Opportunity signal self-positioning method based on time-frequency multi-scale mask pre-training large model

Abstract

The invention discloses an opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-training large model, belongs to the technical field of opportunity signal self-positioning, and solves the problem of insufficient positioning precision caused by limited label data of an existing positioning task. The method comprises the following steps: acquiring time-frequency representation of wavelet packet transform domain sequences of a plurality of signal segments corresponding to each opportunity signal sample; constructing a pre-training set and a labeled data set; training a pre-training neural network based on the random-time frequency mask by using the pre-training set; cutting the pre-trained neural network, adding an average pooling layer and a multi-layer perceptron, and constructing a positioning model; training and verifying the positioning model by using a labeled data set to obtain a successfully verified positioning model; opportunity signals of a point to be positioned are collected in real time, time-frequency representation of the wavelet packet transform domain sequences of the corresponding multiple signal segments is obtained, and coordinates of the point to be positioned are obtained through prediction after processing of the verified positioning model.

Description

Technical Field

[0001] This invention relates to the field of opportunity signal self-localization technology, and in particular to an opportunity signal self-localization method based on a time-frequency multi-scale mask pre-trained large model. Background Technology

[0002] The Global Positioning System (GPS) is currently the most widely used positioning technology, performing excellently in open outdoor environments. However, in indoor environments, urban canyons, or environments with severe multipath effects, GPS positioning accuracy drops significantly or even becomes completely ineffective. To compensate for the limitations of GPS, researchers are constantly exploring other alternative or auxiliary positioning technologies to adapt to more complex application scenarios.

[0003] Opportunistic fingerprint localization offers a feasible path to large-scale, low-cost, indoor / outdoor integrated localization without relying on additional deployments. However, despite its ability to adapt effectively to complex environments, its practical application remains limited. The main challenge lies in the fact that this method typically requires the prior acquisition of large amounts of labeled data to train the model, a time-consuming and labor-intensive process. Furthermore, the fingerprint database needs frequent updates in dynamic environments, resulting in high system maintenance costs and hindering large-scale application.

[0004] Therefore, reducing reliance on labeled data and improving the generalization ability and robustness of models has become a pressing issue in current opportunity signal localization research. Given the difficulty in obtaining labeled data, while unlabeled data is relatively easy to collect, there is an urgent need to improve the performance of opportunity signal-based localization methods with limited labeled data. Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide an opportunistic signal self-localization method based on a time-frequency multi-scale mask pre-trained large model, in order to solve the problem of insufficient localization accuracy caused by limited label data in existing localization tasks.

[0006] This invention discloses a method for self-localization of opportunistic signals based on a time-frequency multi-scale mask pre-trained large model, the method comprising:

[0007] Obtain the time-frequency representation of the wavelet packet transform domain sequence of multiple signal segments corresponding to each opportunity signal sample; select a portion of the time-frequency representations to construct a pre-training set; construct a labeled dataset based on the remaining time-frequency representations and the coordinate labels of the corresponding opportunity signal samples.

[0008] The pre-training set is used to train the pre-trained neural network based on random-time-frequency masking;

[0009] The pre-trained neural network is pruned, and an average pooling layer and a multilayer perceptron are added to construct a localization model; the localization model is trained and validated using a labeled dataset to obtain a validated localization model.

[0010] Opportunity signals of the point to be located are acquired in real time, and the time-frequency representation of the wavelet packet transform domain sequence of the corresponding multiple signal segments is obtained. The coordinates of the point to be located are then predicted after processing by the validated positioning model.

[0011] Based on the above solution, the present invention also makes the following improvements:

[0012] Furthermore, the pre-trained neural network based on random-time-frequency masking is trained using the pre-training set, and the following steps are performed:

[0013] The pre-trained neural network is trained multiple times using the time-frequency representations in the pre-training set until the training round threshold is reached or the average loss is less than the loss threshold. The training ends and a successfully trained pre-trained neural network is obtained.

[0014] Furthermore, during each training round of the pre-trained neural network, the following operations are performed:

[0015] Two patch embedding layers are used to process the time-frequency representation separately to obtain several embedded feature blocks;

[0016] A random mask and a time-frequency mask are respectively used to perform random masking and time-frequency masking on the plurality of embedded feature blocks;

[0017] The unmasked embedded feature blocks after processing with random mask and time-frequency mask are input into the multi-scale transform encoder for feature extraction;

[0018] The feature extraction results corresponding to the random mask and the time-frequency mask are linearized by using a linear layer to obtain the corresponding linearized embedded feature blocks;

[0019] The embedded feature blocks corresponding to the linear representations of the random mask and the time-frequency mask are input into the multi-scale transform decoder for time-frequency representation reconstruction.

