Modulation adaptive deep learning spectrum sensing method

Abstract

This invention discloses a modulation-adaptive deep learning spectrum sensing method, belonging to the field of wireless communication technology. This method constructs an adaptive convolutional gating mechanism, combines a lightweight LSTM to dynamically generate the convolutional kernel offset matrix, and introduces gradient-aware gating to achieve update / pass-through mode switching; simultaneously, a meta-learning optimization engine is run to periodically adjust the learning rate, forming a multi-level adaptive closed-loop system. A sparsity module is inserted after the third convolutional layer during the inference stage to reduce computational load. This invention solves the problems of poor generalization ability, insufficient real-time performance, and high computational overhead of fixed models under unknown modulation types, stably controlling the false detection rate to within 8%, and the processing latency to less than 52 milliseconds. While reducing the computational load by 40%, it only loses 1.5% accuracy, significantly improving the generalization ability, real-time performance, and energy efficiency balance of spectrum sensing.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a modulation-adaptive deep learning spectrum sensing method, which is particularly suitable for real-time spectrum sensing of unknown modulation signals in cognitive radio systems. Background Technology

[0002] With the in-depth development of cognitive radio technology, broadband spectrum sensing, as a key link in realizing dynamic spectrum sharing, directly determines the spectrum utilization efficiency and the reliability of communication systems. In complex and ever-changing wireless environments, deep learning methods have been widely introduced to improve sensing accuracy. Among them, models based on convolutional neural networks (CNNs) achieve automatic identification of spectrum occupancy status by processing short-time Fourier transform (STFT) spectra. However, such methods generally rely on fixed network architectures, such as ResNet or VGG. When faced with modulation types not covered during the training phase, their generalization ability drops significantly, making it difficult to meet the real-time sensing requirements of the highly dynamic spectrum environment in the 5G / 6G era.

[0003] Modulation-adaptive spectrum sensing has become one of the core research directions. This technology aims to enable the sensing system to dynamically adjust its feature extraction strategy based on the modulation characteristics of the input signal, thereby maintaining high detection accuracy even when unknown modulation types (such as 256QAM or 1024QAM) are present. Ideally, the system should have the ability to respond quickly to key features in the spectrum, such as phase jumps and energy abrupt changes, and complete local optimization of model parameters within milliseconds to avoid false detections or missed detections due to feature distortion.

[0004] Existing technologies suffer from multiple structural flaws in addressing the aforementioned challenges. Traditional CNN models, due to their fixed convolutional kernel weights, cannot adapt to the spectral distribution characteristics of newly modulated signals, leading to a false detection rate soaring to 38% in scenarios such as 256QAM. While some solutions attempt to expand the training set through data augmentation, these rely on prior knowledge and cannot adapt to sudden signals in real time. Network fine-tuning strategies, while offering some flexibility, suffer from excessive computational overhead, making it difficult to meet the 10ms-level real-time requirements. Methods incorporating attention mechanisms optimize feature selection but do not address the dynamic reconstruction of the convolutional kernel itself, resulting in a response latency exceeding 15ms to spectral changes. The root cause of these problems lies in the lack of a lightweight, fast, and accurate online optimization mechanism for convolutional kernel parameters, causing a sharp decline in the reliability of existing systems in high-density base station deployments and high-speed mobile scenarios. Therefore, there is an urgent need for a deep learning spectrum sensing method capable of millisecond-level adaptive adjustment on edge devices to overcome the generalization bottleneck caused by fixed model structures. Summary of the Invention

[0005] The purpose of this invention is to provide a modulation-adaptive deep learning spectrum sensing method that can effectively solve the problems mentioned in the background art, such as the significant decrease in generalization ability, insufficient real-time performance, and excessive computational overhead when the fixed model structure is faced with unknown modulation types.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a modulation-adaptive deep learning spectrum sensing method, comprising the following specific steps: Step 1: Acquire spectrum data. Receive wireless signals through the radio frequency front-end, perform short-time Fourier transform processing on the received signals, and generate a spectrum time-frequency diagram.

