A positioning method based on FM signal knowledge distillation

Abstract

The application provides a positioning method based on FM signal knowledge distillation, aiming to improve the accuracy and stability of positioning in real-time resource limited application scenarios. First, the FM signals in the positioning area are collected, and the training set and validation set of the FM signals are constructed through random interception and power normalization processing. Then, the teacher and student convolutional neural networks are constructed, the spatial attention mechanism is used to enhance the spatial expression ability of the model, and the deep features and classification knowledge of the teacher model are transmitted to the student model through the knowledge distillation method. The student model learns the feature extraction ability of the teacher model and optimizes its own parameters to approach the positioning accuracy of the teacher model. Finally, the target position is predicted through the Bayesian position estimation combined with the confidence of the network output. The method improves the accuracy and stability of positioning through wavelet packet transform and knowledge distillation, reduces the calculation cost and parameter quantity of the model, and is suitable for resource limited devices and scenes.

Description

Technical Field

[0001] This invention relates to the field of positioning and navigation technology, and in particular to a positioning method based on FM knowledge distillation. Background Technology

[0002] With the rapid development of the Internet of Things (IoT), the demand for accurate, reliable, and real-time positioning technologies is increasing dramatically across various fields, including autonomous driving, logistics tracking, and health monitoring. Global navigation satellite systems, such as the Global Positioning System (GPS), perform well in outdoor environments with strong signals, providing high-precision positioning services. However, GPS also has significant limitations. First, satellite positioning systems rely on precise signal synchronization, determining location by measuring signal propagation time. Microsecond-level time synchronization errors can lead to significant positioning deviations, thus requiring highly accurate time synchronization. Second, the accuracy of these systems is greatly affected in complex environments (such as indoors or densely populated cities). In urban environments, signals are susceptible to electromagnetic interference, obstruction by buildings, and signal attenuation, reducing positioning accuracy and stability, often failing to meet users' demands for accurate and real-time positioning services. Frequency Modulation (FM) signals, as an opportunistic navigation source, utilize non-navigational radio frequency signals in the environment for positioning purposes and have demonstrated excellent positioning accuracy. However, traditional methods primarily rely on received signal strength indicators as fingerprint features. Focusing solely on energy characteristics limits their ability to fully capture the rich information in the raw signal. Furthermore, since these methods primarily use traditional machine learning techniques, they may lack the ability to learn complex, deep features. Existing deep learning-based methods impose a significant computational burden. Therefore, how to reduce computational intensity and improve overall efficiency while ensuring positioning accuracy remains a pressing issue in the field. Summary of the Invention

[0003] This invention aims to provide a localization method based on FM signal knowledge distillation. By analyzing the time-frequency domain information of FM signals, it improves the accuracy of fingerprint localization while avoiding the complex manual feature extraction and excessive computational complexity of traditional fingerprint localization algorithms. The localization method based on FM signal knowledge distillation disclosed in this invention includes the following steps:

[0004] Step 1: Divide the grid and collect FM signals;

[0005] The positioning area is divided into K grids, each grid representing an independent location unit. FM signal samples are collected in each grid, with each sampling lasting for a period of time, and the sampling is repeated multiple times.

[0006] Step 2: Preprocess the acquired FM signal samples;

[0007] The acquired FM signal samples are processed, including random segmentation and power normalization of signal segments. First, a signal segment of fixed length N (N is less than the original signal length) is randomly segmented from the original FM signal sample. Then, the power of the signal segment is normalized. This process is repeated so that each grid point obtains a sequence of l FM signal segments.

[0008] Step 3: Calculate the wavelet packet transform matrix of the FM signal.

[0009] Perform wavelet packet transform on each FM signal segment sequence, and construct the wavelet packet transform matrix based on the transformed high and low frequency detail components;

[0010] Step 4: Construct FM signal training and validation sets

[0011] All FM signal segments are randomly divided into training and validation sets according to a certain ratio. The sample data in the training and validation sets include wavelet packet transform matrices and sample labels. The sample labels correspond to the grid positions where the FM signal segment is collected.