[0020] The average loss is calculated based on the reconstruction results of the time-frequency representations corresponding to the random mask and the time-frequency mask, and the model parameters of the pre-trained neural network are updated.

[0021] Furthermore, the pre-trained neural network comprises two patch embedding layers, a random mask, a time-frequency mask, a multi-scale transform encoder, a linear layer, a multi-scale transform decoder, and a reconstruction layer.

[0022] Furthermore, the multi-scale converter encoder consists of multiple cascaded encoders; each encoder consists of a first random mask branch, a first time-frequency mask branch, and a first cross-branch attention module.

[0023] Furthermore, the multi-scale transformer decoder consists of multiple cascaded decoders; each decoder consists of a second random mask branch, a second time-frequency mask branch, and a second cross-branch attention module.

[0024] Furthermore, the construction of the localization model involves:

[0025] The pre-trained neural network is pruned, retaining the linear layers and the parts preceding them, along with their parameters.

[0026] After the linear layer, an average pooling layer and a multilayer perceptron are added in sequence to construct the localization model.

[0027] Furthermore, the labeled dataset includes a training set and a validation set; the process of training and validating the localization model using the labeled dataset involves:

[0028] The localization model is trained by using the time-frequency representations in the training set as inputs and the coordinate labels of the corresponding opportunity signal samples as labels, thereby obtaining a successfully trained localization model.

[0029] Each time-frequency representation in the validation set is input into the localization model. The average error between the coordinate labels predicted by the localization model and the coordinate labels of the corresponding opportunity signal samples in the validation set is calculated. If the average error is less than the error threshold, the localization model that has passed the validation is obtained.

[0030] Furthermore, the coordinates of the point to be located are predicted after processing by the validated localization model, and then the following steps are executed:

[0031] The time-frequency representation of the wavelet packet transform domain sequence of each signal segment is input into the validated localization model, and the localization model outputs the predicted coordinate values.

[0032] Calculate the average of the predicted coordinates for all signal segments, and use this average as the coordinates of the point to be located.

[0033] Furthermore, the step of obtaining the time-frequency representation of the wavelet packet transform domain sequence of multiple signal segments corresponding to each opportunity signal sample is performed as follows:

[0034] For each chance signal sample, a signal segment is randomly extracted to obtain multiple signal segments; each signal segment is then normalized to obtain a normalized signal segment.

[0035] Wavelet packet transform and time-frequency conversion are performed on each normalized signal segment to obtain the time-frequency representation of the wavelet packet transform domain sequence.

[0036] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0037] The present invention provides an opportunistic signal self-localization method based on a time-frequency multi-scale mask pre-trained large model. This method improves the accuracy of localization tasks by combining the time-frequency representation of opportunistic signals with a semi-supervised approach, while also addressing the problem of insufficient localization accuracy when label data is limited. Specific effects are described below.

[0038] 1. This invention completes the training of a semi-supervised learning model by converting the opportunistic signal into a wavelet packet transform domain sequence as input, which greatly reduces the requirement for the number of labeled opportunistic signal training samples for opportunistic signal-based localization and improves the generalization ability and robustness of the model.

[0039] 2. This invention learns the feature information of opportunistic signal wavelet packet transform domain sequences through a dual-branch structure. The model can effectively extract features at different scales and improves the ability to model complex signal patterns while reducing computational burden. Furthermore, the use of cross-attention to fully integrate the multi-scale features learned from the two branches is beneficial for improving feature learning performance.

[0040] 3. The present invention employs a dual masking mechanism for the opportunity signal samples of the location point, which is beneficial for the neural network to further learn the characteristics of the opportunity signal during the pre-training stage, and provides better network parameters for the subsequent supervised fine-tuning stage.

[0041] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0042] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0043] Figure 1 The flowchart illustrates the opportunistic signal self-localization method based on a time-frequency multi-scale mask pre-trained large model provided in this embodiment of the invention.

[0044] Figure 2 This is a neural network architecture diagram for a localization task provided in an embodiment of the present invention.

[0045] Figure 3The following is a detailed structural diagram of MS-ViT provided in an embodiment of the present invention.

