Kan-based differentiable channel knowledge map construction method

Abstract

The application provides a KAN-based differentiable channel knowledge map construction method, relates to the field of unmanned aerial vehicle communication, and comprises the following steps: S1, sampling base station received signal strength values at different positions in a target city environment task area and constructing a sampling data set; S2, performing KNN interpolation on the sampling data set to obtain a denser interpolation data set, so as to capture global channel characteristics; S3, combining the sampling data set and the interpolation data set to construct an enhanced data set; and S4, training the initialized KAN by using the enhanced data set, and predicting the base station received signal strength in a target range by using the trained KAN, so as to construct a differentiable channel knowledge map. The application can ensure differentiability and improve the construction precision of the channel knowledge map.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) communication, and in particular to a method for constructing a knowledge map based on KAN. Background Technology

[0002] With the profound evolution of mobile communication technology, communication scenarios are becoming increasingly complex, placing higher demands on the accuracy, real-time performance, and intelligence of channel awareness. Traditional communication networks rely on real-time Channel State Information (CSI) feedback to obtain current channel information. However, in dynamic and complex scenarios, real-time CSI suffers from high acquisition latency, easy obsolescence, high overhead, and difficulty in prediction. Channel Knowledge Map (CKM) is an emerging technology that integrates and correlates multi-dimensional information such as channel characteristics, environmental factors, and communication performance indicators to achieve environment-aware wireless communication. It uses a database with channel knowledge specific to a given location to facilitate or even avoid real-time CSI acquisition, thus solving a series of challenges faced by traditional communication networks that rely on real-time CSI feedback. How to effectively construct an accurate channel knowledge map based on limited user locations and corresponding channel data is a key issue in CKM applications.

[0003] Existing methods for constructing channel gain (CKM) include traditional methods and deep learning methods. Traditional methods, such as K-Nearest Neighbors (KNN) and Kriging, are based on interpolation techniques and require measuring the received signal strength (RSS) over a large area, resulting in high complexity and limited accuracy. Deep learning-based methods utilize advanced neural architectures to simulate complex channel patterns; for example, RadioUNet and RadioDiff can achieve accurate CKM reconstruction. However, most existing neural networks rely on discrete topologies and are not differentiable relative to location, making them unsuitable for gradient-based optimization methods. Multilayer Perceptrons (MLPs) can model channel gain as a continuous function; however, their simple architecture results in low reconstruction accuracy. Summary of the Invention

[0004] The main objective of this invention is to propose a KAN-based method for constructing a channel knowledge map that can improve the accuracy of channel knowledge map construction while ensuring differentiability.

[0005] This invention is achieved through the following technical solution:

[0006] The method for constructing a WeChat knowledge map based on KAN includes the following steps:

[0007] Step S1: Sample the received signal strength values ​​of base stations at different locations within the target urban environmental task area and construct a sampling dataset;

[0008] Step S2: Perform KNN interpolation on the sampled dataset to obtain a denser interpolated dataset in order to capture global channel features;

[0009] Step S3: Merge the sampled dataset and the interpolated dataset to construct the augmented dataset;

[0010] Step S4: Train the initialized KAN with the augmented dataset, and use the trained KAN to predict the signal strength received by base stations within the target range, so as to realize the construction of the WeChat knowledge map.

[0011] Furthermore, in step S1, the task area is... A rectangular urban area and These represent the maximum index values ​​of the task region grid centers in the x-axis and y-axis directions, respectively. The sampled dataset is represented as... ,in, This indicates the nth position sampled. , N s The number of sampled data, P is the base station transmit power. Let be the large-scale fading coefficient at the nth position. Let V be the Gaussian noise variance at the nth position.

[0012] Furthermore, in step S1, based on the sampling rate Perform sampling. , .

[0013] Furthermore, in step S2, the interpolation dataset is represented as , This indicates the total number of interpolation points. , To perform KNN interpolation target location The K nearest neighbors, For target location Its k-th nearest neighbor Euclidean distance.

[0014] Furthermore, in step S3, the augmented dataset is represented as The augmented dataset takes location coordinates as input and base station received signal strength as the target.

[0015] Furthermore, in step S4, initializing KAN specifically includes: targeting the augmented dataset. All samples in the dataset are analyzed based on their corresponding input position coordinates. The channel gain of the sample is calculated from the results after L KAN layer transformations. , This is the one-dimensional function matrix layer corresponding to the l-th KAN layer. , This indicates a connection between multiple KAN layers, the first... The j-th component of the output of a KAN layer is represented as , To make the first The activation function connecting the i-th input of the -1 KAN layer to the j-th output of the l-th KAN layer is parameterized as a k-th order B-spline, denoted as: , For the learnable coefficients that serve as control points for the spline shape, To define the number of intervals in the spline grid, Let m be the k-th order B-spline basis functions, in The k-order B-splines on each interval have There are k-1 continuous basis functions. Indicates the first -1 is the number of inputs to a KAN layer.

