A multi-dimensional radio frequency fingerprint enhancement method based on discrete wavelet transform

Abstract

A multi-dimensional radio frequency fingerprint enhancement method based on discrete wavelet transform comprises the following steps: 1) calculating the positions of all APs by using the positions of known fingerprint collection reference points and the RSSIs received by the corresponding reference points; 2) arranging the collected RSSI sequences according to the AP position information obtained in step 1) to form multi-dimensional fingerprint sequences and preliminarily enhance the original fingerprint; 3) converting the multi-dimensional fingerprint sequences in step 2) into gray-scale images and decomposing the gray-scale images by using discrete wavelet transform to obtain detail coefficients and approximation coefficients; 4) exchanging the detail coefficients of the fingerprint gray-scale images collected at different positions or in different time periods within the same reference point range; and 5) obtaining the finally enhanced fingerprint by inverse wavelet transform of the exchanged detail coefficients and the unexchanged approximation coefficients in step 4). The present application has the advantages of simple design, reduced dependence of the positioning model on the number of tag fingerprints, improved positioning classification accuracy, enhanced robustness of the positioning model, and practical value and realistic promotion significance.

Description

Technical Field

[0001] This invention belongs to the fields of indoor positioning and electronic information science, and in particular, wavelet transform-enhanced radio frequency fingerprinting. Background Technology

[0002] Wi-Fi fingerprint positioning, due to its simple and feasible positioning framework, eliminates the need for additional positioning beacons and signal receiving equipment, making it a preferred solution for indoor positioning. However, in practical applications, Wi-Fi fingerprint positioning requires pre-collecting fingerprints of known locations as reference fingerprints and storing them in a fingerprint database. Then, real-time fingerprints are compared with those in the database to determine the current location. Building a fingerprint database is a tedious process, requiring not only fingerprint collection but also measurement of the collection points' location information. Currently, fingerprint enhancement can alleviate the burden of fingerprint collection, mainly through feature extraction and fingerprint generation methods.

[0003] Feature extraction-based approaches enhance fingerprints by constructing higher-performance feature extraction models to extract deeper features. Lan T. et al. (Lan T., Wang X., Chen Z., et al, Fingerprint Augment Based on Super-Resolution for WiFi Fingerprint Based Indoor Localization[J].IEEE Sensors Journal,22(12):12152–12162,2022.) proposed a fingerprint enhancement framework that transforms the fingerprint enhancement problem into a fingerprint image super-resolution problem. However, the performance of this enhancement framework is closely related to many parameters, such as different reference point distributions and channel characteristics in different environments. Zhang B. et al. (Zhang B., Sifaou H., and Li Y., CSI-Fingerprinting Indoor Localization via Attention-Augmented Residual Convolutional Neural Network[J].IEEE Transactions on Wireless Communications,22(8):5583–5597,2023.) proposed a novel convolutional neural network with attention-enhanced residuals to fully utilize the spatiotemporal features in channel state information to improve the performance of localization networks.

[0004] Fingerprint generation constructs new fingerprints based on existing fingerprint distribution and other information to expand the fingerprint database. Junoh S. et al. (Junoh S. and Pyun J., Enhancing Indoor Localization with Semi-Crowdsourced Fingerprinting and GAN-Based Data Augmentation[J].IEEE Internet Things Journal,11(7):11945–11959,2024.) introduced a generative adversarial neural network-based method, which increased the amount of training data collected for each reference point to enrich the original dataset and reduce human input. Ssekidde P. et al. (Ssekidde, P. Eyobu O., Han D., et al., Augmented CWT Features for Deep Learning-Based Indoor Localization Using WiFi RSSIData[J]. Applied Sciences,11(4):1806,2021.) proposed using continuous wavelet transform to extract a new feature set of RSSI data and using deep learning methods for indoor positioning. Masato Sugasak proposed a novel data augmentation method for indoor localization based on inter-class learning. This method focuses on the positional relationships between each sampled data point in unsampled locations and constructs a generative model of the Wi-Fi signal for each access point using sparse data, thus forming a synthetic augmented fingerprint dataset. Numan P. et al. (Numan P., Park H., Laoudias C., et al, Dropout Autoencoder Fingerprint Augmentation for Enhanced Wi-Fi FTM-RSS Indoor Localization[J].IEEE Communications Letters,27(7):1759–1763,2023.) proposed a fingerprint augmentation method based on a random deactivated autoencoder. By reconstructing unrecorded signal features and applying deep neural network regression, it significantly improves the accuracy of Wi-Fi indoor localization.

