Radio frequency fingerprint-based positioning model training method, positioning model and positioning method

Abstract

The application discloses a positioning model construction method, a positioning model and a positioning method based on a radio frequency fingerprint. The model construction method comprises the following steps: constructing a radio frequency image training set according to a plurality of first original radio frequency fingerprint information; determining a first loss function according to the radio frequency image training set, so as to construct and train a feature encoder; determining corresponding first-view radio frequency images, second-view radio frequency images and position label information according to a plurality of second original radio frequency fingerprint information; and determining a second loss function according to the corresponding first-view radio frequency images, the second-view radio frequency images and the position label information, so as to construct and train a position predictor. The application constructs a training set based on two views of amplitudes and phases, further trains the positioning model by comparison and self-supervision, so that the model learns the potential invariance representation in the radio frequency fingerprint data, better adapts to the dynamic change of the environment, essentially improves the universality of the signal feature representation, and improves the generalization ability of the positioning model.

Description

Technical Field

[0001] This application relates to a method for constructing a positioning model based on radio frequency fingerprints, a positioning model based on radio frequency fingerprints, and a positioning method based on radio frequency fingerprints, belonging to the field of wireless fingerprint positioning technology. Background Technology

[0002] Deep radio frequency fingerprinting technology can train a deep network to extract radio frequency fingerprint features by collecting a large amount of wireless fingerprint data with labeled location information. By utilizing the learning and representation capabilities of the deep network, it can establish an intrinsic mapping relationship between signal features and location, and estimate location information by using feature matching, thereby improving the positioning accuracy in specific application scenarios.

[0003] However, in localization methods using deep radio frequency fingerprinting, the process of extracting signal features is highly dependent on the environment. Therefore, localization methods based on deep transfer learning can be further adopted to reduce the differences in feature distribution caused by environmental changes, thereby improving localization stability. However, localization methods based on deep transfer learning are mostly only applicable to a single specific localization scenario, and their generalization ability is poor when deployed on a large scale. Summary of the Invention

[0004] This application discloses a method for constructing a localization model based on radio frequency fingerprints, a localization model based on radio frequency fingerprints, and a localization method based on radio frequency fingerprints.

[0005] The localization model construction method based on radio frequency fingerprinting in this application constructs a localization model including a feature encoder and a location predictor. The method includes: A radio frequency image training set is constructed based on multiple first original radio frequency fingerprint information. The radio frequency image training set includes a first-view radio frequency image and a second-view radio frequency image. The first view is an amplitude view or a phase view, and the second view is an amplitude view or a phase view. The first view and the second view are different. Each first original radio frequency fingerprint information corresponds to a first-view radio frequency image and a second-view radio frequency image. Based on contrastive self-supervised learning, a first loss function is determined according to the radio frequency image training set in order to construct and train the feature encoder; Based on multiple second original radio frequency fingerprint information, determine the first-view radio frequency image, the second-view radio frequency image, and the location tag information corresponding to the second original radio frequency fingerprint information; Based on the first-view radio frequency image, the second-view radio frequency image, and the location label information corresponding to the second original radio frequency fingerprint information, a second loss function is determined to construct and train the location predictor.

[0006] In some implementations, constructing a radio frequency image training set based on multiple first original radio frequency fingerprint information includes: Based on the first original radio frequency fingerprint information, obtain the first view matrix and the second view matrix; Generate a first-view radio frequency image based on the first-view matrix; Based on the second viewpoint matrix, a second viewpoint radio frequency image is generated to construct the radio frequency image training set.

[0007] In some implementations, determining a first loss function based on the radio frequency image training set to construct and train the feature encoder includes: Based on the radio frequency image training set, the first viewpoint feature embedding and the second viewpoint feature embedding are determined; A first initial loss function is determined based on the radio frequency image training set, the first viewpoint feature embedding, and the second viewpoint feature embedding; Based on the non-backpropagation of gradients in a neural network, the first initial loss function is optimized and determined to construct and train the feature encoder.

[0008] In some implementations, determining the first viewpoint feature embedding and the second viewpoint feature embedding based on the radio frequency image training set includes: Based on the convolutional neural network, the corresponding first-view feature vector is determined according to the first-view radio frequency image; Based on a convolutional neural network, the corresponding second-view feature vector is determined according to the second-view radio frequency image; Based on a multilayer perceptron network, the first viewpoint feature embedding is determined according to the first viewpoint feature vector; Based on a multilayer perceptron network, the second-view feature embedding is determined according to the second-view feature vector.

