Distributed spectrum sensing noise stripping method based on wavelet gradient

Abstract

The invention discloses a distributed spectrum sensing noise stripping method based on wavelet gradient, and belongs to the technical field of communication and artificial intelligence crossing. The method comprises the following steps: acquiring original spectrum data through a distributed cognitive radio node; performing sliding window segmentation, Hanning window weighting and power spectrum density calculation on the data; the preprocessed data are input into a differentiable wavelet noise separation layer, the layer serves as a learnable one-dimensional convolution kernel through a parameterized biorthogonal wavelet filter, and end-to-end trainable multi-resolution analysis is achieved in a deep learning framework; calculating the power spectrum density gradient of the high-frequency detail coefficient in real time, and quantifying the noise abrupt change intensity based on a central difference method; dynamically generating a gating weight, and selectively suppressing the wavelet coefficient of the high-gradient frequency point; inputting the processed data into a one-dimensional convolutional neural network with residual connection for classification; and results are fused by adopting a clustering type collaborative decision-making mechanism.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of wireless communication and artificial intelligence, specifically involving a distributed spectrum sensing noise stripping method based on wavelet gradient. It is particularly suitable for environments with extremely low signal-to-noise ratio (SNR≤-15dB). It achieves precise suppression of noise mutation frequency points through differentiable wavelet transform and gradient gating mechanism, while preserving the structural features of useful signals. Background Technology

[0002] With the in-depth development of cognitive radio technology, distributed spectrum sensing, as a core mechanism for achieving dynamic spectrum access, has shown great potential in improving spectrum utilization efficiency. This technology aims to accurately identify unoccupied frequency band holes by licensed users through multi-node collaborative acquisition and analysis of spectrum data, thereby supporting intelligent access for secondary users. However, in practical deployments, especially in complex electromagnetic environments with extremely low signal-to-noise ratios (SNR≤-15dB), existing sensing schemes face severe challenges. In such environments, white noise and target signals (such as OFDM modulated signals) highly overlap in the frequency domain, severely obscuring the effective features in the original I / Q sampling data and greatly weakening the reliability and robustness of the sensing system.

[0003] Traditional methods based on cooperative energy detection rely directly on raw sampled data for decision-making, lacking the ability to finely model noise structures. While emerging deep learning models (such as 1D-CNN) possess certain feature extraction capabilities, they are prone to misjudgments when noise and signal spectral characteristics are highly similar due to the lack of integration with prior physical layer knowledge. Experimental results show that at a -15dB SNR, the false alarm rate of such models reaches as high as 40.2%, far exceeding the 10% limit stipulated by 3GPP, severely restricting their feasibility for application in weak signal scenarios such as military communications and deep-sea exploration.

[0004] Existing noise suppression techniques also have significant limitations. Classical wavelet thresholding denoising methods employ fixed threshold strategies, which cannot adapt to dynamically changing noise environments and destroy signal phase information during denoising, leading to a significant decrease in detection rate. While end-to-end deep learning models can automatically learn features, the separation of noise suppression and feature extraction processes prevents joint optimization. Crucially, the non-differentiable nature of traditional wavelet denoising modules hinders end-to-end training with neural networks, making it difficult for models to distinguish noise abrupt changes from real signal edges (such as OFDM symbol boundaries). Therefore, there is an urgent need for a distributed spectrum sensing method that can deeply integrate physical priors, possess adaptive noise stripping capabilities, and support end-to-end optimization, in order to achieve both low false alarm rates and high detection rates under extremely low signal-to-noise ratios. Summary of the Invention

[0005] The purpose of this invention is to provide a distributed spectrum-sensing noise stripping method based on wavelet gradient, which can effectively solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The distributed spectrum-aware noise stripping method based on wavelet gradients includes the following specific steps:

[0008] Step 1: Collect raw spectrum data through distributed cognitive radio nodes. Each node is configured with a software-defined radio front end to receive signals in orthogonal frequency division multiplexing modulation. The sampling rate is set to 20 MHz, the center frequency covers the 400 MHz to 6 GHz band, and each sampling frame contains 1024 complex sampling points.

[0009] Step 2: Preprocess the raw spectrum data. Use a sliding window mechanism to segment the continuous sampled data. The window length is 256 sampling points and the overlap rate is 50%. Perform Hanning window weighting operation on each data segment to suppress spectral leakage and calculate its power spectral density.