[0046] Figure 4 The structural diagram of Cross-Attention provided in the embodiments of the present invention. Detailed Implementation

[0047] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0048] A specific embodiment of the present invention discloses a method for self-localization of opportunistic signals based on a time-frequency multi-scale mask pre-trained large model, the flowchart of which is shown below. Figure 1 As shown, the specific explanation is as follows.

[0049] Step S1: Obtain the time-frequency representation of the wavelet packet transform domain sequence of multiple signal segments corresponding to each opportunity signal sample; select a portion of the time-frequency representations to construct a pre-training set; construct a labeled dataset based on the remaining time-frequency representations and the coordinate labels of the corresponding opportunity signal samples.

[0050] In this embodiment, by dividing the data into grid points, opportunity signal samples are collected for each grid point, thereby obtaining multiple opportunity signal samples and their coordinate labels (coordinates of the grid points). Each opportunity signal sample is a discretized digital signal sequence of equal length, and the sampled opportunity signal sequence s(n) = s(t)δ T δ(t), where s(t) is the analog domain chance signal, δ T (t) is the impact function, T = 1 / F s T is the sampling period, F s The sampling frequency.

[0051] Step S11: Randomly extract signal segments from each opportunity signal sample to obtain multiple signal segments; normalize each signal segment to obtain a normalized signal segment.

[0052] Specifically, signal segments are randomly extracted from the opportunity signal sample to obtain multiple signal segments. The length of each extracted opportunity signal segment is fixed at N and is less than the length of the original signal. The extracted signal segment can be represented as s. t (n) = s(n+b), where n = 0, 1, ..., N-1, and N is the length N of the original chance signal sample sequence s(n). S A positive integer, b is a randomly generated integer less than N. S -N is a non-negative integer.

[0053] Through normalization, each truncated signal segment is normalized to 1. The normalized signal segment can be represented as:

[0054] Step S12: Perform wavelet packet transform and time-frequency conversion on each normalized signal segment to obtain the time-frequency representation of the wavelet packet transform domain sequence.

[0055] Specifically, a P-order discrete wavelet packet transform is performed on the normalized signal segment sequence x(n). The mother wavelet uses the "haar" wavelet, where P is greater than 1 and less than floor(log2(N)), where floor(log2(N)) represents the largest integer not exceeding log2(N). The approximate components and detail coefficients after each order wavelet packet transform are concatenated into a time-frequency representation, thus obtaining the time-frequency representation of the wavelet packet transform domain sequence.

[0056] Then, a portion of the time-frequency representations can be selected to construct a pre-training set; based on the remaining time-frequency representations and the coordinate labels of the corresponding opportunistic signal samples, a labeled dataset can be constructed.

[0057] It should be noted that the pre-training set does not contain corresponding coordinate labels, reflecting the limited label data available for current localization tasks. This embodiment uses this as a basis for subsequent pre-training, training, and validation operations, effectively addressing the problem of insufficient localization accuracy caused by limited label data in existing localization tasks. Specifically, the neural network is pre-trained using the pre-training set to achieve effective learning of feature representations and improve generalization capabilities. The localization model is trained and validated using a labeled dataset (including training and validation sets) to support subsequent spatial location estimation based on signal features. For example, the pre-training set, training set, and validation set can be divided in a 6:3:1 ratio.

[0058] Accordingly, this embodiment constructs a pre-trained network for opportunity signal pre-training (corresponding to the pre-training stage) and a neural network for localization tasks (corresponding to the supervised fine-tuning stage). Detailed explanations are as follows.

[0059] Step S2: Train the pre-trained neural network based on random-time-frequency mask using the pre-training set.

[0060] Preferably, in this embodiment, the pre-trained neural network based on random-time-frequency masking includes two patch embedding layers, a random mask and a time-frequency mask, a multi-scale transform encoder, a linear layer, a multi-scale transform decoder, and a reconstruction layer. The structure of the pre-trained neural network is shown in Table 1.

[0061] Table 1. Structure of the pre-trained network

[0062]

[0063] In the specific implementation process, the pre-training set is used to train the pre-trained neural network based on random-time-frequency mask. The process is as follows: the pre-training neural network is trained in multiple rounds using the time-frequency representation in the pre-training set; training ends when the training round threshold is reached or the average loss is less than the loss threshold, and the trained pre-trained neural network is obtained.

[0064] The principle of pre-trained model training is as follows: the input time-frequency representation is first divided into fixed-size embedded feature blocks, which are then processed by random masking and time-frequency masking to generate two sets of inputs with complementary features. The unmasked tokens are input to the encoder to extract features, and then sent to the decoder for time-frequency representation reconstruction.