[0016] Furthermore, in step S4, the B-spline basis function is recursively defined using a stable Cox-de-Boor formula. When k=1, the B-spline basis function is expressed as follows: When k>1, the B-spline basis function is expressed as: Terms with zero factors are considered zero. It is a non-decreasing sequence of nodes.

[0017] Furthermore, in step S4, when training the initialized KAN, the loss function is set to... The KAN parameters are updated to minimize the loss function via gradient descent.

[0018] Furthermore, it also includes step S5: based on the trained KAN, according to the formula For different positions gradient The solution is performed to verify the differentiability of the constructed channel knowledge map, where, , , , .

[0019] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0020] This invention first samples the received signal strength values ​​of base stations at different locations within the target urban environment task area and constructs a sampling dataset. Then, it performs KNN interpolation on the sampling dataset to obtain a denser interpolated dataset to capture global channel features. Next, it merges the sampling dataset and the interpolated dataset to construct an augmented dataset. Finally, it trains the initialized KAN with the augmented dataset and uses the trained KAN to predict the received signal strength of base stations within the target range, thereby constructing a channel knowledge map. This overcomes the problems of non-differentiability of location and inapplicability to gradient optimization in traditional deep learning methods. The constructed channel knowledge map has a lower NMSE under various measurement sampling densities, which effectively improves the construction accuracy of the channel knowledge map. Attached Figure Description

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Figure 1 This is a flowchart of the present invention.

[0023] Figure 2 This is a schematic diagram comparing the NMSE of the channel knowledge maps constructed by the present invention and the comparative scheme at different sampling rates.

[0024] Figure 3 This diagram illustrates the comparison between the channel knowledge map construction using this invention and other benchmark methods.

[0025] Figure 4 This is a comparison diagram of the predicted channel knowledge map and the actual channel knowledge map obtained by the present invention at different RSS sampling rates. Detailed Implementation

[0026] The present invention will be further described below through specific embodiments.

[0027] like Figure 1 As shown, the method for constructing a WeChat knowledge map based on KAN includes the following steps:

[0028] Step S1: Sample the received signal strength values ​​of base stations at different locations within the target urban environmental task area and construct a sampling dataset;

[0029] The task area is A rectangular urban area and These represent the maximum index values ​​at the center of the task area grid in the x-axis and y-axis directions, respectively, based on the terminal device at different locations, with a sampling rate. Sampling is performed by sampling the received signal strength (RSS) of the base station based on the reported channel quality information. The number of sampled data is... The dataset formed by the sampled data is represented as ,in, This indicates the nth position sampled. , , This represents the normalized received signal power, where P is the base station transmit power. Let be the large-scale fading coefficient at the nth position. Let V be the Gaussian noise variance at the nth position.

[0030] In this embodiment, , The path loss, shadows, and reflections and diffractions from obstacles were modeled based on the 2.4 GHz frequency band using the ray tracing tool Wireless InSite.

[0031] Step S2: Perform KNN interpolation on the sampled dataset to obtain a denser interpolated dataset in order to capture global channel features;

[0032] Specifically, the interpolated dataset is represented as , This indicates the total number of interpolation points. , To perform KNN interpolation target location The K nearest neighbors, The normalized received signal power of the k-th nearest neighbor is... For target location Its k-th nearest neighbor Euclidean distance.

[0033] Step S3: Merge the sampled dataset and the interpolated dataset to construct the augmented dataset;

[0034] Specifically, the augmented dataset is represented as The augmented dataset takes location coordinates as input and base station received signal strength as the target.

[0035] Step S4: Train the initialized KAN with the augmented dataset, and use the trained KAN to predict the signal strength received by base stations within the target range, so as to realize the construction of a knowledge map.

[0036] The initialization of KAN (Kolmogorov-Arnold network) specifically includes: targeting the augmentation dataset. All samples in the dataset are processed according to their corresponding input coordinates. The channel gain of the sample is calculated from the results after L KAN layer transformations. , This is the one-dimensional function matrix layer corresponding to the l-th KAN layer. , This indicates a connection between multiple KAN layers, the first... The j-th component of the output of a KAN layer is represented as , To make the first The activation function connecting the i-th input of the -1 KAN layer to the j-th output of the l-th KAN layer is parameterized as a k-th order B-spline, denoted as: , For the learnable coefficients that serve as control points for the spline shape, To define the number of intervals in the spline grid, Let m be the k-th order B-spline basis functions, in The k-order B-splines on each interval have There are k basis functions, and the B-spline basis functions are continuous of order k-1. Indicates the first -1 is the number of inputs to a KAN layer, 0 ≤ m ≤ .

[0037] B-spline basis functions are recursively defined using stable Cox-de-Boor formulas. When k=1, the B-spline basis functions are expressed as follows: When k>1, the B-spline basis function is expressed as: Terms with zero factors are considered zero. It is a non-decreasing sequence of nodes. During computation, it will... In Replace with .

[0038] When training the initialized KAN, the loss function is set to The KAN parameters are updated to minimize the loss function via gradient descent.

[0039] In this embodiment, the KAN includes 2 input neurons, 3 hidden layers, and 1 output neuron.