[0010] Step 3: Input the preprocessed data into the differentiable wavelet noise separation layer for noise stripping. The differentiable wavelet noise separation layer is implemented in the deep learning framework using the Conv1d module by parameterizing the analysis filter of the biorthogonal wavelet into a learnable one-dimensional convolution kernel. The low-pass filter is initialized with the reconstructed low-pass filter coefficients of the biorthogonal wavelet 'bior3.5', the high-pass filter is derived from the low-pass filter through the QMF relation, downsampling is achieved through convolution with stride=2, and upsampling is completed by convolution after zero-placing. All operations support automatic differentiation. This layer performs 5-level discrete wavelet decomposition on the input signal, using biorthogonal wavelet basis functions with vanishing moment of 3 and support length of 9, retaining approximation coefficients and detail coefficients, and calculating the power spectral density gradient of the highest frequency detail coefficient D1. The gradient calculation adopts the central difference method, and the difference step size is 1 frequency unit.

[0011] Step 4: Dynamically generate gated weights based on the power spectral density gradient, assuming the D1 detail coefficient sequence is... Its corresponding power spectral density gradient vector After normalization, we get gnorm=||g||∞-1·∣g|; initial gating weights Calculated by element

[0012] winit[i]=σ(α·(gnorm[i]-β))

[0013] in

[0014] σ(x)=1 / (1+ex), α=10, β=0.5

[0015] Smoothing winit with a 3-point moving average yields wsmoooth[i] = (winit[i-1] + winit[i] + winit[i+1]) / 3 (padding the boundaries with zeros); the final suppressed coefficient d′[i] = d[i] × (1 - wsmooth[i]);

[0016] Step 5: Reconstruct the noise-removed data and recalculate the power spectral density. Input the data into a deep neural network for classification. This network adopts a one-dimensional convolutional neural network structure, which contains 3 convolutional layers and 2 fully connected layers. The first convolutional layer outputs a 64-channel feature map, and the third convolutional layer outputs a 256-channel feature map. The output of the first layer is upscaled to 256 channels through a 1×1 convolution and then added to the output of the third layer to form a residual connection. Each convolutional layer is followed by a batch normalization layer and a modified linear unit activation function. The number of neurons in the fully connected layer is 512 and 256, respectively. The output layer uses the softmax function to generate the classification probability of the spectral occupancy state.

[0017] Step 6: Distributed collaborative decision-making based on classification probability, using a clustered network topology. Each node in a cluster sends its local hard decision result to the cluster head node. Cluster head election is based on the node's remaining energy and signal quality indicators, including the local false alarm rate estimate or wavelet gradient stability index. The cluster head uses a hard decision fusion rule. When more than 60% of the nodes determine that the channel is occupied, they report "occupancy". The central fusion node applies the 60% threshold again to all cluster head reports for global decision-making.

[0018] Step 7: Monitor and optimize system performance in real time, and record the false alarm rate and detection rate for each sensing task. When the false alarm rate exceeds 15% or the detection rate is lower than 80% in 10 consecutive sensing tasks, trigger the parameter adjustment mechanism: if the false alarm rate is >15% and the detection rate is ≥80%, increase β by 0.05; if the detection rate is <80% and the false alarm rate is ≤15%, decrease β by 0.05; if both exceed the limit, decrease β by 0.05 and increase the learning rate from 0.001 to 0.005, and restore the original learning rate after 5 sensing cycles.

[0019] Preferably, in step 1, the distributed cognitive radio node adopts a hardware architecture combining a general-purpose software-defined radio peripheral and an embedded artificial intelligence computing platform. The software-defined radio front-end supports real-time bandwidth reconfiguration, with a maximum instantaneous bandwidth of 56 MHz and an analog-to-digital converter resolution of 14 bits. The embedded artificial intelligence computing platform is equipped with a 128-core graphics processor and a 4-core central processing unit, with a single node power consumption of no more than 10 watts. This hardware architecture supports parallel processing of wavelet decomposition and convolutional neural network inference with a power spectral density of 256 points.

[0020] Preferably, the preprocessing process in step 2 further includes DC offset compensation and gain calibration of the acquired raw spectrum data, and the use of a minimum mean square adaptive filter to eliminate the influence of frequency selective fading. The filter order is set to 16 and the convergence factor is 0.01.

[0021] Preferably, in the discrete wavelet transform decomposition process of the differentiable wavelet noise separation layer in step 3, the approximation coefficients and other detail coefficients D2 to D5 remain unchanged, and only the D1 detail coefficient is subjected to gating weight suppression.

[0022] Preferably, the 1×1 convolution upscaling operation of the deep neural network in step 5 does not include an activation function, ensuring the linear superposition characteristic of the residual branches.

[0023] Preferably, the wavelet gradient stability index in the signal quality index in step 6 is defined as the mean standard deviation of the power spectral density gradient vector in the most recent 5 sensing tasks. The smaller the standard deviation, the more stable the sensing environment.