[0012] Step 5: Construct the FM positioning teacher and student model

[0013] Teacher and student models for FM localization are constructed. The teacher model incorporates a spatial attention mechanism in its head to enhance its spatial representation ability, enabling it to focus on features at different locations and improve recognition accuracy. The student model has fewer parameters than the teacher model. The input to both the teacher and student models is the wavelet packet transform matrix of the FM signal, and the output layer is a classification layer with the same number of hidden nodes as the number of grid points.

[0014] Step 6: Train the FM positioning teacher model

[0015] The teacher model was trained and validated using the FM signal training set and validation set, and its network parameters were optimized to obtain the trained and validated teacher model.

[0016] Step 7: Train the FM positioning student model

[0017] During training, both the teacher and student models receive the same input (FM signal training and validation sets). The total loss of the student model is a weighted sum of the student loss and the distillation loss. Through iterative optimization, the parameters of the student model are updated according to the gradient of the loss function. The distillation error only affects the update of the student network parameters, not the teacher network parameters. Furthermore, the student loss and distillation loss are balanced to ensure that the student model learns both the true label information of the data and useful patterns from the teacher model. By iterating through the loss function, the student model can learn the knowledge of the teacher model, thereby improving its localization accuracy and approaching the performance of the teacher model.

[0018] Step 8: Acquire the FM signal of the point to be located.

[0019] Collect the FM signal of the point to be located, with no less than N signal sampling points;

[0020] Step 9: Extract the FM signal segment of the point to be located.

[0021] The FM signal of the location to be located is randomly segmented and the power is normalized. The segment length is fixed at N each time and the power is normalized to 1. This process is repeated multiple times to obtain M FM signal segments of the location to be located.

[0022] Step 10: Calculate the wavelet packet transform matrix of the FM signal at the point to be located.

[0023] For the FM signal segment sequences of the M points to be located, the wavelet packet transform domain sequence of each signal segment sequence is calculated, and the wavelet packet transform matrix is ​​constructed based on the transformed approximate components and detail components.

[0024] Step 11: Estimate the location of the point to be located.

[0025] The M wavelet packet transform matrices of the point to be located are input into the trained student model. The position corresponding to the sample is calculated based on the confidence score and grid coordinates output by the student model. The final position is obtained by averaging the positions of the M samples.

[0026] Compared with the prior art, the present invention can achieve at least the following beneficial effects:

[0027] 1. This invention employs wavelet packet transform to decompose signals at multiple scales, making it particularly suitable for capturing the local features of non-stationary signals. Compared to traditional wavelet analysis, wavelet packet analysis can achieve finer frequency band division, thereby improving the accuracy and efficiency of time-frequency analysis and revealing the multi-level characteristics of the signal.

[0028] 2. This invention transfers knowledge from a complex teacher model to a simple student model through knowledge distillation, which not only significantly reduces the computational cost and number of parameters of the model, but also makes the model lighter and more suitable for devices and scenarios with limited resources while maintaining positioning accuracy.

[0029] 3. This invention randomly segments the FM signal samples of the location to be located and inputs them into a trained student convolutional neural network model. The decision is further optimized by combining confidence level fusion. Bayesian estimation based on grid coordinates and confidence levels effectively reduces location estimation errors and improves positioning accuracy.

[0030] Other features and advantages of the 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 the invention may be realized and obtained from what is particularly pointed out in the description and the drawings. Attached Figure Description

[0031] 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.

[0032] Figure 1 This is a flowchart of the fingerprint localization method based on FM signal knowledge distillation according to the present invention.

[0033] Figure 2 The FM positioning teacher model is constructed.

[0034] Figure 3 The FM positioning student model is constructed.