[0009] In some implementations, determining the first initial loss function based on the first viewpoint feature embedding and the second viewpoint feature embedding includes: Based on the prediction network, the first viewpoint feature prediction is determined according to the first viewpoint feature embedding; Based on the prediction network, the second perspective feature prediction is determined according to the second perspective feature embedding; The first initial loss function is determined based on the similarity between the first viewpoint feature embedding and the second viewpoint feature prediction, and the similarity between the first viewpoint feature prediction and the second viewpoint feature embedding.

[0010] In some implementations, optimizing the first initial loss function based on neural network gradient non-backpropagation and determining the first loss function includes: Based on the first initial loss function, gradient non-backpropagation processing is performed on the first viewpoint feature embedding and the second viewpoint feature embedding to determine the first loss function.

[0011] In some implementations, determining a second loss function based on the first-view radio frequency image, the second-view radio frequency image, and the location tag information corresponding to the second original radio frequency fingerprint information to construct and train the location predictor includes: Based on a convolutional neural network, a first feature vector is determined according to the first-view radio frequency image corresponding to the second original radio frequency fingerprint information; Based on a convolutional neural network, a second feature vector is determined according to the second-view radio frequency image corresponding to the second original radio frequency fingerprint information. Based on cross-entropy loss, the second loss function is determined according to the first feature vector, the second feature vector, and the location label information.

[0012] The radio frequency fingerprint-based localization model in this application is constructed according to the radio frequency fingerprint-based localization model construction method in the above embodiments.

[0013] The radio frequency fingerprint-based localization method in this application is implemented based on a localization model constructed according to the radio frequency fingerprint-based localization model construction method in the above embodiments. The method includes: Based on the radio frequency fingerprint information of the object to be located, generate a first-view radio frequency image and a second-view radio frequency image data corresponding to the object to be located. Based on the feature encoder, a corresponding first-view feature vector is determined according to the first-view radio frequency image, and a corresponding second-view feature vector is determined according to the second-view radio frequency image data. Based on the location predictor, the estimated location of the object to be located is determined according to the first view feature vector, the second view feature vector, and the location label information saved when constructing and training the location predictor.

[0014] In some implementations, determining the estimated location of the object to be located based on the location predictor, according to the first view feature vector, the second view feature vector, and the location label information, includes: Based on the location predictor, the output posterior probability of the location predictor is determined according to the first view feature vector and the second view feature vector. The estimated location of the object to be located is determined based on the posterior probability and the location label information saved when constructing and training the location predictor.

[0015] The beneficial effects of this application are: This application constructs a training set based on two perspectives, amplitude and phase, and then further trains the localization model through comparative self-supervision, so that the model learns the potential invariant representations in radio frequency fingerprint data and better adapts to the dynamic changes in the environment, thereby fundamentally improving the universality of signal feature representation and enhancing the generalization ability of the localization model. Attached Figure Description

[0016] Figure 1 This is one of the flowcharts illustrating the location model construction method based on radio frequency fingerprinting in the embodiments of this application; Figure 2 This is the second flowchart illustrating the method for constructing a localization model based on radio frequency fingerprinting in the embodiments of this application; Figure 3 This is the third flowchart illustrating the method for constructing a localization model based on radio frequency fingerprinting in the embodiments of this application; Figure 4 This is the fourth flowchart illustrating the method for constructing a localization model based on radio frequency fingerprinting in the embodiments of this application; Figure 5 This is the fifth flowchart illustrating the method for constructing a localization model based on radio frequency fingerprinting in the embodiments of this application; Figure 6 This is the sixth flowchart illustrating the method for constructing a localization model based on radio frequency fingerprinting in the embodiments of this application; Figure 7 This is the seventh flowchart of the method for constructing a localization model based on radio frequency fingerprints in the embodiments of this application; Figure 8 This is the eighth flowchart of the method for constructing a localization model based on radio frequency fingerprints in the embodiments of this application. Detailed Implementation

[0017] Please see Figure 1 The localization model construction method based on radio frequency fingerprinting in this application constructs a localization model including a feature encoder and a location predictor. The method includes the following steps: Step 01: Construct an RF image training set based on multiple initial RF fingerprint information. The radio frequency image training set includes a first-view radio frequency image and a second-view radio frequency image. The first view is an amplitude view or a phase view, and the second view is an amplitude view or a phase view. The first view and the second view are different. Each first original radio frequency fingerprint information corresponds to a first-view radio frequency image and a second-view radio frequency image.