[0024] Preferably, it also includes the visualization and storage of spectrum sensing results, storing the raw data, processing data, decision results and performance indicators of each sensing task into a time series database, supporting multi-dimensional queries by time range and geographical region, and the data retention period is 30 days.

[0025] Preferably, the method supports the detection of various modulation signals, including orthogonal frequency division multiplexing signals, single-carrier frequency division multiple access signals, and global mobile communication system signals. By configuring different signal detection thresholds and feature extraction parameters, adaptive spectrum sensing in heterogeneous network environments can be achieved.

[0026] Compared with existing technologies, this invention has the following advantages: By introducing a differentiable wavelet noise separation layer, it achieves deep integration of physical layer prior knowledge and deep learning models. This layer uses parameterized wavelet filters as learnable convolutional kernels, making the entire system end-to-end trainable. The dynamic gating mechanism based on power spectral density gradient can accurately distinguish between noise abrupt changes and signal edges, selectively suppressing high gradient frequencies without disrupting the OFDM subcarrier structure. The closed-loop adaptive optimization mechanism intelligently adjusts the gradient threshold according to the direction of performance index deviation, ensuring that the system maintains a low false alarm rate and a high detection rate even under unknown interference abrupt changes. Experimental results show that at a signal-to-noise ratio of -18dB, the false alarm rate of this invention is 8.7%, and the detection rate reaches 92.3%. Compared with the traditional wavelet + CNN cascaded system, the false alarm rate is reduced by 31.5%, and the detection rate is increased by 17.8%, achieving unexpected technical results. Attached Figure Description

[0027] Figure 1This is a schematic diagram of the overall technical architecture of the distributed spectrum sensing noise stripping method based on wavelet gradient proposed in this invention.

[0028] Figure 2 This is a schematic diagram of the core principle framework of the differentiable wavelet noise separation layer and gradient gating mechanism in this invention;

[0029] Figure 3 This is a flowchart illustrating the logical process of raw spectrum data preprocessing and power spectral density calculation in this invention.

[0030] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between deep neural network classification and distributed collaborative decision-making in this invention;

[0031] Figure 5 This is a logical flowchart of the system performance monitoring and end-to-end adaptive optimization in this invention; Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0033] Currently, in the interdisciplinary field of wireless communication and artificial intelligence, distributed spectrum sensing, as the core mechanism for dynamic spectrum access in cognitive radio, has shown great potential in improving spectrum utilization efficiency. This technology aims to accurately identify unoccupied frequency band holes by licensed users through multi-node collaborative acquisition and analysis of spectrum data, thereby supporting intelligent access for secondary users. However, in practical deployments, especially in complex electromagnetic environments with extremely low signal-to-noise ratios (SNR≤-15dB), existing sensing schemes face severe challenges. In such environments, white noise and target signals (such as orthogonal frequency division multiplexing modulation signals) highly overlap in the frequency domain, severely obscuring the effective features in the original I / Q sampling data and greatly weakening the reliability and robustness of the sensing system. To address the aforementioned technical issues, this invention proposes a distributed spectrum sensing noise stripping method based on wavelet gradients. By introducing a differentiable wavelet noise separation layer, it achieves deep integration of physical layer prior knowledge and deep learning models. A dynamic gating mechanism based on power spectral density gradients can accurately identify and suppress noise mutation frequency points while preserving useful signal features. The end-to-end differentiable architecture enables joint optimization of noise suppression strategies and classification tasks, significantly improving detection performance in extremely low signal-to-noise ratio environments. A distributed collaborative decision-making mechanism combined with hardware compatibility design ensures the reliability and practicality of the system in actual deployments. Ultimately, it maintains a high detection rate while keeping the false alarm rate at a low level, meeting the stringent requirements of spectrum sensing in complex electromagnetic environments. This method is applied to a distributed spectrum sensing noise stripping method based on wavelet gradients.

[0034] refer to Figure 1The overall technical architecture of the wavelet gradient-based distributed spectrum sensing noise stripping method proposed in this invention includes distributed cognitive radio nodes, a differentiable wavelet noise separation layer, a deep neural network classifier, a distributed collaborative decision-making module, and a system performance monitoring and optimization unit. These modules work collaboratively to form an end-to-end trainable and adaptively optimized spectrum sensing system. The following will provide a detailed engineering explanation of each technical aspect based on the seven core steps of this invention, in conjunction with the accompanying drawings.