[0007] This step uses the Hanning window function for time-frequency analysis, with a window length of 256 sampling points, an overlap rate of 50%, a frequency resolution of 15.625 kHz, and a time resolution of 64 microseconds. The specific processing steps are as follows: First, the received time-domain signal is divided into overlapping time-window segments. A Hanning window function is applied to each segment to reduce spectral leakage. Then, a Fast Fourier Transform (FFT) is performed to convert the time-domain signal into a frequency-domain representation. The key formula for calculating the Short-Time Fourier Transform (SFT) is:

[0008] in, For time-domain signals, For the Hanning window function, For the length of the window, For time frame indexing, The frequency index is then used. The power spectral density is subsequently calculated and subjected to logarithmic compression and normalization to generate a 128×128 pixel two-dimensional time-frequency spectrum. This time-frequency spectrum uses grayscale values ​​to represent the energy intensity of different time-frequency units, with the values ​​normalized to the 0-1 range, serving as the basic input data for subsequent feature extraction.

[0009] Step 2: Construct an adaptive convolutional gating mechanism, reconstruct the underlying logic of the convolutional neural network feature extraction layer based on meta-learning driven approach, and realize online dynamic optimization of convolutional kernel parameters.

[0010] This step addresses the generalization collapse problem caused by fixed convolution kernels and unknown modulation types. It reconstructs the underlying logic of the CNN feature extraction layer, enabling dynamic adjustment of the convolution kernel parameters based on the characteristics of the input signal. Specifically, the standard convolution operation is replaced with a parameterizable gated convolution operation: the original convolution kernel weights serve as the base template, while the effective weights participating in the convolution operation are obtained by adding the base template to a dynamically generated offset matrix. The core formula for this dynamic adjustment is:

[0011] in, The base convolutional kernel weights for pre-training, The offset matrix is ​​dynamically generated. The gating coefficients are used. This design transforms the feature extraction layer from a static mapping function into a dynamic system driven by external control signals. When the modulation type of the input signal changes, the system can automatically adjust the convolution kernel parameters to adapt to the new signal feature distribution, thereby significantly improving the ability to identify unknown modulation types.

[0012] Step 3: Perform dynamic weight generation, use a lightweight long short-term memory network to analyze the spectral change features in the time-frequency spectrum in real time, and output the convolution kernel offset matrix.

[0013] This step uses a lightweight LSTM network with 16 hidden layer units as the dynamic weight generator. This network has a single-layer, unidirectional structure. The input is a 128-dimensional sequence (128 frequency points per frame) unfolded from a 128×128 frequency spectrum over time frames, which is normalized and then fed into the LSTM. The LSTM hidden state... Ultimately, it is achieved through a fully connected layer. The hidden state at the last time step is mapped to a 9-dimensional vector, and then reshaped into a 3×3 convolution kernel offset matrix. This fully connected layer is jointly trained with LSTM, and the loss function is... To balance classification accuracy and offset sparsity, this network analyzes the phase transitions and energy abrupt changes of signals in the time-frequency domain to capture key criteria for different modulation types. For example, higher-order QAM signals typically exhibit denser energy distributions and more complex phase change patterns. The processing latency of the entire dynamic weight generation process is strictly controlled within 2 milliseconds, ensuring the system's real-time response capability.

[0014] Step 4: Implement gradient-aware gating. Construct a binary switch matrix based on a specific activation function. When the spectral energy distribution deviates from the threshold set in the training set, trigger the update mode; otherwise, maintain the pass-through mode.

[0015] This step introduces an intelligent decision-making mechanism to avoid unnecessary computational overhead. Specifically, the system first calculates the global energy mean of the current frequency spectrum and compares it with a baseline threshold derived from the training set. An sigmoid growth curve function is used as the activation function.

[0016] in, This represents the relative deviation between the current average energy and the baseline threshold. This is the steepness coefficient. The offset center is used. When the spectral energy distribution deviates from the training set threshold by ±15%, the system automatically triggers the update mode; otherwise, if the deviation is within the threshold range, the pass-through mode is maintained, and the original fixed convolutional kernel is used directly. This binary switching mechanism has a response time of 1.5 milliseconds, ensuring a fast response to spectral changes while avoiding frequent updates to the convolutional kernel parameters in stable signal environments, thus balancing system performance and computational overhead.