[0035] Table 1 shows the experimental results of the fingerprint localization method based on FM signal knowledge distillation. Detailed Implementation

[0036] The 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 implementation examples of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0037] like Figure 1 As shown, the present invention provides a positioning method based on FM signal knowledge distillation, comprising the following steps:

[0038] Step S1: Divide the grid and collect FM signals

[0039] The positioning area is divided into K grids, each representing an independent location unit. FM signal samples are collected in each grid, with each sampling lasting for a certain period of time. Multiple samples are collected, each resulting in a discretized digital signal sequence of equal length. The sampled FM signal sequence is s(n) = s(t)δ. T (t), where s(t) is the analog domain FM signal, δ T (t) is the impact function, T = 1 / F s For the sampling period, F s The sampling frequency.

[0040] Step S2: Preprocess FM signal samples

[0041] The acquired FM signal samples are processed, including random segment extraction and power normalization. First, a signal segment of fixed length N is randomly extracted from the original FM signal sample (N is less than the original signal length N).s Next, the power of the signal segment is normalized, and the process is repeated so that each grid point obtains l FM signal segment sequences.

[0042] The specific process for each step is as follows.

[0043] Step S2-1: Randomly extract signal segments

[0044] A segment of the original FM signal is randomly extracted from the original signal sample. The length of the extracted segment is fixed at N and is less than the length of the original signal. The extracted FM signal segment can be represented as s. t (n) = s(n+b), where n = 0, 1, ..., N-1, and the length N of the original signal sequence s(n) is N. s A positive integer, b is a randomly generated integer less than N. s -N is a non-negative integer.

[0045] Step S2-2 Power Normalization

[0046] The extracted signal segment undergoes power normalization, being normalized to 1. The normalized signal segment is represented as follows:

[0047] Step S3: Calculate the wavelet packet transform matrix of the FM signal.

[0048] For the l FM signal segment sequences, calculate the wavelet packet transform matrix for each signal segment sequence. The calculation process for the FM wavelet packet transform sequence is as follows: For the power-normalized FM signal segment sequence s... w (n) Perform a P-order wavelet transform, using the "Haar" wavelet as the wavelet basis, where P is greater than 1 and less than floor(log2(N)), where floor(log2(N)) represents the largest integer not exceeding log2(N). Then, concatenate the high and low frequency detail components of the root node after the wavelet packet transform into a matrix according to the root node coefficients in ascending order, obtaining the wavelet packet transform matrix of the FM signal. This yields the FM wavelet packet transform vector matrix x. w (n);

[0049] Step S4: Construct the FM signal training and validation sets.

[0050] All FM signal segments are randomly divided into training and validation sets according to a certain ratio. The sample data in the training and validation sets include wavelet packet transform matrices and sample labels. The sample labels correspond to the grid positions where the FM signal segment is collected.

[0051] Step S5: Construct the FM positioning teacher and student model

[0052] Construct teacher and student models for FM positioning, as follows: Figure 2 and Figure 3 As shown, the teacher model incorporates a spatial attention mechanism in its head to enhance its spatial representation ability, enabling it to focus on features at different locations and improve recognition accuracy. The student model has fewer parameters compared to the teacher model. The input to both the teacher and student models is the wavelet packet transform matrix of the FM signal, and the output layer is a classification layer with the same number of hidden nodes as the number of grid points.

[0053] The teacher model uses a convolutional neural network such as Figure 2 As shown. "ResnextBlock(64)" indicates that this convolutional layer is a residual network block, a residual block (Bottleneck) with 64 filters. The convolutional neural network used in the student model is as follows: Figure 3 As shown in the figure. "Conv,32,3×3" indicates that this convolutional layer is a traditional convolution with 32 kernels and a kernel size of 3×3. The "Conv" layer in the figure corresponds to a sequence convolutional layer + batch normalization layer + activation layer, where the activation function is ReLU. "Max_Pool" is a max pooling layer with a stride of (1,2), "Global_Avgpool" is global average pooling, and "FC" is a linear layer with an output of K (number of grids).