[0035] In the aforementioned distributed spectrum sensing noise stripping method based on wavelet gradients, step 1 involves acquiring raw spectrum data through distributed cognitive radio nodes. Specifically, each distributed cognitive radio node is configured with a software-defined radio front-end, receiving signals using orthogonal frequency division multiplexing modulation (OFDM), with a sampling rate set to 20 MHz, a center frequency covering the 400 MHz to 6 GHz band, and each sampling frame containing 1024 complex sampling points. The distributed cognitive radio node adopts a hardware architecture combining general-purpose software radio peripherals and an embedded artificial intelligence computing platform. The software-defined radio front-end supports real-time bandwidth reconfiguration, with a maximum instantaneous bandwidth of 56 MHz and an analog-to-digital converter (ADC) resolution of 14 bits. The embedded artificial intelligence computing platform is equipped with a 128-core graphics processor and a 4-core CPU, with a single node power consumption not exceeding 10 watts. During data acquisition, the software-defined radio front-end first down-converts the RF signal to baseband, and then the 14-bit ADC samples it at a rate of 20 MHz, generating orthogonal complex sampling sequences in I and Q channels. Each sampling frame lasts 51.2 microseconds (1024 / 20 MHz), a time window sufficient to capture a complete orthogonal frequency division multiplexing symbol and its cyclic prefix. To ensure data synchronization among multiple nodes, each node has a built-in high-precision clock source and is periodically calibrated using a GPS timing signal, keeping the time synchronization error within 100 nanoseconds. The acquired raw spectrum data is temporarily stored as a complex array in the cache of the embedded artificial intelligence computing platform, awaiting subsequent preprocessing module calls.

[0036] In the aforementioned distributed spectrum-aware noise stripping method based on wavelet gradients, step 2 involves preprocessing the original spectrum data. Specifically, a sliding window mechanism is used to segment the continuously sampled data, with a window length of 256 sampling points and an overlap rate of 50%. A Hanning window weighting operation is performed on each segment to suppress spectral leakage, and its power spectral density is calculated. The preprocessing also includes DC offset compensation and gain calibration of the acquired original spectrum data, and the use of a minimum mean square adaptive filter to eliminate the influence of frequency-selective fading. The filter order is set to 16, and the convergence factor is 0.01.

[0037] refer to Figure 3The preprocessing logic begins with DC offset compensation: First, the mean of the real and imaginary parts of all complex sampling points within the current sampling frame is calculated. Then, this mean is subtracted from each sampling point to eliminate the DC component introduced by hardware non-ideals. Next, gain calibration is performed. Using a built-in calibration signal source, the gain of the receiving link is dynamically adjusted to ensure consistent amplitude output from different nodes under the same signal strength. Then, a 16th-order least mean square adaptive filter is activated. Its weight vector is updated in real time based on the input signal and the desired response (usually the delayed input signal). The convergence factor is set to 0.01 to balance convergence speed and steady-state error, effectively compensating for frequency-selective fading caused by multipath effects. After the above compensation is completed, a sliding window mechanism is initiated. The window slides across the continuous data stream in steps of 128 sampling points (256 × 50%), truncating 256 sampling points each time to form a data segment. For each data segment, a Hanning window function is applied for weighting, and its expression is...

[0038] w(n) = 0.5 - 0.5cos(2πn / 255)

[0039] Where n is an index from 0 to 255.

[0040] The weighted data segment is transformed to the frequency domain through a fast Fourier transform to obtain a complex spectrum of 256 points. Then, its power spectral density is calculated, which is the square of the modulus of the complex value at each frequency point. Finally, a real vector of length 256 is output as the input to the subsequent differentiable wavelet noise separation layer.

[0041] In the aforementioned distributed spectrum-aware noise stripping method based on wavelet gradients, step 3 involves inputting the preprocessed data into a differentiable wavelet noise separation layer for noise stripping. Specifically, the differentiable wavelet noise separation layer is implemented in the PyTorch framework using the Conv1d module, which parameterizes the analysis filter of the bioorthogonal wavelet into a learnable one-dimensional convolution kernel. The low-pass filter h0 is initialized with the reconstructed low-pass filter coefficients of the bioorthogonal wavelet 'bior3.5' [0.0378, -0.0955, ..., 0.3379], and the high-pass filter h1 is derived from h0 through QMF relationships. Downsampling is achieved using convolution with stride = 2, and upsampling is completed by convolution after zero-placing. All operations support automatic differentiation, thus forming an end-to-end trainable wavelet decomposition-reconstruction path. This layer performs a 5-layer discrete wavelet decomposition on the input signal, using a bioorthogonal wavelet basis function with a vanishing moment of 3 and a support length of 9. During the decomposition process, approximation coefficients and detail coefficients are preserved, and the power spectral density gradient of the highest frequency detail coefficient D1 is calculated. The gradient calculation adopts the central difference method. For the k-th element in the sequence, its gradient is g[k]=(p[k+1]-p[k-1]) / 2, where p is the power spectral density of D1, and the difference step size is 1 frequency unit.