[0065] During each training round of the pre-trained neural network, perform the following operations:

[0066] Step S21: Use two patch embedding layers to process the time-frequency representation to obtain several embedded feature blocks (Tokens);

[0067] Step S22: The random mask and the time-frequency mask perform random masking and time-frequency masking on the plurality of embedded feature blocks, respectively;

[0068] Step S23: Input the unmasked embedded feature blocks after processing with random mask and time-frequency mask into the multi-scale transform encoder for feature extraction;

[0069] Step S24: Linearize the feature extraction results corresponding to the random mask and the time-frequency mask using a linear layer to obtain the corresponding linearized embedded feature blocks;

[0070] Step S25: Input the embedded feature blocks of the linearized representations corresponding to the random mask and the time-frequency mask into the multi-scale transform decoder for time-frequency representation reconstruction;

[0071] Step S26: Calculate the average loss based on the reconstruction results of the time-frequency representations corresponding to the random mask and the time-frequency mask, and update the model parameters of the pre-trained neural network.

[0072] Preferably, in this embodiment, the multi-scale converter encoder consists of multiple cascaded encoders. Each encoder comprises a first random mask branch, a first time-frequency mask branch, and a first cross-attention module. The outputs of the first random mask branch and the first time-frequency mask branch are respectively connected to the input of the first cross-attention module. The first random mask branch and the first time-frequency mask branch are used to encode the received embedded feature blocks, and the first cross-attention module is used for multi-scale feature fusion of the encoding results of the first random mask branch and the first time-frequency mask branch. More specifically, the unmasked embedded feature blocks after processing with random masks and time-frequency masks are respectively input to the first random mask branch and the first time-frequency mask branch in the first-stage encoder; the two outputs of the first cross-attention module of the previous stage encoder are respectively connected to the first random mask branch and the first time-frequency mask branch of the next stage encoder.

[0073] Preferably, in this embodiment, the multi-scale transformer decoder consists of multiple cascaded decoders. Each decoder comprises a second random mask branch, a second time-frequency mask branch, and a second cross-attention module. The outputs of the second random mask branch and the second time-frequency mask branch are respectively connected to the input of the second cross-attention module. The second random mask branch and the second time-frequency mask branch are used to decode the received embedded feature blocks, and the second cross-attention module is used for multi-scale feature fusion of the encoding results of the second random mask branch and the second time-frequency mask branch. More specifically, the linearized representations corresponding to the random mask and the time-frequency mask are respectively input to the second random mask branch and the second time-frequency mask branch in the first-level decoder; the two outputs of the second cross-attention module of the previous stage decoder are respectively connected to the second random mask branch and the first time-frequency mask branch of the next stage decoder.

[0074] As described above, in this embodiment, both the encoder and decoder employ a multi-scale vision transformer (MS-ViT). The encoder structure based on MS-ViT is as follows: Figure 3 As shown, the decoder structure is similar. The overall model consists of K stacked MS-ViT modules, each containing two branches. Branch A is a random masking branch, employing a random masking strategy to enhance the model's robustness to local perturbations and effectively capture small but important local features; Branch B is a time-frequency masking branch, using a random masking strategy on the time and frequency axes to further enhance the model's ability to model spectral structure and redundant patterns.

[0075] To fully integrate the multi-scale features learned from the two branches, this embodiment utilizes a cross-branch attention module. Figure 4 This demonstrates the cross-attention module in branch A. Let the CLS Token of branch A be... The Patch Token comes from branch B and is denoted as... first, Projection function f with dimension alignment A (·) Map the input dimensions, then use As the only query item; and The concatenation is used as input to the cross-attention module to perform the cross-attention operation. To enhance representation capabilities, layer normalization and residual connection mechanisms are added after the cross-attention module, and their calculation process is shown below: Where f A (·) and g A (·) represents the projection and anti-projection operations respectively, || represents the concatenation operation, LN(·) is the layer normalization function, and MCA(·) represents the multi-head cross attention mechanism, which only uses the CLS Token as the query, so the generated attention has a linear complexity of O(N), rather than the quadratic complexity of O(N) in the traditional full attention mechanism. 2 This significantly improves computational efficiency. Simultaneously, it resembles the standard Transformer structure, employing a multi-head mechanism, but unlike the self-attention module, it does not introduce FFN after cross-attention. Finally, the output of the cross-attention module is as follows:

[0076] During the pre-training phase, a pre-training set was used for training. The batch size was set to 40, the number of training epochs was 400, and the AdamW optimizer was used with a weight decay factor of 0.05. The learning rate adopted the Cosine annealing strategy with an initial learning rate of 0.001. In the first 50 epochs, the Step Decay (StepLR) method was used for learning rate warm-up, with the specific parameters being: step size η = 5, decay factor γ = 0.5, and the initial learning rate also set to 0.001. The pre-training samples were unlabeled samples from the first day of the corresponding indoor and outdoor scenes, respectively, as the pre-training dataset.