[0018] Specifically, the method for constructing a localization model based on radio frequency fingerprints in this application mainly consists of two parts. The first part is to train the feature encoder, which aims to train the feature encoder to learn general features that can reflect the essence of the location. This allows the localization model to extract the aforementioned general features based on the radio frequency fingerprint of the target to be located when performing localization. The second part is to train the location predictor, which aims to train the location predictor to use radio frequency fingerprints in an indoor environment using a small number of radio frequency fingerprint samples with location information labels. This enables the location predictor to achieve location estimation through feature matching. Thus, the localization capability of the localization model is realized through training. Finally, a complete localization model based on radio frequency fingerprints is constructed through the above two parts of training.

[0019] First, training the feature encoder requires a pre-constructed training dataset. Therefore, in some examples, multiple sets of first raw radio frequency (RF) fingerprint information can be pre-collected. For example, the first raw RF fingerprint information can be Channel State Information (CSI) fingerprints for multiple antennas and subcarriers, hereinafter referred to as first raw CSI fingerprint information. Based on each set of first raw CSI fingerprint information, an amplitude-view RF image and a phase-view RF image can be generated. The amplitude-view RF images and phase-view RF images from the same set of first raw CSI fingerprint information are positive sample pairs. The combination of RF images corresponding to all first raw CSI fingerprint information forms the RF image training set for training the feature encoder.

[0020] For example, please refer to Figure 2 In some implementations, step 01 includes: Step 011: Based on the first original radio frequency fingerprint information, obtain the first view matrix and the second view matrix; Step 012: Generate a first-view radio frequency image based on the first-view matrix; Step 013: Generate second-view radio frequency images based on the second-view matrix to construct a radio frequency image training set.

[0021] Specifically, in some examples, the CSI fingerprint information used to generate the radio frequency image training set is generally in the form shown in Equation 1.

[0022] ………… Formula 1 in H The fingerprint matrix corresponding to the first received raw CSI fingerprint information. , N For multiple subcarriers, MThe number of antennas. A This is a matrix of amplitude measurement values. , where e is the base of the natural logarithm. j The imaginary unit, This is a matrix of phase measurement values. Amplitude measurement matrix A and phase measurement matrix This corresponds to the first viewpoint matrix and the second viewpoint matrix. For ease of description, in the following embodiments, the amplitude viewpoint corresponds to the first viewpoint and the phase viewpoint corresponds to the second viewpoint. In practical applications, the above correspondence can be interchanged, and this application will not repeat the explanation.

[0023] Then, using image transformation algorithms from current related technologies, the amplitude measurement matrix is ​​used. A and phase measurement matrix Based on this, image transformation is performed to obtain the amplitude measurement matrix. A Corresponding amplitude-view radio frequency image and phase measurement matrix The corresponding phase-view radio frequency (RF) images. In the above case, each set of first original CSI fingerprint information corresponds to a pair of amplitude-view RF images and phase-view RF images. Therefore, all the amplitude-view and phase-view RF images corresponding to the first original CSI fingerprint information together constitute the RF image training set used to train the feature encoder. The characteristic of the RF image training set is that it only includes the image features of amplitude-view and phase-view, but does not include the image or the location label information of the object corresponding to the first original CSI fingerprint information.

[0024] Please refer to further information. Figure 1 The method for constructing a localization model based on radio frequency fingerprints in the embodiments of this application further includes: Step 02: Based on contrastive self-supervised learning, determine the first loss function according to the radio frequency image training set to construct and train the feature encoder.

[0025] Specifically, based on the above implementation, when a radio frequency image training set is determined, the feature encoder is further trained using the images in the set as the data basis and a contrastive self-supervised learning mechanism. By maximizing the similarity of the dual-view feature embeddings, the feature encoder learns a consistent representation that is invariant to the viewpoint, thereby enabling the feature encoder to learn general features that can reflect the essence of the location.