[0005] It can be seen that although some researchers have used continuous wavelet transform to extract fingerprint features, the continuous wavelet transform process is more complex than the discrete wavelet transform. With the advent of deep learning, most current research on Wi-Fi fingerprint enhancement mainly relies on deep learning and similar schemes to learn fingerprint distributions to expand the fingerprint database. However, limited by the quantity and quality of training samples, the fingerprint distribution extracted by deep learning models is not necessarily accurate, and the generated new fingerprints may be detrimental to the accuracy of localization. At the same time, the feature distribution of fingerprints changes dynamically, and new samples generated based solely on the current fingerprints may not adapt to the changed environment. In addition, deep learning methods have other drawbacks, such as long training time and significant differences between the generated fingerprints and actual fingerprints. Therefore, designing a lightweight fingerprint enhancement scheme that fully utilizes the similarities and differences in the spatiotemporal features of fingerprint samples to adapt to dynamic environmental changes and improve the accuracy of localization and classification is a problem that urgently needs to be solved. Summary of the Invention

[0006] To address the problems existing in current technologies, this invention provides a multidimensional radio frequency fingerprint enhancement method based on discrete wavelet transform. This method reduces the dependence of the localization model on tag fingerprints, further improving the accuracy of Wi-Fi fingerprint localization and classification.

[0007] This invention is achieved through the following technical solution.

[0008] The multidimensional radio frequency fingerprint enhancement method based on discrete wavelet transform described in this invention includes the following steps:

[0009] Step 1: Calculate the positions of all Access Points (APs) using a weighted average of the known fingerprint acquisition reference point locations and the corresponding Received Signal Strength Indication (RSSI) received by each reference point. If the positions of the nr reference points are {P} ri |0 <i≤nr,i∈N +}, then the position of AP can be represented as:

[0010]

[0011] in, For the estimated location of the j-th AP, P rk Let w be the position of the k-th known reference point. jk This indicates that the j-th AP is at the k-th reference point P. rk The weighted value, This represents the average received signal strength of the j-th AP at the m-th reference point.

[0012] Step 2: Arrange the collected RSSI sequences according to the AP location information obtained in Step 1 to form a multidimensional fingerprint sequence, which initially enhances the original fingerprint and enriches the spatial features of the fingerprint data. The specific construction process is as follows: (1) Receive the signal strength of na APs for n periods at a reference point in the positioning area, and calculate the average signal strength of different APs per period; sort them alternately in the order of maximum, minimum, second largest, second smallest, etc., to form a 1-dimensional feature of the positioning fingerprint. (2) According to the estimated three-dimensional position of the APs, sort the APs according to the horizontal direction (left-to-right direction), vertical direction (up-and-down direction), and depth direction (front-to-back direction) respectively, so as to obtain the other 3-dimensional features of the fingerprint. Finally, merge the multidimensional feature sequence to obtain a fingerprint array of size 4×na.

[0013] Step 3: Normalize the multidimensional fingerprint sequence values ​​from Step 2 to a range of 0-255 and save them as a grayscale image, such as... Figure 1 As shown. Then, the grayscale image is decomposed using discrete wavelet decomposition, as follows. Figure 2 As shown.

[0014] If the grayscale image f(x,y) has dimensions M×N, then the corresponding discrete wavelet decomposition expression is:

[0015]

[0016] Where CA(j0,m,n) are approximation coefficients, and CD l (j,m,n) are the detail coefficients, l∈{H,V,D}, j0 is the scale at which the decomposition begins, and this invention chooses j0=0; the fingerprint size is 4×na, therefore M=4, N=n a j = 0, 1, ... min(1, log₂n) a -1), m=n=0,1,…,2 j -1.

[0017] For two-dimensional scaling functions:

[0018]

[0019] For two-dimensional wavelet functions:

[0020]

[0021] ψ D (x,y)=ψ(x)ψ(y)

[0022] in, Let ψ(x) be the scaling function and ψ(x) be the wavelet function.