[0077] The average loss (reconstruction loss) uses mean-square error (MSE), which only calculates the reconstruction error of the masked region, guiding the model to learn the global structure from local observations and obtain a more generalized time-frequency feature representation.

[0078] It should be noted that the decoder is only used in the pre-training phase.

[0079] Step S3: Prune the pre-trained neural network and add an average pooling layer and a multilayer perceptron to construct a localization model; train and validate the localization model using a labeled dataset to obtain a validated localization model.

[0080] Preferably, the localization model is constructed by: pruning the pre-trained neural network, retaining only the linear layers and preceding layers and their parameters; then, adding an average pooling layer and a multilayer perceptron sequentially after the linear layers to construct the localization model. Specifically, in this stage, the decoder module used in the pre-training stage is removed, retaining only the encoder for extracting multi-scale time-frequency feature representations. An average pooling layer and a multilayer perceptron are added, and the encoder parameters are imported, and the network parameters of the localization model are retrained. Subsequently, the extracted features are fused using an average pooling operation to form a unified global representation vector. This global feature vector is input into the multilayer perceptron to achieve the final task prediction. The entire training process is based on supervised data. Compared to a completely new initialization, fine-tuning the network based on the pre-trained encoder network parameters results in better network training performance.

[0081] In this stage, a fully supervised training strategy is adopted. By introducing training and validation sets, the model can be further optimized for specific task objectives. Specifically, the localization model is trained in a supervised manner using the training set, and then validated using the validation set to obtain a validated localization model. Specifically, the localization model is trained using the time-frequency representations from the training set as inputs and the coordinate labels of the corresponding opportunity signal samples as labels, resulting in a successfully trained localization model. Then, the time-frequency representations from the validation set are input into the localization model, and the average error between the coordinate labels of the predicted output of the localization model and the coordinate labels of the corresponding opportunity signal samples in the validation set is calculated. If the average error is less than the error threshold, the validated localization model is considered successful.

[0082] Specifically, the localization model is trained using the training set, and its performance is validated using a validation dataset after a period of training. In the supervised fine-tuning phase, the training set is used with a batch size of 25 and 20 training epochs. The AdamW optimizer and a weight decay of 0.05 are used, and the learning rate strategy continues to employ cosine annealing (initial value 0.001). The warm-up period is 5 epochs, and the learning rate is linearly increased to 0.001. The drop-path is set to 0.1 in this phase to enhance the model's regularization effect.

[0083] Step S4: Real-time acquisition of opportunity signals of the point to be located, obtaining time-frequency representations of wavelet packet transform domain sequences of multiple corresponding signal segments, and then predicting the coordinates of the point to be located after processing by the validated positioning model.

[0084] In this embodiment, the process of processing the opportunity signal of the point to be located to obtain the time-frequency representation of the wavelet packet transform domain sequence of the corresponding multiple signal segments is similar to the processing process of the opportunity signal sample in step S1, and will not be described again here.

[0085] In step S4, the following operations are performed: the time-frequency representation of the wavelet packet transform domain sequence of each signal segment is input into the verified positioning model, and the positioning model outputs the predicted coordinate values; the average value of the predicted coordinate values ​​corresponding to all signal segments is calculated as the coordinates of the point to be located.

[0086] Assume there are L signal segments, p i Let the predicted coordinates of the wavelet packet transform domain sequence of the i-th signal segment represent the time-frequency representation of the signal segment. Then the coordinates of the point to be located are...