[0042] refer to Figure 2 The core of the differentiable wavelet noise separation layer lies in its fully differentiable implementation. The input power spectral density vector is first treated as a one-dimensional signal and fed into a 5-layer discrete wavelet transform module. This module employs biorthogonal wavelet basis functions, with its analysis filter and synthesis filter having vanishing moment of 3 and support length of 9, respectively. This design ensures good time-frequency localization while preserving the phase information of the signal to the greatest extent. In each layer of decomposition, the signal is processed by low-pass and high-pass filters respectively, and then downsampled to obtain approximation coefficients and detail coefficients. After 5 layers of decomposition, the original signal is decomposed into one low-frequency approximation coefficient sequence (A5) and five high-frequency detail coefficient sequences (D1 to D5). This invention focuses on the highest frequency detail coefficient D1, as it best reflects the instantaneous changes in the signal, while the energy of noise in the high-frequency band is usually much higher than that of the useful signal. For the D1 sequence, the system calculates its power spectral density gradient. The gradient is calculated using the central difference method. For the k-th element in the sequence, its gradient g[k] = (p[k+1] - p[k-1]) / 2, where the difference step size is one frequency unit. This gradient value directly quantifies the drastic change in power spectral density at that frequency unit, i.e., the intensity of noise abrupt changes. Since the entire wavelet transform and gradient calculation process consists of differentiable convolutions and arithmetic operations, this layer can be seamlessly embedded into deep learning frameworks, supporting end-to-end backpropagation training.

[0043] In the aforementioned distributed spectrum-aware noise stripping method based on wavelet gradients, step 4 involves dynamically generating gate weights based on the power spectral density gradient, and adaptively suppressing noise abrupt frequency points through a gradient gating mechanism. Specifically, let the D1 detail coefficient sequence be... Its corresponding power spectral density gradient vector After normalization, we get gnorm=||g||∞-1·∣g|; initial gating weights Calculated by element

[0044] winit[i]=σ(α·(gnorm[i]-β))

[0045] Where σ(x)=1 / (1+ex), α=10 is the steepness parameter, and β=0.5 is the suppression threshold; after smoothing by a 3-point moving average filter, we get wsmooth[i]=(winit[i-1]+winit[i]+winit[i+1]) / 3 (zeros are padded at the boundaries); the final generated gating weights are multiplied point by point with the wavelet coefficients of the corresponding frequency points to achieve selective suppression, i.e., d′[i]=d[i]×(1-wsmooth[i]).

[0046] refer to Figure 2The execution flow of the gradient gating mechanism is as follows: First, the power spectral density gradient vector g calculated in step 3 is normalized. The normalization factor is the maximum absolute value of vector g, so that the normalized gradient value gnorm∈[0,1]. Next, gnorm[i] is input into the sigmoid function σ(α·(gnorm[i]-β)), where α=10 and β=0.5, thereby generating the initial gating weights. The design of this sigmoid function makes the gating weights approach 1 (no suppression) when the normalized gradient value is close to 0, and the gating weights rapidly decay to 0 (strong suppression) when the gradient value is large. In order to prevent the gating weights from jumping drastically due to a single gradient outlier, the system applies a 3-point moving average filter to winit, i.e., wsmooth[i]=(winit[i-1]+winit[i]+winit[i+1]) / 3 (zero padding at the boundary), thereby obtaining the smoothed final gating weights. Finally, the gate weight is multiplied point-by-point with the original wavelet detail coefficients D1 to obtain the suppressed detail coefficients d′. The approximation coefficient A5 and other detail coefficients D2 to D5 remain unchanged. Subsequently, A5, D1′, and D2 to D5 are reconstructed back into the time domain through a differentiable inverse discrete wavelet transform to obtain the clean signal after noise removal. The power spectral density of this signal is recalculated and used as the input to the deep neural network.