[0026] Please see Figure 3 In some implementations, step 02 specifically includes: Step 021: Determine the first-view feature embedding and the second-view feature embedding based on the radio frequency image training set; Step 022: Determine the first initial loss function based on the radio frequency image training set, the first viewpoint feature embedding, and the second viewpoint feature embedding; Step 023: Based on the non-backpropagation of neural network gradients, optimize the first initial loss function and determine the first loss function to construct and train the feature encoder.

[0027] Specifically, for amplitude-view or phase-view radio frequency images from the same first original CSI fingerprint information in the radio frequency image training set. x First, based on this, the corresponding amplitude feature embedding and phase feature embedding are obtained. In this way, for any first original radio frequency fingerprint information, a pair of feature embeddings in the amplitude view and phase view will be obtained.

[0028] Regarding the method for obtaining feature embeddings, in some implementations, please refer to [link to relevant documentation]. Figure 4 Step 021 specifically includes: Step 0211: Based on the convolutional neural network, determine the corresponding first-view feature vector according to the first-view radio frequency image; Step 0212: Based on the convolutional neural network, determine the corresponding second-view feature vector according to the second-view radio frequency image; Step 0213: Based on the multilayer perceptron network, determine the first viewpoint feature embedding according to the first viewpoint feature vector; Step 0214: Based on the multilayer perceptron network, determine the second viewpoint feature embedding according to the second viewpoint feature vector.

[0029] It should be noted that the execution order between steps 0211 and 0212, and between steps 0213 and 0214, is arbitrary. Figure 4 This is for illustrative purposes only and should not be interpreted as a limitation on the order of execution.

[0030] Specifically, the feature embedding process generally consists of two steps. First, based on a convolutional neural network, arbitrary amplitude-view or phase-view radio frequency images are used in the radio frequency image training set. x Perform feature extraction, as shown in Formula 2, to obtain the corresponding amplitude feature vector or phase feature vector.

[0031] ………… Formula 2 in For image The corresponding amplitude-view feature vector or phase-view radio frequency image, Represents the feature encoder. This represents a convolutional neural network.

[0032] It is important to note that after the localization model is trained, the feature vector output by the feature encoder based on the given CSI fingerprint is also based on the output of the aforementioned convolutional neural network. The training process of the feature encoder actually optimizes the network parameters of the aforementioned convolutional neural network.

[0033] Next, based on the obtained amplitude feature vector or phase feature vector, a nonlinear transformation is performed on a multilayer perceptron (MLP) designed with a two-layer fully connected neural network to obtain the amplitude feature embedding or phase feature embedding corresponding to the amplitude feature vector or phase feature vector, as shown in Formula 3.

[0034] ………… Formula 3 in For image The corresponding amplitude-view feature embedding or phase-view feature embedding, This represents a multilayer perceptron. Next, using... This represents the amplitude-view feature embedding corresponding to the amplitude-view radio frequency image. The phase view feature embedding corresponding to the phase view radio frequency image is further explained.

[0035] In some implementations, please refer to Figure 5 Step 022 further includes: Step 0221: Based on the prediction network, determine the first-view feature prediction according to the first-view feature embedding; Step 0222: Based on the prediction network, determine the second-view feature prediction according to the second-view feature embedding; Step 0223: Determine the first initial loss function based on the similarity between the first-view feature embedding and the second-view feature prediction, and the similarity between the first-view feature prediction and the second-view feature embedding.

[0036] It should be noted that the execution order between steps 0221 and 0222 is arbitrary. Figure 5 This is for illustrative purposes only and should not be interpreted as a limitation on the order of execution.

[0037] Specifically, based on the above implementation method, the amplitude-viewpoint feature embedding has been determined. and phase view feature embedding In this case, the next step is to embed features from the amplitude perspective. and phase view feature embedding Based on this, we construct the first loss function for training the feature encoder.

[0038] For example, firstly, based on the amplitude feature vector or phase feature vector in the above embodiments, the amplitude feature vector or phase feature vector is predicted and calculated using two preset prediction networks composed of fully connected layers, to obtain the corresponding amplitude feature prediction and phase feature prediction. This is specifically shown in Formula 4.

[0039] ………… Formula 4 in The prediction network described above has one branch corresponding to each of the amplitude and phase perspectives. For amplitude perspective feature prediction, Predicting phase view features.