[0087] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0088] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for opportunity signal self-localization based on pre-training a large model with time-frequency multi-scale masks, characterized in that, The method comprises: obtaining the time-frequency representation of the wavelet packet transform domain sequence of each opportunity signal sample corresponding to multiple signal segments; selecting part of the time-frequency representation to construct a pre-training set; and constructing a labeled data set based on the remaining time-frequency representation and the coordinate label of the corresponding opportunity signal sample; training the pre-training neural network based on the random-time-frequency mask using the pre-training set; trimming the pre-training neural network and adding an average pooling layer and a multilayer perceptron to construct a positioning model; training and verifying the positioning model using the labeled data set to obtain a verified positioning model; real-time acquisition of the opportunity signal of the to-be-positioned point to obtain the time-frequency representation of the wavelet packet transform domain sequence of the corresponding multiple signal segments, and prediction of the coordinate of the to-be-positioned point after processing by the verified positioning model.

2. The opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-trained large model according to claim 1, characterized in that, Training the pre-training neural network based on the random-time-frequency mask using the pre-training set, which performs: using the time-frequency representation in the pre-training set to train the pre-training neural network for multiple rounds; until the training round threshold is reached, or the average loss is less than the loss threshold, the training is ended, and a trained pre-training neural network is obtained.

3. The opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-trained large model according to claim 2, characterized in that, In each round of training of the pre-training neural network, the following operations are performed: using two patch embedding layers to process the time-frequency representation to obtain a plurality of embedded feature blocks; the random masker and the time-frequency masker perform random masking and time-frequency masking on the embedded feature blocks, respectively; inputting the embedded feature blocks that are not occluded after random masking and time-frequency masking into a multi-scale transform encoder for feature extraction; using linear layers to perform linearization processing on the feature extraction results corresponding to the random masking and the time-frequency masking, respectively, to obtain linearly represented embedded feature blocks corresponding to the random masking and the time-frequency masking; inputting the linearly represented embedded feature blocks corresponding to the random masking and the time-frequency masking into a multi-scale transform decoder for time-frequency representation reconstruction; calculating the average loss according to the time-frequency representation reconstruction results corresponding to the random masking and the time-frequency masking, and updating the model parameters of the pre-training neural network.

4. The opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-trained large model according to claim 3, characterized in that, The pre-training neural network comprises two patch embedding layers, a random masker, a time-frequency masker, a multi-scale transform encoder, a linear layer, a multi-scale transform decoder, and a reconstruction layer.

5. The opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-trained large model according to claim 4, characterized in that, The multi-scale transform encoder is composed of a plurality of cascaded encoders; each encoder is composed of a first random mask branch, a first time-frequency mask branch, and a first cross-branch cross-attention module.

6. The opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-trained large model according to claim 5, characterized in that, The multi-scale transform decoder is composed of a plurality of cascaded decoders; each decoder is composed of a second random mask branch, a second time-frequency mask branch, and a second cross-branch cross-attention module.

7. The opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-trained large model according to any one of claims 4-6, characterized in that, The construction of the positioning model performs: trimming the pre-training neural network to retain the linear layer and the part before the linear layer and the parameters thereof in the pre-training neural network; adding an average pooling layer and a multilayer perceptron in sequence after the linear layer to construct a positioning model.

8. The opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-trained large model according to claim 7, characterized in that, The labeled data set comprises a training set and a verification set; the training and verification of the positioning model using the labeled data set perform: The positioning model is trained by taking each time-frequency representation in the training set as an input of the positioning model and taking the coordinate label of the corresponding opportunity signal sample as a label of the positioning model, to obtain a training-passed positioning model; The positioning model is input with each time-frequency representation in the verification set, and an average error between the coordinate label predicted by the positioning model and the coordinate label of the corresponding opportunity signal sample in the verification set is calculated, and if the average error is less than an error threshold, a verification-passed positioning model is obtained.

9. The opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-trained large model according to claim 8, characterized in that, The coordinate of the to-be-positioned point is predicted by the verification-passed positioning model, and the following is performed: The time-frequency representation of the wavelet packet transform domain sequence of each signal segment is input into the verification-passed positioning model, and a predicted coordinate value is output by the positioning model; An average of the predicted coordinate values corresponding to all signal segments is obtained as the coordinate of the to-be-positioned point.

10. The opportunity signal self-positioning method based on a time-frequency multi-scale mask pre-trained large model according to claim 1, characterized in that, The time-frequency representation of the wavelet packet transform domain sequence of each opportunity signal sample is obtained by performing the following steps: Each opportunity signal sample is randomly intercepted to obtain a plurality of signal segments, and each signal segment is normalized to obtain a normalized signal segment; Each normalized signal segment is subjected to wavelet packet transform and time-frequency conversion to obtain a time-frequency representation of a wavelet packet transform domain sequence.

Images