[0047] In the aforementioned distributed spectrum-aware noise stripping method based on wavelet gradients, step 5 involves inputting the noise-stripped data into a deep neural network for classification. Specifically, this network employs a one-dimensional convolutional neural network structure, comprising three convolutional layers and two fully connected layers. The first convolutional layer uses 64 one-dimensional convolutional kernels of size 7, outputting a 64-channel feature map; the second convolutional layer uses 128 convolutional kernels of size 5; and the third convolutional layer uses 256 convolutional kernels of size 3. Each convolutional operation is followed by a batch normalization layer to accelerate training and improve model stability, followed by a correction of the nonlinearity introduced by the linear unit activation function. After the third convolutional layer, a residual branch is introduced: the feature map output by the first convolutional layer... Channel upscaling is performed using a 1×1 convolution kernel (256 in total) to obtain... The 1×1 convolution has no activation function; then x1′ is connected to the output of the third convolutional layer. Element-wise addition, i.e., xout = x3 + x1′, is then fed into subsequent batch normalization and ReLU activation. The output of the convolutional layer is flattened into a one-dimensional vector and fed into the first fully connected layer, which contains 512 neurons, followed by batch normalization and corrected linear units; the second fully connected layer contains 256 neurons. Finally, the output layer contains two neurons, corresponding to the "channel idle" and "channel occupied" categories, respectively, and the output is converted into a probability distribution using the softmax function. This network is jointly trained with the front-end differentiable wavelet noise separation layer, using cross-entropy loss as the loss function, Adam as the optimizer, and an initial learning rate of 0.001.

[0048] In the aforementioned distributed spectrum sensing noise stripping method based on wavelet gradients, step 6 involves distributed collaborative decision-making based on classification probabilities. Specifically, a clustered network topology is adopted, dividing the sensing nodes into multiple clusters. Each cluster elects a cluster head node responsible for local fusion. The cluster head node then sends the fusion result to the central fusion node for global decision-making. The cluster head election is based on the node's remaining energy and signal quality indicators. The signal quality indicators include the local false alarm rate estimate or the wavelet gradient stability index. The wavelet gradient stability index is defined as the mean standard deviation of the power spectral density gradient vector in the last five sensing tasks.

[0049] refer to Figure 4 The distributed collaborative decision-making adopts a two-level fusion architecture. Within the local cluster, assuming there are N sensing nodes, each node makes a hard decision of "occupied" based on the probability output in step 5. If the probability of "channel occupancy" is greater than 0.5, it is "idle." These binary decision results are sent to the cluster head node via a short-range wireless link (such as ZigBee). The cluster head node adopts the same hard decision fusion rule as the global cluster: if more than 60% of the nodes in the cluster decide "occupied," the cluster head reports "occupied" to the central fusion node. The election of the cluster head is a dynamic process, comprehensively considering the node's remaining battery energy and its reported signal quality indicators (such as received signal strength indication or wavelet gradient stability), prioritizing nodes with sufficient energy and good signal quality to extend network lifetime and improve fusion reliability. The central fusion node collects the decisions of all cluster heads and applies the 60% threshold again for the final decision. This clustered design effectively reduces the communication load and computational complexity of the central node and improves the scalability of large-scale networks.

[0050] In the aforementioned distributed spectrum sensing noise stripping method based on wavelet gradients, step 7 involves real-time monitoring and optimization of system performance. Specifically, the false alarm rate and detection rate for each sensing task are recorded. When performance indicators deviate from a preset range, the parameters of the differentiable wavelet noise separation layer are automatically adjusted to achieve end-to-end adaptive optimization. The system performance monitoring includes real-time calculation of the false alarm rate and detection rate. When the false alarm rate exceeds 15% or the detection rate is below 80% in 10 consecutive sensing tasks, a parameter adjustment mechanism is triggered: if the false alarm rate is >15% and the detection rate is ≥80%, it is determined that the useful signal is over-suppressed, and the sigmoid offset β is increased by 0.05; if the detection rate is <80% and the false alarm rate is ≤15%, it is determined that the noise suppression is insufficient, and β is decreased by 0.05; if both exceed the limit simultaneously, β is preferentially reduced to improve the detection rate, and the learning rate is simultaneously increased from 0.001 to 0.005, initiating an online fine-tuning mode, which continues for 5 sensing cycles before restoring the original learning rate. In addition, it includes the visualization and storage of spectrum sensing results. The raw data, processing data, decision results and performance indicators of each sensing task are stored in the database. The database adopts a time-series database structure, supports multi-dimensional queries by time range and geographical region, and the data retention period is 30 days.