[0017] Step 5: Run the meta-learning optimization engine, which periodically estimates the gradient direction online based on the cached samples using the model-independent meta-learning algorithm, and dynamically adjusts the learning rate parameter.

[0018] This step serves as the long-term memory and optimization hub of the entire adaptive mechanism, executing an optimization cycle every 50 milliseconds. The buffer adopts a circular queue structure with a maximum capacity of 500 samples, each labeled with its modulation type or marked as "unknown". Every 50 milliseconds, the meta-learning optimization engine randomly samples 4 tasks from this buffer. Each task consists of 5 support samples (containing at least 2 different modulation types, one of which can be "unknown") and 10 query samples. The support set is used for inner loop updates. The query set is used for outer loop computation. If the proportion of "unknown" samples in the cache exceeds 30%, tasks containing "unknown" samples will be sampled first. Based on the model-independent meta-learning algorithm, the engine estimates the optimal initial parameter direction that enables the model to converge quickly on new tasks by simulating gradient descent steps on the support set. The key meta-learning optimization formula is:

[0019]

[0020] in, For model parameters, For the inner loop learning rate, The loss function is defined as follows. Based on this, the engine dynamically adjusts the learning rate parameter, controlling it within the range of 0.0005 to 0.002, effectively avoiding overfitting. Furthermore, a gradient pruning mechanism is introduced, pruning the gradient when its norm exceeds a set threshold of 2.0, and employing a stochastic gradient descent algorithm with a momentum coefficient of 0.9 to ensure the stability of the optimization process.

[0021] Step 6: Complete spectrum status recognition. Use the optimized convolutional neural network to classify and determine the processed spectrum features, and output the spectrum occupancy status detection results.

[0022] This step employs an optimized network with three convolutional layers and two fully connected layers for final classification. All convolutional layers use 3×3 kernels with a stride of 1 and employ the same padding method to maintain the feature map size. Before the fully connected layers, a feature sparsification module is inserted: the feature map output from the third convolutional layer is then subjected to global average pooling to obtain a 128-dimensional vector. This vector serves as the input to the sparsification module; during the inference phase, it is used for... The 51 elements with the smallest absolute values ​​(i.e., 40%) are set to zero, while the rest are retained. This operation is not involved in gradient calculation during training; it is only enabled during deployment to reduce the number of multiply-accumulate operations in subsequent fully connected layers. The number of neurons in the fully connected layers is 128 and 64, respectively. Finally, the classification probability of the spectral occupancy state is output through the Softmax function.

[0023] in, For the output value of class k, This represents the total number of categories. The entire recognition process is performed on an adaptively adjusted network, whose input is a feature map processed by a gating mechanism, thus exhibiting strong robustness to unknown modulation types. Testing showed that this sparsity operation resulted in only a 1.5% loss in accuracy, while reducing computational load by 40%.

[0024] Preferably, in step 1, the short-time Fourier transform processing uses the Hanning window function with a window length of 256 sampling points, an overlap rate of 50%, a frequency resolution of 15.625 kHz, and a time resolution of 64 microseconds, generating a two-dimensional spectrum time-frequency map with a size of 128 x 128 pixels. This parameter setting achieves a good balance between time and frequency resolution, effectively capturing common modulation signal characteristics in modern communication systems.

[0025] Preferably, the adaptive convolutional gating mechanism in step 2 addresses the generalization collapse problem caused by fixed convolutional kernels for unknown modulation types. Real-world testing confirmed a performance gap where the false detection rate increased from 5% to 38%. This mechanism is specifically designed to maintain the adaptability of the feature extractor even when the training set does not cover all modulation types. By dynamically adjusting the convolutional kernel parameters, the network can extract the most discriminative features for different modulation types, thereby significantly improving the system's generalization ability.

[0026] Preferably, in step 3, the lightweight long short-term memory network has 16 hidden layer units, analyzes phase transitions and energy mutations in the time-frequency spectrum in real time, outputs a convolutional kernel offset of 3x3, and has a processing latency of no more than 2 milliseconds. This lightweight design ensures the real-time performance of the system on edge devices while retaining sufficient expressive power to capture key signal features.