[0023] Step 4: Exchange the detail coefficients of fingerprint grayscale images collected from different locations or time periods within the same reference point range. If the set of detail coefficients after fingerprint decomposition is as follows:

[0024] CD = [CD] H CD V CD D ]

[0025] The entire set of coefficients can then be represented as:

[0026] SC = [CD, CA].

[0027] (1) Exchange of detail coefficients of fingerprint grayscale images at different time periods:

[0028] Assume that the fingerprint coefficient sets collected at different times t1 and t2 within the same reference point range are as follows:

[0029]

[0030] Then, the detail coefficients are interchanged:

[0031]

[0032] in, express Medium detail coefficient Replaced with

[0033] (2) Exchange of detail coefficients of fingerprint grayscale images at different locations within the same reference point range:

[0034]

[0035] in, This represents the set of coefficients after the fingerprint detail coefficients at position p1 are swapped with those at position p2.

[0036] Step 5: The set of detail coefficients SC exchanged in Step 4 is transformed using inverse discrete wavelet transform to obtain the final enhanced fingerprint grayscale image, as shown below. Figure 3 As shown.

[0037] Discrete wavelet inverse transform:

[0038]

[0039] Among them, f R (x,y) represents the final enhanced fingerprint grayscale image. The coefficients are the replaced two-dimensional detail coefficients, l∈{H,V,D}, and the other symbols are the same as those of the wavelet forward transform.

[0040] This invention provides a fingerprint enhancement method based on wavelet transform. First, spatial information is embedded into the fingerprint sequence using the virtual location of a Wi-Fi access point (AP), expanding the one-dimensional fingerprint sequence into a multi-dimensional fingerprint grayscale image. Then, based on discrete wavelet transform, different spatiotemporal detail coefficients within the same category are exchanged, thereby expanding the fingerprint database. The proposed wavelet transform-based Wi-Fi fingerprint enhancement scheme effectively reduces the dependence on tagged fingerprints without significantly increasing time complexity, making it suitable for dynamic environments and improving positioning accuracy. This invention is simple in design, reduces the dependence of the positioning model on the number of tagged fingerprints, improves the accuracy of positioning classification, enhances the robustness of the positioning model, and has flexible practical value and significant implications for real-world application. Attached Figure Description

[0041] Figure 1 This is the grayscale image of the multidimensional fingerprint in this invention.

[0042] Figure 2 This refers to the discrete wavelet decomposition of the multidimensional fingerprint grayscale image in this invention.

[0043] Figure 3 This is the enhanced fingerprint grayscale image based on discrete wavelet transform in this invention. (a) is the fingerprint grayscale image before enhancement, and (b) is the fingerprint grayscale image after enhancement.

[0044] Figure 4 This is a diagram of the CNN structure used for classification prediction in this invention.

[0045] Figure 5 This is a schematic diagram of the original 1D fingerprint replication and expansion in this invention.

[0046] Figure 6 These are the test results for 1D and 4D fingerprints in this invention.

[0047] Figure 7 This is a bar chart showing the test results of multidimensional fingerprint enhancement based on discrete wavelet transform in this invention.

[0048] Figure 8 This is a flowchart of the present invention. Detailed Implementation

[0049] To verify the feasibility and effectiveness of the present invention, the present invention will be described in further detail below. The specific embodiments described herein are only for explaining the present invention and are not intended to limit the invention.

[0050] This invention uses the open-source Wi-Fi indoor UJIIndoorLoc dataset (denoted as dataset U) and a self-built dataset (denoted as dataset S) for verification testing. Dataset U contains data from 520 access points (APs) across 3 buildings. To avoid loss of generality, this invention selects data with BUILDINGID 0 and FLOOR 0 for data augmentation testing, resulting in 1059 original data entries. However, in some locations, all APs had no signal, or some APs had no signal at all locations. By filtering out these invalid locations and APs, and then selecting the top 16 AP signals based on the AP's coverage area (the number of reference points capable of receiving the AP's signal) as the test fingerprint database, 246 fingerprint data entries remained from 14 locations. Dataset S collected fingerprint data from 10 reference locations, collected in two batches: 200 fingerprint data entries in the morning (denoted as T0 field) and 200 fingerprint data entries in the afternoon (denoted as T1 field).