[0040] Step S5: Based on the trained KAN, according to the formula... For different positions gradient The solution is performed to verify the differentiability of the constructed channel knowledge map, where, , , , , where terms with zero factors are considered zero.

[0041] Figure 2This is a schematic diagram comparing the NMSE (Normalized Mean Square Error) of the channel knowledge maps constructed by the present invention and the comparison scheme at different sampling rates. Figure 2 In the graph, the horizontal axis represents the sampling rate γ, with values ​​of [5%, 10%, 15%, 20%, 25%, 30%], and the vertical axis represents the NMSE of the generated channel knowledge map and the actual channel knowledge map. Figure 2 It can be seen that as the sampling rate γ increases, the channel knowledge map generation error of each scheme gradually decreases, but the NMSE of the present invention (K-Nearest Neighbor Enhanced KAN) remains the smallest and shows the most significant improvement at a lower sampling rate, highlighting its robustness in sparse data scenarios.

[0042] Figure 3 In the diagram, the sampling rates corresponding to each column from left to right are 5%, 10%, 20%, and 30%, respectively, and each row from top to bottom represents a different construction scheme. It can be seen that the result image corresponding to this invention has richer details and better sharpening effect. Especially when γ=30%, the results corresponding to other methods will show obvious artifacts or excessive smoothing. It can be seen that the NMSE of this invention is the smallest.

[0043] from Figure 4 As can be seen, as the sampling rate γ increases, the CKM gradient magnitude predicted by this invention gradually becomes more closely aligned with the original CKM details.

[0044] In this invention, the terms "first," "second," and "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. The use of terms such as "upper," "lower," "left," "right," "front," and "rear" to indicate orientation or positional relationships is based on the orientation or positional relationships shown in the accompanying drawings and is only for the convenience of describing the invention, not to indicate or imply that the device referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation on the scope of protection of this invention. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0045] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0046] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A method for constructing a WeChat knowledge map based on KAN, characterized by: Includes the following steps: Step S1: Sample the received signal strength values ​​of base stations at different locations within the target urban environmental task area and construct a sampling dataset; Step S2: Perform KNN interpolation on the sampled dataset to obtain a denser interpolated dataset in order to capture global channel features; Step S3: Merge the sampled dataset and the interpolated dataset to construct the augmented dataset; Step S4: Train the initialized KAN with the augmented dataset, and use the trained KAN to predict the signal strength received by base stations within the target range, so as to realize the construction of a knowledge map. In step S1, the task area is A rectangular urban area and These represent the maximum index values ​​of the task region grid centers in the x-axis and y-axis directions, respectively. The sampled dataset is represented as... ,in, This indicates the nth position sampled. , N s The number of sampled data, P is the base station transmit power. Let be the large-scale fading coefficient at the nth position. Let Variance be the Gaussian noise variance at the nth position; In step S2, the interpolation dataset is represented as follows: , This indicates the total number of interpolation points. , To perform KNN interpolation target location The K nearest neighbors, For target location Its k-th nearest neighbor Euclidean distance.

2. The method for constructing a WeChat knowledge map based on KAN according to claim 1, characterized in that: In step S1, based on the sampling rate Perform sampling. , .

3. The method for constructing a WeChat knowledge map based on KAN according to claim 1 or 2, characterized in that: In step S3, the augmented dataset is represented as follows: .

4. The method for constructing a WeChat knowledge map based on KAN according to claim 3, characterized in that: In step S4, initializing KAN specifically includes: targeting the augmented dataset. All samples in the dataset are analyzed based on their corresponding input position coordinates. The channel gain of the sample is calculated from the results after L KAN layer transformations. , This is the one-dimensional function matrix layer corresponding to the l-th KAN layer. , This indicates a connection between multiple KAN layers, the first... The j-th component of the output of a KAN layer is represented as , To make the first The activation function connecting the i-th input of the -1 KAN layer to the j-th output of the l-th KAN layer is parameterized as a k-th order B-spline, denoted as: , For the learnable coefficients that serve as control points for the spline shape, To define the number of intervals in the spline grid, Let m be the k-th order B-spline basis functions, in The k-order B-splines on each interval have There are k-1 continuous basis functions. Indicates the first -1 is the number of inputs to a KAN layer.

5. The method for constructing a WeChat knowledge map based on KAN according to claim 4, characterized in that: In step S4, the B-spline basis function is recursively defined using a stable Cox-de-Boor formula. When k=1, the B-spline basis function is expressed as follows: When k>1, the B-spline basis function is expressed as: Terms with zero factors are considered zero. It is a non-decreasing sequence of nodes.

6. The method for constructing a WeChat knowledge map based on KAN according to claim 5, characterized in that: In step S4, when training the initialized KAN, the loss function is set to... The KAN parameters are updated to minimize the loss function via gradient descent.

7. The method for constructing a WeChat knowledge map based on KAN according to claim 6, characterized in that: It also includes step S5: based on the trained KAN, according to the formula For different positions gradient The solution is performed to verify the differentiability of the constructed channel knowledge map, where, , , , .

Images