[0027] Preferably, the specific activation function in step 4 is an S-shaped growth curve function. When the spectral energy distribution deviates from the training set threshold by ±15%, an update mode is triggered. The response time of the binary switch matrix is ​​1.5 milliseconds, ensuring a rapid response to spectral changes. This threshold setting has been verified through extensive experiments, achieving an optimal balance between avoiding frequent updates and timely response to real signal changes.

[0028] Preferably, in step 5, the model-independent meta-learning algorithm is optimized every 50 milliseconds, estimating the gradient direction online based on cached support set samples. The dynamically adjusted learning rate parameter ranges from 0.0005 to 0.002, effectively avoiding overfitting. This periodic optimization strategy enables the system to continuously adapt to long-term environmental changes without forgetting previously learned knowledge.

[0029] Preferably, the meta-learning optimization engine also includes a gradient pruning mechanism, which prunes the gradient when the gradient norm exceeds a set threshold of 2.0, and employs a stochastic gradient descent algorithm with a momentum coefficient of 0.9 to ensure the stability of the optimization process. These technical details guarantee the robustness of the system in various complex wireless environments.

[0030] Preferably, the optimized convolutional neural network in step 6 comprises 3 convolutional layers and 2 fully connected layers, with a convolutional kernel size of 3x3, a stride of 1, and the same padding method. The number of neurons in the fully connected layers is 128 and 64, respectively. Finally, the classification probability of the spectral occupancy state is output through the Softmax function. This network structure achieves a good balance between recognition accuracy and computational complexity.

[0031] Preferably, the method also includes a feature sparsification step, inserting a sparsification module before the fully connected layer of the convolutional neural network. By setting a sparsity rate of 40%, the computational load is reduced, and the accuracy loss is only 1.5% according to tests. This technology significantly improves the deployment feasibility of the system on resource-constrained edge devices.

[0032] Preferably, the method is deployed on an edge computing device, employing a digital signal processor supporting floating-point operations with a clock frequency of no less than 1.2 GHz and a memory capacity of no less than 2 gigabytes, ensuring that the overall processing latency is controlled within 52 milliseconds. This hardware configuration meets the stringent requirements of 5G / 6G systems for real-time spectrum sensing.

[0033] Preferably, the system also includes an online performance monitoring module. This module uses a sliding window to analyze the most recent 200 spectrum detection results. If the proportion of those results judged as "occupied" but subsequently verified as idle (through collaborative sensing or channel idle detection) exceeds 8%, a recalibration is initiated. The system automatically caches the most recent 50 such false detection samples and manually or through an auxiliary channel labels their true status. Subsequently, these samples are used to fine-tune the last two fully connected layers in three rounds, with a fixed learning rate of 0.001 and a batch size of 10 per round. This module acts as a "health monitor" for the system, ensuring the long-term reliability of the system.

[0034] Preferably, the incremental learning employs an elastic weight consolidation algorithm with an importance weight decay coefficient set to 0.95. Each update adjusts only the parameters of the last two fully connected layers, ensuring a balance between knowledge retention and adaptability. This strategy effectively prevents catastrophic forgetting, enabling the system to continuously learn new knowledge without losing existing capabilities.

[0035] Preferably, the method is validated in the 5G New Radio band with a center frequency of 3.5 GHz and a bandwidth of 100 MHz, supporting various modulation formats for orthogonal frequency division multiplexing (OFDM) signals, including orthogonal phase shift keying (QPSK) and orthogonal amplitude modulation (AEM). Experimental results show that this method can effectively handle various modulation signals in modern communication systems.

[0036] Preferably, the interface also includes a visualization of the spectrum sensing results, displaying the spectrum occupancy status in the form of a heatmap, with different colors representing different signal strength levels, and marking detected abnormal areas and confidence levels. This interface provides intuitive decision support for network operations and maintenance personnel.