[0051] refer to Figure 5 The system performance monitoring unit continuously tracks two core metrics: false alarm rate (the probability of misclassifying an idle channel as occupied) and detection rate (the probability of correctly detecting an occupied channel). These metrics are calculated by comparing them with known ground truth values ​​(obtained during testing or calibration). The monitoring unit maintains a sliding window of length 10, recording the performance metrics of the most recent 10 sensing tasks. Once the average false alarm rate within the window exceeds 15% or the average detection rate falls below 80%, the system determines that a significant change has occurred in the current environment (such as a sudden change in noise characteristics or the emergence of new interference signals), triggering an adaptive optimization mechanism. This mechanism first fine-tunes the sigmoid function offset β (i.e., the gradient threshold) in the differentiable wavelet noise separation layer, for example, adjusting it from 0.5 to 0.45 to enhance noise suppression, or adjusting it to 0.55 to avoid over-suppressing useful signals. Simultaneously, the learning rate of the deep neural network is temporarily increased (e.g., from 0.001 to 0.005) to accelerate the model's adaptation to the new environment. The entire adjustment process is performed in the background and does not affect the system's real-time sensing function. All raw data, intermediate processing results (such as wavelet coefficients and gating weights), final decisions, and performance metrics are serialized and stored in a time-series-based database (such as InfluxDB). Each record is accompanied by a precise timestamp and node geographic location information, supporting efficient multidimensional queries and historical backtracking. The data retention period is 30 days.

[0052] To illustrate the application effects of this invention more specifically, a scenario is constructed: 20 distributed cognitive radio nodes as described in this invention are deployed in a military exercise area to monitor for enemy communication signals in the 400 MHz to 6 GHz frequency band. The environmental signal-to-noise ratio is -18 dB, with strong broadband white noise and sudden narrowband interference. Each node operates at a 20 MHz sampling rate, and the collected orthogonal frequency division multiplexing (OFDM) signals are fed into the processing flow of this invention. After preprocessing and differentiable wavelet noise stripping, the original noise-saturated signal spectrum becomes clear, and the unique subcarrier structure of the OFDM signal is revealed. Based on this clean spectrum, a deep neural network correctly determines that the channel is occupied with a probability of 92.3%. Each node sends its decision to the cluster head, and the fusion center, after receiving 13 (over 60%) "occupancy" decisions, confirms that the channel is unavailable. Throughout the process, the system recorded a false alarm rate of 8.7%, far below the 10% upper limit specified by 3GPP, fully verifying the superior performance of this invention under extremely low signal-to-noise ratios. Comparative experiments show that the traditional wavelet denoising + CNN scheme has a false alarm rate of 40.2% and a detection rate of 74.5% in the same scenario; the pure 1D-CNN scheme has a false alarm rate of 38.7% and a detection rate of 76.1%. Compared with the two, the present invention reduces the false alarm rate by more than 31.5% and improves the detection rate by more than 16.2%, achieving unexpected technical results. Furthermore, the method described in this invention supports the detection of various modulated signals, including orthogonal frequency division multiplexing (OFDM) signals, single-carrier frequency division multiple access (SCDMA) signals, and Global System for Mobile Communications (GSMA) signals. By configuring different signal detection thresholds and feature extraction parameters, adaptive spectrum sensing in heterogeneous network environments is achieved.

[0053] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A distributed spectrum-sensing noise stripping method based on wavelet gradient, characterized in that: The specific steps include the following: Step 1: Collect raw spectrum data through distributed cognitive radio nodes. Each node is configured with a software-defined radio front end to receive signals in orthogonal frequency division multiplexing modulation. The sampling rate is set to 20 MHz, the center frequency covers the 400 MHz to 6 GHz band, and each sampling frame contains 1024 complex sampling points. Step 2: Preprocess the raw spectrum data. Use a sliding window mechanism to segment the continuous sampled data. The window length is 256 sampling points and the overlap rate is 50%. Perform Hanning window weighting operation on each data segment to suppress spectral leakage and calculate its power spectral density. Step 3: Input the preprocessed data into the differentiable wavelet noise separation layer for noise stripping. The differentiable wavelet noise separation layer is implemented in the deep learning framework using the Conv1d module by parameterizing the analysis filter of the biorthogonal wavelet into a learnable one-dimensional convolution kernel. The low-pass filter is initialized with the reconstructed low-pass filter coefficients of the biorthogonal wavelet 'bior3.5', the high-pass filter is derived from the low-pass filter through the QMF relation, downsampling is achieved through convolution with stride=2, and upsampling is completed by convolution after zero-placing. All operations support automatic differentiation. This layer performs 5-level discrete wavelet decomposition on the input signal, using biorthogonal wavelet basis functions with vanishing moment of 3 and support length of 9, retaining approximation coefficients and detail coefficients, and calculating the power spectral density gradient of the highest frequency detail coefficient D1. The gradient calculation adopts the central difference method, and the difference step size is 1 frequency unit. Step 4: Dynamically generate gated weights based on the power spectral density gradient, assuming the D1 detail coefficient sequence is... Its corresponding power spectral density gradient vector After normalization, we get gnorm=||g||∞-1·∣g|; initial gating weights Calculated by element winit[i]=σ(α·(gnorm[i]-β)) in σ(x)=1 / (1+ex), α=10, β=0.