[0054] Step S6: Train the FM localization teacher model

[0055] The teacher model was trained and validated using the FM signal training and validation sets, respectively, and its network parameters were optimized to obtain the trained and validated teacher model. The convolutional neural network was trained using the training set, and its performance was periodically evaluated using the validation set during training. The AdamW optimizer was used for training, with the cross-entropy function as the loss function. To improve training effectiveness, the learning rate was decayed to half its original value every two training epochs, with the initial learning rate set to 10. -3 Throughout the training process, the model parameters that achieve the highest accuracy on the validation set are saved as the optimal parameters for the convolutional neural network.

[0056] Step S7: Train the FM localization student model

[0057] During training, knowledge distillation is used. Both the teacher and student models receive the same input (FM signal training and validation sets). The total loss of the student model is a weighted sum of the student loss and the distillation loss. Through iterative optimization, the parameters of the student model are updated according to the gradient of the loss function. The distillation loss only affects the update of the student network parameters, not the teacher network parameters. Furthermore, the student and distillation losses are balanced to ensure that the student model learns both the true label information of the data and useful patterns from the teacher model. By iteratively applying the loss function, the student model can learn the knowledge of the teacher model, thereby improving its localization accuracy and approaching the performance of the teacher model.

[0058] In the knowledge distillation process, the teacher model first performs inference on the training set to generate soft labels, which are labels that contain the predicted probability distribution of the input features. Where, q i z represents the probability of each class output by the model. i Let represent the Logits value of the i-th category, where T is the temperature coefficient. The student model uses these soft labels as the training objective loss function to learn the knowledge of the teacher model while maintaining simplicity and low computational cost; the loss function consists of the KL divergence loss function (for soft labels) and the cross-entropy (CE) loss function (for hard labels). total =α·L KL +(1-α)·L CE L KL =KL(q) i teacher |q i student ), q i teacher q represents the softening probability of the i-th category output by the teacher model. i student h represents the softening probability of the i-th category output by the student model. i It is the one-hot encoding of the real label, q 1 i This represents the probability of the student model outputting the i-th class when T=1. α is a hyperparameter between 0 and 1, used to balance the importance of KL divergence loss and cross-entropy loss in the total loss. The distillation temperature T is 5, and α is 0.5 during training.

[0059] Step S8: Acquire the FM signal of the point to be located.

[0060] Acquire the FM signal to be detected to obtain y(n), n=0,1,...,N y -1, its length N y >N.

[0061] Step S9: Extract the FM signal segment of the point to be located.

[0062] The FM signal of the location to be located is randomly truncated and the power is normalized. The length of each truncation is fixed at N, and the power is normalized to 1. This process is repeated multiple times to obtain M FM signal segments of the location to be estimated. The specific truncation and normalization operations are as follows.

[0063] Step S9-1: Randomly extract signal segments

[0064] A signal segment is randomly selected from the FM signal samples y(n) of the point to be located. The length of the selected signal segment is fixed at N. The selected FM signal segment can be represented as y t (n) = y(n+c), where n = 0, 1, ..., N-1, and c is a randomly generated number less than N. y -N is a non-negative integer.

[0065] Step S9-2 Power Normalization

[0066] The power of the intercepted signal is normalized to 1, and the normalized FM signal segment for the location to be estimated is:

[0067] Step S10: Calculate the wavelet packet transform matrix of the FM signal at the point to be located.

[0068] For the M FM signal segment sequences to be estimated, the wavelet packet transform domain sequence of each signal segment sequence is calculated. Based on the transformation, the high and low frequency detail components of the root node are all spliced ​​into a matrix according to the root node coefficient from small to large. The specific calculation process is the same as in step S3.

[0069] Step S11: Estimate the location of the point to be located.