[0040] So, after obtaining the amplitude perspective feature prediction Phase-view feature prediction Amplitude-view feature embedding and phase view feature embedding In this case, the amplitude perspective feature prediction is then calculated using a contrastive self-supervised approach. Embedded with phase view features Similarity between them, and phase view feature prediction Embedding with amplitude and viewpoint features The similarity between them. For example, the above similarity can generally be taken as negative cosine similarity, as shown in Formula 5.

[0041] ………… Formula 5 in This indicates L2 normalization.

[0042] Furthermore, based on the two sets of symmetrical similarities mentioned above, the sum of the similarities is calculated, which yields the first initial loss function, as shown in Formula 6.

[0043] ………… Formula 6 in This represents the first initial loss function described above. Indicates the first i The initial loss corresponding to the first set of original CSI fingerprint information. K The total number of original CSI fingerprints. For the first i Prediction of amplitude-view feature corresponding to the first original CSI fingerprint information in group 1. For the first i Prediction of phase view features corresponding to the first original CSI fingerprint information in group 1 For the first i Phase view feature embedding corresponding to the first original CSI fingerprint information in group 1 For the first i The amplitude-view feature embedding is based on the first original CSI fingerprint information of the group.

[0044] In this way, by calculating the similarity based on the data obtained from the two prediction networks and using it as a way to optimize the loss, the network parameters of the two prediction networks can be optimized simultaneously, thereby constructing a feature encoder that meets the requirements.

[0045] Next, the process of obtaining the first loss function of the feature encoder as a whole based on the first initial loss function, in some embodiments, step 023 specifically includes: Based on the first initial loss function, gradient non-backpropagation processing is performed on the first viewpoint feature prediction and the second viewpoint feature prediction to determine the first loss function.

[0046] Specifically, under the aforementioned first initial loss function, when the prediction network backpropagates, the loss will simultaneously update all parameters of both the amplitude branch and the phase branch. At this point, in order to... Maximum, may make proactively To move closer in direction, similarly, in order to Maximum, may make proactively They moved in the same direction, so that... and They will become increasingly similar or even completely convergent, indicating that the model has "collapsed".

[0047] Therefore, for example, based on the first initial loss function described above, one prediction network can be fixed and the other prediction network optimized separately by using gradient-free backpropagation. This emphasizes the cross-perspective prediction ability learning during the learning process and can, to some extent, avoid model collapse. For example, regarding the first initial loss function described above, for... and Gradient backpropagation is performed to obtain the optimized first loss function, as shown in Equation 7.

[0048] ………… Formula 7 in For the first loss function mentioned above, Indicates the first i The loss corresponding to the first original CSI fingerprint information in group , This indicates that the gradient is not backpropagated.

[0049] by To illustrate this with an example, under the condition of gradient non-backpropagation, Backpropagation of the gradient is blocked, therefore The corresponding network parameters are fixed, which is equivalent to a fixed reference standard. In this case, the loss can only be achieved by updating... The corresponding network parameters of the prediction network are used to let proactively To move closer. In this way, and It will not converge due to bidirectional gradient updates, thus forcing the feature encoder to capture universal feature representations during training, ultimately improving the generalization ability of the localization model.

[0050] Please continue reading. Figure 1 The location model construction method based on radio frequency fingerprinting in the embodiments of this application further includes: Step 03: Based on multiple second original radio frequency fingerprint information, determine the first-view radio frequency image, the second-view radio frequency image, and the location tag information corresponding to the second original radio frequency fingerprint information.

[0051] Specifically, based on the completion of the feature encoder training in the above embodiments, the training of the position predictor first generates corresponding amplitude-view radio frequency images and phase-view radio frequency images based on the second original radio frequency fingerprint information including position label information. The second original radio frequency fingerprint information is generally also CSI fingerprint information, hereinafter referred to as the second original CSI fingerprint information. Its quantity is generally much smaller than the first original CSI fingerprint information, but unlike the first original CSI fingerprint information, it includes the position information label of the object corresponding to the information. The specific method for generating the radio frequency images can refer to the process of generating corresponding radio frequency images based on the first original CSI fingerprint information in the above embodiments, and will not be repeated here.

[0052] Please continue reading. Figure 1 The location model construction method based on radio frequency fingerprinting in the embodiments of this application further includes: Step 04: Based on the first-view radio frequency image, the second-view radio frequency image, and the location label information corresponding to the second original radio frequency fingerprint information, determine the second loss function to construct and train the location predictor.