5. Perform a 3-point moving average smoothing on winit to get wsmooth[i]=(winit[i-1]+winit[i]+winit[i+1]) / 3 (zero padding at the boundaries); the final suppressed coefficient d′[i]=d[i]×(1-wsmooth[i]); Step 5: Reconstruct the noise-removed data and recalculate the power spectral density. Input the data into a deep neural network for classification. This network adopts a one-dimensional convolutional neural network structure, which contains 3 convolutional layers and 2 fully connected layers. The first convolutional layer outputs a 64-channel feature map, and the third convolutional layer outputs a 256-channel feature map. The output of the first layer is upscaled to 256 channels through a 1×1 convolution and then added to the output of the third layer to form a residual connection. Each convolutional layer is followed by a batch normalization layer and a modified linear unit activation function. The number of neurons in the fully connected layer is 512 and 256, respectively. The output layer uses the softmax function to generate the classification probability of the spectral occupancy state. Step 6: Distributed collaborative decision-making based on classification probability, using a clustered network topology. Each node in a cluster sends its local hard decision result to the cluster head node. Cluster head election is based on the node's remaining energy and signal quality indicators, including the local false alarm rate estimate or wavelet gradient stability index. The cluster head uses a hard decision fusion rule. When more than 60% of the nodes determine that the channel is occupied, they report "occupied". The central fusion node applies the 60% threshold again to the results reported by all cluster heads for global decision-making. Step 7: Monitor and optimize system performance in real time, and record the false alarm rate and detection rate for each sensing task. When the false alarm rate exceeds 15% or the detection rate is lower than 80% in 10 consecutive sensing tasks, trigger the parameter adjustment mechanism: if the false alarm rate is >15% and the detection rate is ≥80%, increase β by 0.05; if the detection rate is <80% and the false alarm rate is ≤15%, decrease β by 0.05; if both exceed the limit, decrease β by 0.05 and increase the learning rate from 0.001 to 0.005, and restore the original learning rate after 5 sensing cycles.

2. The distributed spectrum-aware noise stripping method based on wavelet gradient according to claim 1, characterized in that: The distributed cognitive radio node adopts a hardware architecture combining general-purpose software-defined radio peripherals and an embedded artificial intelligence computing platform. The software-defined radio front-end supports real-time bandwidth reconfiguration with a maximum instantaneous bandwidth of 56 MHz and an analog-to-digital converter resolution of 14 bits. The embedded artificial intelligence computing platform is equipped with a 128-core graphics processor and a 4-core central processing unit, with a single node power consumption of no more than 10 watts. This hardware architecture supports parallel processing of wavelet decomposition and convolutional neural network inference with 256-point power spectral density.

3. The distributed spectrum-aware noise stripping method based on wavelet gradient according to claim 1, characterized in that: The preprocessing process also includes DC offset compensation and gain calibration of the acquired raw spectrum data, and the use of a minimum mean square adaptive filter to eliminate the influence of frequency selective fading. The filter order is set to 16 and the convergence factor is 0.

01.

4. The distributed spectrum-aware noise stripping method based on wavelet gradient according to claim 1, characterized in that: During the discrete wavelet transform decomposition of the differentiable wavelet noise separation layer, the approximation coefficients and other detail coefficients D2 to D5 remain unchanged, and only the D1 detail coefficient is subjected to gating weight suppression.

5. The distributed spectrum-aware noise stripping method based on wavelet gradient according to claim 1, characterized in that: The 1×1 convolution upscaling operation of the deep neural network does not include an activation function, ensuring the linear superposition property of the residual branches.

6. The distributed spectrum-aware noise stripping method based on wavelet gradient according to claim 1, characterized in that: The wavelet gradient stability index in the signal quality index is defined as the mean standard deviation of the power spectral density gradient vector in the most recent 5 sensing tasks. The smaller the standard deviation, the more stable the sensing environment.

7. The distributed spectrum-aware noise stripping method based on wavelet gradient according to claim 1, characterized in that: The system performance monitoring also includes storing the raw data, processing data, decision results and performance indicators of each sensing task into a time-series database, supporting multi-dimensional queries by time range and geographical region, with a data retention period of 30 days.

8. The distributed spectrum-aware noise stripping method based on wavelet gradient according to claim 1, characterized in that: The method supports the detection of various modulation signals, including orthogonal frequency division multiplexing signals, single-carrier frequency division multiple access signals, and global mobile communication system signals. By configuring different signal detection thresholds and feature extraction parameters, adaptive spectrum sensing in heterogeneous network environments can be achieved.

Images