[0051] This invention underwent all data processing and testing on a desktop computer equipped with an Intel Core i7-13700KF CPU and an NVIDIA RTX 4080 GPU. Discrete wavelet transform enhancement was performed using the Wavelet Toolbox in MATLAB 2020b software. A Convolutional Neural Network (CNN) was constructed using the PyTorch framework (Python 3.9) as a classification prediction model, and enhanced fingerprints were used for classification prediction. The CNN contains two convolutional layers and two fully connected layers. First, the input single-channel image passes through the first convolutional layer (Conv1), which increases the number of input channels from 1 to 16 and performs convolution operations using a 3x3 kernel, maintaining the output size. Then, it is downsampled through a max-pooling layer (P1) to reduce the size of the feature map. Next, it passes through the second convolutional layer (Conv2), increasing the number of channels to 32, and again performs convolution operations using a 3x3 kernel, maintaining the output size, followed by another max-pooling downsampling operation (P2). After the feature map is flattened into a one-dimensional vector, it is input into the first fully connected layer (Fc1), then through the second fully connected layer (Fc2), and finally outputs the classification result. The CNN structure is as follows: Figure 4 As shown.

[0052] During training, we set the maximum number of training epochs to 500, the initial learning rate to 0.001, and the batch size to 4. We used the Adam optimizer for training, and accuracy was used as the evaluation metric for location recognition classification. In testing, we initially selected Haar wavelets for fingerprint enhancement. To simulate insufficient fingerprint data, the samples were divided into training and test sets in a 2:8 ratio. In the following embodiments, unless otherwise specified, the testing conditions are consistent with those described above.

[0053] Example 1: Verification test of multidimensional fingerprints.

[0054] Table 1 Comparison of classification accuracy between 1D and 4D fingerprints

[0055]

[0056] To verify the effectiveness of the multidimensional fingerprint designed in this invention, tests were conducted on datasets U and S, and compared with commonly used 1D fingerprints. However, the CNN used in this invention requires two-dimensional input, and 1D fingerprints cannot be directly input. To reasonably compare 1D fingerprints with the multidimensional fingerprint (4D fingerprint) proposed in this invention, the corresponding original 1D fingerprint was directly copied and expanded into 4D (e.g., ...). Figure 5 As shown), 1D fingerprints are used as input to the CNN. The test results for 1D and 4D fingerprints are as follows. Figure 6 As shown in Table 1, the 4D fingerprint proposed in this invention significantly improves classification accuracy compared to 1D fingerprints. This is because the 4D fingerprint incorporates the three-dimensional spatial information of the AP points, which is beneficial for the recognition of the localization model. Compared to the original 1D fingerprint, the 4D fingerprint designed in this invention improves the classification accuracy by 19.19% in dataset U and by 13.85% in dataset S. These test results verify the effectiveness of the multi-dimensional fingerprint designed in this invention.

[0057] Example 2: Verification test of multidimensional fingerprint enhancement based on discrete wavelet transform.

[0058] In the test, all samples were dataset S, with 20% of the T0 domain fingerprints used as the training set. The training samples in the training set were enhanced using a 4D fingerprint from each reference point in the T1 domain based on discrete wavelet transform. The test samples comprised 80% of the T1 domain fingerprints. For comparison, the selected T1 domain samples used for discrete wavelet enhancement were directly added to the original 1D fingerprint database (already copied and expanded to 4D) and the T0 domain 4D fingerprint database without discrete wavelet enhancement, denoted as 1DF and 4DF respectively. Furthermore, due to the diverse types of wavelet basis functions for discrete wavelet transform, this invention selected commonly used wavelet basis functions such as haar, dbN (N=1,2,3,4), symN (N=1,2,3,4), and coifN (N=1,2,3,4) for enhancement comparison. The classification accuracy was as follows: Figure 7 As shown in Table 2.