[0053] Specifically, based on the amplitude-view radio frequency image and phase-view radio frequency image corresponding to the second original CSI fingerprint information, the next step is to construct a second loss function for training the location predictor using the above two sets of images and the location label information included in the second original CSI fingerprint information.

[0054] For the construction method of the second loss function, please refer to [link / reference] in some implementations. Figure 6 Step 04 specifically includes: Step 041: Based on the convolutional neural network, determine the first feature vector according to the first-view radio frequency image corresponding to the second original radio frequency fingerprint information; Step 042: Based on the convolutional neural network, determine the second feature vector according to the second-view radio frequency image corresponding to the second original radio frequency fingerprint information; Step 043: Based on cross-entropy loss, determine the second loss function according to the first feature vector, the second feature vector, and the position label information.

[0055] It should be noted that the execution order between steps 041 and 042 is arbitrary. Figure 6 This is for illustrative purposes only and should not be interpreted as a limitation on the order of execution.

[0056] Specifically, firstly, based on the amplitude-view RF image and phase-view RF image generated from the second original CSI fingerprint information, feature vector acquisition is performed using the feature encoder trained in the above embodiment to obtain the corresponding amplitude-view feature vector and phase-view feature vector, respectively. Then, these feature vectors are combined with the location label information included in the second original CSI fingerprint information to construct the second loss function, as shown in Equation 8. It should be noted that the training process of the location predictor generally uses the softmax algorithm, and loss optimization is performed based on the aforementioned second loss function, which is constructed based on cross-entropy loss.

[0057] ………… Formula 8 in, This represents the second loss function mentioned above. k The total number of second original CSI fingerprint information. , Represents cross-entropy loss, Indicates the first j The amplitude-view radio frequency image corresponding to the second set of original CSI fingerprint information. Indicates the first j Phase-view radio frequency image corresponding to the second set of original CSI fingerprint information. This refers to the feature encoder in the above embodiment, whose output is the feature vector corresponding to the radio frequency image within parentheses. `concat` indicates concatenation. This represents the location predictor, whose output value represents the probability distribution of the dual-view fused features belonging to each location label (within square brackets). Indicates the first j The location tag information included in the second original second CSI fingerprint information of the group.

[0058] The core purpose of the second loss function constructed by Formula 8 is to use a small number of second original CSI fingerprint information as training samples to train a position predictor capable of determining the object's position based on dual-view fusion features. This is applied to the feature encoder trained in the above implementation, and the feature vectors of the amplitude view and phase view are concatenated. Compared to a single view, this results in more comprehensive features and stronger resistance to environmental interference. Furthermore, by minimizing the sum of cross-entropy losses, the position predictor... The system learns to map fused features to the correct location labels. Based on this, the training of the location predictor is performed using a second loss function, aiming to minimize this second loss function, ultimately resulting in a localization model that meets the localization requirements.

[0059] Thus, this application constructs a training set based on both amplitude and phase perspectives, and further trains the localization model through comparative self-supervision, enabling the model to learn the latent invariant representations in radio frequency fingerprint data and better adapt to dynamic changes in the environment, thereby fundamentally improving the universality of signal feature representation and enhancing the generalization ability of the localization model.

[0060] The radio frequency fingerprint-based localization model in this application is constructed according to the radio frequency fingerprint-based localization model construction method in the above embodiments.

[0061] Please see Figure 7 The radio frequency fingerprint-based localization method in this application is implemented based on the localization model constructed according to the radio frequency fingerprint-based localization model construction method in the above embodiments, and the method includes: Step 001: Based on the radio frequency fingerprint information of the object to be located, generate the first-view radio frequency image and the second-view radio frequency image data corresponding to the object to be located; Step 002: Based on the feature encoder, determine the corresponding first-view feature vector according to the first-view radio frequency image, and determine the corresponding second-view feature vector according to the second-view radio frequency image data. Step 003: Based on the location predictor, determine the estimated location of the object to be located according to the first-view feature vector, the second-view feature vector, and the location label information saved when constructing the training location predictor.

[0062] Specifically, based on the above embodiments, the localization model trained according to the radio frequency fingerprint-based localization model construction method described in the above embodiments can be tested using the following method, or the same method can be used to apply the localization model to perform actual localization tasks. The specific method is as follows.