[0037] Compared with existing technologies, this invention has the following advantages: Through a meta-learning-driven adaptive convolutional gating mechanism, online dynamic optimization of convolutional kernel parameters is achieved, significantly improving the model's generalization ability to unknown modulation types and keeping the false detection rate stably below 8%, a reduction of 30.7 percentage points compared to traditional solutions. Simultaneously, through lightweight design and optimized algorithms, processing latency is controlled within 52 milliseconds, meeting the requirements of real-time spectrum sensing. Combined with feature sparsity technology, the computational load is reduced by 40% while incurring only a 1.5% accuracy loss, achieving an optimized balance between performance and efficiency. More importantly, gradient-aware gating, LSTM dynamic weight generation, and the meta-learning optimization engine form a synergistic closed loop, producing unexpected technical effects: the gating mechanism avoids invalid LSTM calls, the lightweight nature of LSTM allows it to be embedded in edge devices, meta-learning prevents system degradation over time, and the overall system maintains high robustness even under unknown modulation types. Attached Figure Description

[0038] Figure 1This is a schematic diagram of the overall technical architecture of the modulation adaptive deep learning spectrum sensing method proposed in this invention. Figure 2 This is a schematic diagram illustrating the core principle framework of the adaptive convolutional gating mechanism in this invention; Figure 3 This is a logical flowchart of the spectrum time-frequency diagram generation and dynamic weight generation stages in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between gradient-aware gating and meta-learning optimization engine in this invention; Figure 5 This is a logical flow diagram of the spectrum state identification and online performance monitoring module in this invention; Detailed Implementation

[0039] Example 1 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.

[0040] Reference Appendix Figure 1 The overall technical architecture of the modulation-adaptive deep learning spectrum sensing method proposed in this invention includes an RF front-end module, a spectrum time-frequency graph generation unit, an adaptive convolutional gating mechanism, a dynamic weight generation module, a gradient-aware gating unit, a meta-learning optimization engine, a spectrum state recognition module, and an online performance monitoring module. The entire system is deployed on an edge computing device, using a digital signal processor that supports floating-point operations with a main frequency of no less than 1.2 GHz and a memory capacity of no less than 2 gigabytes.

[0041] In step 1, the RF front-end captures a wireless signal with a center frequency of 3.5 GHz and a bandwidth of 100 MHz. This signal contains various modulation types under the orthogonal frequency division multiplexing (OFDM) format. The signal is sent to the spectrum time-frequency diagram generation unit for short-time Fourier transform processing. This processing uses a Hanning window function with a window length of 256 sampling points and an overlap rate of 50%, resulting in a time resolution of 64 microseconds and a frequency resolution of 15.625 kHz. After this processing, the original time-domain signal is converted into a 128×128 pixel two-dimensional spectrum time-frequency diagram, where the grayscale value of each pixel corresponds to the energy intensity of a specific time-frequency unit.

[0042] In step 2, the adaptive convolutional gating mechanism addresses the generalization collapse problem caused by the fixed convolutional kernel for unknown modulation types. Experimental data confirms that when the input signal modulation type changes from 64QAM in the training set to the uncovered 256QAM, the false detection rate of the traditional fixed model increases sharply from 5% to 38%. To bridge this performance gap, the adaptive convolutional gating mechanism designed in this invention replaces the standard convolution operation with a parameterizable gated convolution operation, using the formula... This mechanism enables online dynamic optimization of convolutional kernel parameters. It transforms the feature extraction layer of a convolutional neural network from a static mapping function into a dynamic system, allowing for real-time adjustment of the feature extraction strategy based on the characteristics of the input signal.

[0043] In step 3, the dynamic weight generation module consists of a lightweight LSTM network with 16 hidden layer units. This network is a single-layer, unidirectional structure. The input is a 128-dimensional sequence (128 frequency points per frame) unfolded from a 128×128 frequency spectrum over time frames, which is then normalized and fed into the LSTM. The LSTM hidden state... Ultimately, it is achieved through a fully connected layer. The hidden state at the last time step is mapped to a 9-dimensional vector, and then reshaped into a 3×3 convolution kernel offset matrix. This fully connected layer is jointly trained with LSTM, and the loss function is... To balance classification accuracy and offset sparsity, this network takes a 128×128 pixel spectral time-frequency map as input and extracts contextual information through methods such as sliding windowing or global pooling, focusing on capturing phase transitions and energy abrupt changes in the spectrum. These features are key criteria for distinguishing different modulation types; for example, higher-order QAM signals typically exhibit denser energy distributions and more complex phase change patterns. After training, the LSTM network encodes these high-level semantic features into low-dimensional vectors and ultimately decodes and outputs a 3×3 convolutional kernel offset matrix.

