represents the embedding vector in the channel dictionary, and K represents the size of the RF dictionary.

6. The radio frequency fingerprint positioning method based on the self-supervised pre-training model according to claim 3 is characterized in that: The momentum update mechanism of the self-supervised pre-training model uses the contrast loss function L to perform self-supervised learning to optimize the network model and update f through back propagation. q , g q The network parameters of f are updated through the momentum mechanism k , g k The network parameters of the momentum mechanism are as follows: This method uses a momentum mechanism to update model parameters based on a sliding average: i k ←mθ k +(1_m)θ q where m∈[0,1) represents the weight updated by the momentum mechanism, θ q ,θ k They are the feature encoder f q and momentum encoder f k The learnable parameters of θ q The model parameters are optimized by back propagation, and θ k It is updated through the momentum mechanism; At the same time, the parameters of the mapping network are updated in the same mechanism: where φ q and φ k Represent the mapping network g q and g k The learnable parameters, φ q will be updated via backpropagation, and Will be updated through the momentum mechanism.

7. The radio frequency fingerprint positioning method based on the self-supervised pre-training model according to claim 1 is characterized in that: For the specific scenario of downstream positioning tasks, this method directly migrates the feature encoder to obtain a high-quality, highly robust and highly generalized feature representation vector: h=f q (x) This method then uses the high-quality signal feature representation vector generated by the pre-trained model to predict the location; specifically, this method first uses a small number of labeled small sample RF fingerprints Train the downstream position predictor: p i =softmax(MLP(f q (x i ))) Where MLP stands for Multi-layer Perceptron, L p is the loss function of the softmax classifier, L ce represents the cross entropy loss function, y i for y i The corresponding one-hot vector; In the positioning stage, the RF image x is generated by the channel state information provided by the user, and the pre-trained feature encoder is used to obtain the high-quality feature vector h = f q (x), then, the posterior probability of the downstream position predictor is calculated: Pr(y|h)=softmax(MLP(h)) Then calculate the position estimate as Where L y is the position coordinate of the reference point, and y is its corresponding position label.

Images