[0063] For example, firstly, the CSI fingerprint information corresponding to the object to be located is input into the localization model. This CSI fingerprint information typically includes amplitude-view angle information and phase-view angle information. Then, based on the amplitude-view angle information and phase-view angle information, and referring to the above implementation method, the input quantity of the feature encoder in the localization model, i.e., the amplitude-view angle radio frequency image, can be determined. and phase-view radio frequency images .

[0064] Next, based on the feature encoder in the localization model, according to the amplitude-view radio frequency image... and phase-view radio frequency images By extracting feature vectors, amplitude-view feature vectors can be obtained. and phase view feature vector Then, based on the above... as well as By combining the location label information saved during the training of the location predictor, the estimated location value of the object to be located is determined, thereby achieving localization.

[0065] For details on how to determine the location estimate, please refer to [link / reference]. Figure 8 In some implementations, step 003 includes: Step 0031: Based on the position predictor, determine the output posterior probability of the position predictor according to the first view feature vector and the second view feature vector; Step 0032: Determine the estimated location of the object to be located based on the posterior probability and the location label information saved when constructing the training location predictor.

[0066] Specifically, based on the above implementation method, firstly, according to the obtained amplitude-view feature vector... and phase view feature vector The splicing is a dual-view fusion feature, as shown in Formula 9.

[0067] ………… Formula 9 in It is a dual-view fusion feature.

[0068] Then, based on the location predictor and using the aforementioned dual-view fusion features, softmax regression is employed to calculate the output posterior probability of the location predictor. P Specifically, as shown in Formula 10.

[0069] ………… Formula 10 Output posterior probability PSpecifically, it means: given dual-view fusion features In this case, Corresponding to location label The probability distribution, that is This is the output value of the location predictor for the object to be located.

[0070] Furthermore, for all the location labels saved during the training of the location predictor, a weighted average of the output posterior probabilities described above is applied to obtain the final location estimate of the localization model. This location estimate describes the location information corresponding to the location of the object to be located with the highest probability. See Equation 11 for details.

[0071] ………… Formula 11 in The above location estimates, This indicates the position predictor's stored first... q There are 10 location labels, because the location predictor was trained with a total of 1000 inputs. k There are a maximum of 1,000 second original CSI fingerprints, and each second original CSI fingerprint corresponds to a location tag. Therefore, the location predictor can store a maximum of 1,000 locations. k Location label, This indicates the position predictor's stored first... q The actual location information corresponding to each location label, such as location coordinates.

[0072] The process of obtaining the location estimate described above can be used as a testing process to test the localization model trained and constructed in the above embodiments, or as a practical localization method when the localization model is actually applied. During testing, the CSI fingerprint information corresponding to the object to be located should include the location label or the actual location information corresponding to the location label, to verify the output location estimate. However, in actual localization, the CSI fingerprint information corresponding to the object to be located does not need to include the location label information.

[0073] The above description is merely a preferred embodiment of this application and is not intended to limit this application in any way. Although this application has disclosed the preferred embodiment as above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the technical solution of this application, based on the technical essence of this application and within the spirit and principles of this application, shall still fall within the protection scope of the technical solution of this application.