[0070] The M wavelet packet transform matrices of the point to be located are sequentially input into the trained student model. Based on the confidence score and grid coordinates of the model output, the position corresponding to the sample is calculated. The final position is obtained by averaging the positions of the M samples. This includes the following steps: calculating the confidence score for each grid point based on each sample, and using Bayesian estimation to obtain the predicted position based on the coordinate values ​​of each grid point. Where (x) i ,y i () represents the collected grid coordinate values. For each sample's predicted location, p(i) is the confidence level of the i-th grid point in the model output, which is q. 1 i The final position is obtained by averaging the positions of the M samples. Experimental data

[0071] This method was compared with the FM-Pnet model using four performance metrics: Mean Distance Error (MDE), which reflects the average Euclidean distance between the estimated and actual locations; a smaller MDE value indicates higher positioning accuracy. Standard Deviation (STD) measures the dispersion of positioning errors; a smaller STD value indicates better stability of the positioning results. Floating-point operations (#FLOPs) measure the computational complexity of the model; a smaller #FLOPs value means that the model requires fewer computational resources when processing data. The number of parameters (#Params) is a metric for model size, representing the total number of trainable parameters in the model; fewer parameters mean a lighter model.

[0072] The proposed method achieves lower MDE (Mean Deficit Error) than FM-Pnet in both indoor and outdoor environments, indicating higher localization accuracy. STD (Stability Deficit) is also lower than FM-Pnet in both indoor and outdoor environments, suggesting more stable localization results. The FM-Pnet model has a floating-point computation cost of 2848M, while the proposed model has only 117M, representing a reduction of approximately 95.9%. Similarly, the FM-Pnet model has 14M parameters, compared to 0.10M in the proposed model, representing a reduction of approximately 99.3%.

[0073] Table 1

[0074]

[0075] Those skilled in the art will understand that all or part of the above-described implementation methods can be accomplished by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. This computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0076] 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 positioning method based on FM signal knowledge distillation, characterized in that, Includes the following steps: Step 1: Divide the grid and collect FM signals; Step 2: Preprocess the acquired FM signal samples; Step 3: Calculate the wavelet packet transform matrix of the FM signal; Perform wavelet packet transform on each FM signal segment sequence, and construct the wavelet packet transform matrix based on the transformed high and low frequency detail components; A P-order wavelet packet transform is performed on the FM signal segment sequence, where P is greater than 1 and less than floor(log2(N)), where floor(log2(N)) represents the largest integer not exceeding log2(N), and N is the fixed length of the signal segment. The high and low frequency detail components of the root node after the wavelet packet transform are concatenated into a matrix according to the root node coefficients in ascending order to obtain the wavelet packet transform matrix of the FM signal. Step 4: Construct the FM signal training and validation sets; All FM signal segments are randomly divided into training and validation sets according to a certain ratio. The sample data in the training and validation sets include wavelet packet transform matrices and sample labels. The sample labels correspond to the grid positions where the FM signal segment is collected. Step 5: Construct the FM positioning teacher and student models; Teacher and student models for FM localization are constructed. The teacher model incorporates a spatial attention mechanism in its head to enhance its spatial representation ability, enabling it to focus on features at different locations and improve recognition accuracy. The student model has fewer parameters than the teacher model. The input to both the teacher and student models is the wavelet packet transform matrix of the FM signal, and the output layer is a classification layer with the same number of hidden nodes as the number of grid points. Step 6: Train the FM positioning teacher model; The teacher model is trained and validated using the FM signal training set and validation set, respectively. The network parameters of the teacher model are optimized to obtain the trained and validated teacher model. Step 7: Train the FM student positioning model; During the training of the student model, both the teacher and student models receive the same input: the FM signal training and validation sets. The total loss of the student model is a weighted sum of the student loss and the distillation loss. Through iterative optimization, the network parameters of the student model are updated according to the gradient of the loss function, balancing the student loss and the distillation loss to ensure that the student model can learn both the true label information of the data and useful patterns from the teacher model. Through iterative learning of the loss function, the student model can learn the knowledge of the teacher model, thereby improving its localization accuracy and approaching the performance of the teacher model. Step 8: Acquire the FM signal of the point to be located; Step 9: Extract the FM signal segment of the point to be located; Step 10: Calculate the wavelet packet transform matrix of the FM signal at the point to be located; For the FM signal segment sequences of the M points to be located, calculate the wavelet packet transform domain sequence of each signal segment sequence, and construct the wavelet packet transform matrix based on the transformed approximate components and detail components; Step 11: Estimate the location of the point to be located; The M wavelet packet transform matrices of the point to be located are input into the trained student model. The position corresponding to the sample is calculated based on the confidence score and grid coordinates output by the student model. The final position is obtained by averaging the positions of the M samples.