[0059] Table 2 Comparison of classification accuracy of different wavelet bases

[0060]

[0061] As shown in Table 2, the discrete wavelet transform enhancement based on coif3, haar, coif4, and sym3 achieved a classification accuracy of over 90%, representing a maximum improvement of 36.14% compared to 1D fingerprint (1DF) and a maximum improvement of 24.09% compared to 4D fingerprint (4DF). 4DF had the second lowest classification accuracy at only 76.25%, because it only enhanced the T0 domain fingerprint features. Although it introduced a small number of the latest T1 domain fingerprints (one new fingerprint per reference point), its features still differed significantly from those in the T1 domain. Nevertheless, 4DF's classification accuracy was still higher than 1DF because it incorporated the spatial relationship of the AP (Aspect-Oriented Fingerprint), and the AP's position is usually fixed, which can mitigate the dynamic changes in the fingerprint to some extent. The above tests demonstrate that the multidimensional RF fingerprint enhancement method based on discrete wavelet transform designed in this invention is effective.

Claims

1. A multidimensional radio frequency fingerprint enhancement method based on discrete wavelet transform, characterized in that... Includes the following steps: Step 1: Calculate the positions of all APs using the known fingerprint acquisition reference point positions and the weighted average of the RSSI received at the corresponding reference points; if the positions of the nr reference points are {P} ri |0 <i≤nr,i∈N + }, then the position of AP can be represented as: in, For the estimated location of the j-th AP, P rk Let w be the position of the k-th known reference point. jk This indicates that the j-th AP is at the k-th reference point P. rk The weighted value, This represents the average received signal strength of the j-th AP at the m-th reference point; Step 2: Arrange the collected RSSI sequences according to the AP location information obtained in Step 1 to form a multidimensional fingerprint sequence, which initially enhances the original fingerprint and thus enriches the spatial features of the fingerprint data. The construction process is as follows: (1) Receive the signal strength of na APs for n cycles at a reference point in the positioning area, and calculate the average signal strength of different APs per cycle; sort them alternately in the order of maximum, minimum, second largest, second smallest, etc. to form the 1-dimensional feature of the positioning fingerprint; (2) Sort the APs according to the estimated 3D position of the APs in the horizontal direction, vertical direction and depth direction respectively, so as to obtain the other 3D features of the fingerprint; Finally, merge the multi-dimensional feature sequence to obtain a fingerprint array of size 4×na; Step 3: Normalize the multidimensional fingerprint sequence values ​​from Step 2 to between 0 and 255, save them as grayscale images, and then perform discrete wavelet decomposition on the grayscale images. If the grayscale image f(x,y) has dimensions M×N, then the corresponding discrete wavelet decomposition expression is: Where CA(j0,m,n) are approximation coefficients, and CD l (j,m,n) are the detail coefficients, l∈{H,V,D}, j0 is the scale at which the decomposition begins, and this invention chooses j0=0; the fingerprint size is 4×na, therefore M=4, N=n a j = 0, 1, ... min(1, log₂n) a -1), m=n=0,1,…,2 j -1; For two-dimensional scaling functions: For two-dimensional wavelet functions: ψ D (x,y)=ψ(x)ψ(y) in, Let ψ(x) be the scaling function, and ψ(x) be the wavelet function. Step 4: Exchange the detail coefficients of fingerprint grayscale images collected from different locations or time periods within the same reference point range; if the set of detail coefficients after fingerprint decomposition is: CD=[CD H ,CD V ,CD D ] The entire set of coefficients can then be represented as: SC = [CD, CA]; (1) Exchange of detail coefficients of fingerprint grayscale images at different time periods: Assume that the fingerprint coefficient sets collected at different times t1 and t2 within the same reference point range are as follows: Then, the detail coefficients are interchanged: in, express Medium detail coefficient Replaced with (2) Exchange of detail coefficients of fingerprint grayscale images at different locations within the same reference point range: in, This represents the set of coefficients after the fingerprint minutia coefficients at position p1 are swapped with those at position p2. Step 5: The set of detail coefficients SC after the exchange in Step 4 is transformed by discrete wavelet inverse transform to obtain the final enhanced fingerprint grayscale image. Discrete wavelet inverse transform: Among them, f R (x,y) represents the final enhanced fingerprint grayscale image. The coefficients are the replaced two-dimensional detail coefficients, l∈{H,V,D}, and the other symbols are the same as those of the wavelet forward transform.

Images