Claims

1. A method for constructing a localization model based on radio frequency fingerprinting, characterized in that, The localization model includes a feature encoder and a location predictor, and the method includes: A radio frequency image training set is constructed based on multiple first original radio frequency fingerprint information, wherein the first original radio frequency fingerprint information is the first original CSI fingerprint information, and the radio frequency image training set includes a first-view radio frequency image and a second-view radio frequency image. The first view is an amplitude view or a phase view, and the second view is an amplitude view or a phase view. The first view and the second view are different, and each first original radio frequency fingerprint information corresponds to a first-view radio frequency image and a second-view radio frequency image. Based on contrastive self-supervised learning, a first loss function is determined according to the radio frequency image training set in order to construct and train the feature encoder; Based on multiple second original radio frequency fingerprint information, determine the first-view radio frequency image, the second-view radio frequency image, and the location tag information corresponding to the second original radio frequency fingerprint information; Based on the first-view radio frequency image, the second-view radio frequency image, and the location label information corresponding to the second original radio frequency fingerprint information, a second loss function is determined to construct and train the location predictor. The method of contrastive self-supervised learning, which determines a first loss function based on the radio frequency image training set to construct and train the feature encoder, includes: Based on the radio frequency image training set, the first viewpoint feature embedding and the second viewpoint feature embedding are determined; A first initial loss function is determined based on the radio frequency image training set, the first viewpoint feature embedding, and the second viewpoint feature embedding; Based on the non-backpropagation of neural network gradients, the first initial loss function is optimized and determined to construct and train the feature encoder; The step of determining the first initial loss function based on the radio frequency image training set, the first viewpoint feature embedding, and the second viewpoint feature embedding includes: Based on the prediction network, the first viewpoint feature prediction is determined according to the first viewpoint feature embedding; Based on the prediction network, the second perspective feature prediction is determined according to the second perspective feature embedding; The first initial loss function is determined based on the similarity between the first viewpoint feature embedding and the second viewpoint feature prediction, and the similarity between the first viewpoint feature prediction and the second viewpoint feature embedding; the formula for the first initial loss function is: ； in This represents the first initial loss function. Indicates the first i The initial loss corresponding to the first original CSI fingerprint information in the group, K This represents the total number of the first original CSI fingerprint information. For the first i Predicting the amplitude-view feature corresponding to the first original CSI fingerprint information in the group. For the first i Predicting the phase view features corresponding to the first original CSI fingerprint information in the group. For the first i Embed the phase view feature corresponding to the first original CSI fingerprint information. For the first i The amplitude-view feature embedding corresponding to the first original CSI fingerprint information is used; The optimization of the first initial loss function based on neural network gradient non-backpropagation, and the determination of the first loss function, include: Based on the first initial loss function, gradient non-backpropagation processing is performed on the first viewpoint feature embedding and the second viewpoint feature embedding to determine the first loss function.

2. The method according to claim 1, characterized in that, The step of constructing a radio frequency image training set based on multiple first original radio frequency fingerprint information includes: Based on the first original radio frequency fingerprint information, obtain the first view matrix and the second view matrix; Generate a first-view radio frequency image based on the first-view matrix; Based on the second viewpoint matrix, a second viewpoint radio frequency image is generated to construct the radio frequency image training set.

3. The method according to claim 1, characterized in that, The step of determining the first viewpoint feature embedding and the second viewpoint feature embedding based on the radio frequency image training set includes: Based on the convolutional neural network, the corresponding first-view feature vector is determined according to the first-view radio frequency image; Based on a convolutional neural network, the corresponding second-view feature vector is determined according to the second-view radio frequency image; Based on a multilayer perceptron network, the first viewpoint feature embedding is determined according to the first viewpoint feature vector; Based on a multilayer perceptron network, the second-view feature embedding is determined according to the second-view feature vector.

4. The method according to claim 1, characterized in that, The step of determining a second loss function based on the first-view radio frequency image, the second-view radio frequency image, and the location label information corresponding to the second original radio frequency fingerprint information, in order to construct and train the location predictor, includes: Based on a convolutional neural network, a first feature vector is determined according to the first-view radio frequency image corresponding to the second original radio frequency fingerprint information; Based on a convolutional neural network, a second feature vector is determined according to the second-view radio frequency image corresponding to the second original radio frequency fingerprint information; Based on cross-entropy loss, the second loss function is determined according to the first feature vector, the second feature vector, and the location label information.

5. A localization model based on radio frequency fingerprinting, characterized in that, The localization model is constructed according to the localization model construction method based on radio frequency fingerprint as described in any one of claims 1-4.

6. A positioning method based on radio frequency fingerprinting, characterized in that, The method is implemented based on a positioning model constructed according to the positioning model construction method based on radio frequency fingerprints as described in any one of claims 1-4, and the method includes: Based on the radio frequency fingerprint information of the object to be located, generate a first-view radio frequency image and a second-view radio frequency image data corresponding to the object to be located. Based on the feature encoder, a corresponding first-view feature vector is determined according to the first-view radio frequency image, and a corresponding second-view feature vector is determined according to the second-view radio frequency image data. Based on the location predictor, the estimated location of the object to be located is determined according to the first view feature vector, the second view feature vector, and the location label information saved when constructing and training the location predictor.

7. The method according to claim 6, characterized in that, The step of determining the estimated location of the object to be located based on the location predictor, according to the first view feature vector, the second view feature vector, and the location label information, includes: Based on the location predictor, the output posterior probability of the location predictor is determined according to the first view feature vector and the second view feature vector. The estimated location of the object to be located is determined based on the posterior probability and the location label information saved when constructing and training the location predictor.

Images