2. The positioning method based on FM signal knowledge distillation according to claim 1, characterized in that, The formula for calculating the spatial attention mechanism in the teacher model is as follows: H s (G s )=σ(f 7×7 ([AvgPool(G s );MaxPool(G s )])) Among them, G s ∈R C*H*W As inputs to spatial attention, AvgPool(·) and MaxPool(·) represent global average pooling and global max pooling, respectively, σ represents the sigmoid function, and H... s For input G s Spatial attention representation, f 7×7 This indicates a convolution operation with a filter size of 7×7.

3. The positioning method based on FM signal knowledge distillation according to claim 1, characterized in that, The structure of the teacher model is as follows: First, SpatialAttention focuses on spatially important regions; then, the model extracts features and models the relationships between channels using a series of ResNeXt blocks, which contain 64, 128, 256, and 512 channels respectively, with the number of repetitions for each channel being [3, 4, 5, 3], to gradually increase the model's expressive power; after processing by the ResNeXt blocks, the feature map is globally averaged using Global_Avg_Pool to compress the spatial dimension and retain key feature information, resulting in a one-dimensional feature vector; subsequently, this one-dimensional feature vector is mapped through a fully connected layer (FC); finally, the Softmax layer converts the output of the fully connected layer into a probability distribution for multi-class tasks, with the output representing the probability that the input belongs to each category; the entire model, through the combination of these layers, can extract key features from the original FM signal and ultimately use them for localization tasks.

4. The positioning method based on FM signal knowledge distillation according to claim 3, characterized in that, The structure of the student model is as follows: Starting from the input raw FM signal, the student model first extracts local features through a 3x3 convolutional layer; then, it downsamples the feature map through a max pooling layer to reduce its spatial size. Subsequently, the model extracts features again through convolutional layers, and finally performs downsampling through max pooling layers. This process is repeated once, followed by convolutional and max pooling layers to extract deeper features and reduce the size of the feature map. After multiple convolutional and pooling operations, the feature map is converted into a one-dimensional feature vector through a global average pooling layer. This vector contains global information of the entire feature map. Then, this feature vector is mapped through a fully connected layer. Finally, the Softmax layer converts the output of the fully connected layer into a probability distribution for multi-class classification tasks, where the output represents the probability that the input belongs to each class.

5. The positioning method based on FM signal knowledge distillation according to claim 1, characterized in that, In the knowledge distillation process of training the student model, the loss function is used. The teacher model first performs inference on the training set to generate soft labels, which are labels containing the predicted probability distribution of the input features. Where, q i z represents the probability of each class output by the model. i Let represent the output Logits value of the i-th category, where T is the temperature coefficient; the student model uses soft labels as the training objective loss function to learn the knowledge of the teacher model while maintaining simplicity and low computational cost; the loss function consists of the KL divergence loss function for soft labels and the cross-entropy loss function for hard labels: L total =α·L KL +(1-α)·L CE L KL =KL(q) i teacher |q i student ), q i teacher q represents the softening probability of the i-th category output by the teacher model. i student y represents the softening probability of the i-th category output by the student model. i It is the one-hot encoding of the real label, q 1 i This represents the probability of the student model outputting the i-th category when T=1, where α is a hyperparameter between 0 and 1.

6. The positioning method based on FM signal knowledge distillation according to claim 1, characterized in that, The final location prediction result is obtained by performing Bayesian estimation based on the grid coordinates and confidence scores. This includes the following steps: calculating the confidence score for each grid point based on each sample, and using Bayesian estimation to obtain the predicted location based on the coordinate values ​​of each grid point. Where (x) i ,y i () represents the collected grid coordinate values. For each sample, the predicted location is given, and p(i) is the confidence level of each grid point in the model output. The final position is obtained by averaging the positions of M samples.

Images