[0044] In step 4, the gradient-aware gating unit intelligently decides whether to enable dynamic offset. This unit first calculates the global energy mean of the current spectral time-frequency plot and compares it with a baseline threshold statistically derived from the training set. An S-shaped growth curve function is used. As the activation function, where , When the spectral energy distribution deviates from the training set threshold by ±15%, an update mode is triggered; otherwise, if the deviation is within the threshold range, the pass-through mode is maintained, and the original fixed convolutional kernel is used directly. The response time of this binary switch matrix is ​​1.5 milliseconds, ensuring a fast response to spectral changes while avoiding unnecessary computational overhead.

[0045] In step 5, the meta-learning optimization engine executes an optimization cycle every 50 milliseconds. The buffer adopts a circular queue structure with a maximum capacity of 500 samples, each labeled with its modulation type or marked as "unknown". Every 50 milliseconds, the meta-learning optimization engine randomly samples 4 tasks from this buffer. Each task consists of 5 support samples (containing at least 2 different modulation types, one of which can be "unknown") and 10 query samples. The support set is used for inner loop updates. The query set is used for outer loop computation. If the proportion of "unknown" samples in the cache exceeds 30%, tasks containing "unknown" samples will be sampled first. Based on the model-independent meta-learning algorithm, the engine uses key formulas... and Parameter optimization is performed. Based on this, the engine dynamically adjusts the learning rate parameter, which is constrained to a range of 0.0005 to 0.002.

[0046] In step 6, the optimized convolutional neural network consists of 3 convolutional layers and 2 fully connected layers. All convolutional layers have a kernel size of 3×3, a stride of 1, and the same padding method. A batch normalization layer and a ReLU activation function are applied after each convolutional layer. Before the fully connected layers, a feature sparsity processing module is inserted: the feature map output from the third convolutional layer is global average pooled to obtain a 128-dimensional vector. This vector serves as the input to the sparsification module; during the inference phase, it is used for... The 51 elements with the smallest absolute values ​​(40%) are set to zero, while the rest are retained. This operation is not involved in gradient calculation during training; it is only enabled during deployment to reduce the number of multiply-accumulate operations in subsequent fully connected layers. The number of neurons in the fully connected layers is 128 and 64 respectively, and finally, the softmax function is applied. Output the classification probability of the spectrum occupancy state.

[0047] In a specific application scenario, when an unknown signal modulated with 1024QAM suddenly appears in the environment, the time-frequency spectrum generated in step 1 will show an abnormally high-density energy distribution. The gradient-aware gating unit in step 4 detects that the energy distribution deviates from the baseline threshold by more than 15% and immediately triggers an update mode. The lightweight LSTM network in step 3 then analyzes the abnormal spectrum and outputs a targeted 3×3 convolutional kernel offset matrix within 2 milliseconds. This offset is applied by the adaptive convolutional gating mechanism in step 2, dynamically adjusting the weights of the first-layer convolutional kernel to more effectively extract the subtle features of the 1024QAM signal. After optimization, the network successfully identifies the signal as "occupied" in step 6, with a false detection rate of only 7.2%, far lower than the 38% of the traditional fixed model.

[0048] 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 modulation-adaptive deep learning spectrum sensing method, characterized in that: The specific steps include the following: Step 1: Acquire spectrum data. Receive wireless signals through the radio frequency front-end, perform short-time Fourier transform on the received signals, and generate a spectrum time-frequency diagram. Specifically, this includes using the Hanning window function with a window length of 256 sampling points, an overlap rate of 50%, a frequency resolution of 15.625 kHz, a time resolution of 64 microseconds, and generating a two-dimensional spectrum time-frequency diagram with a size of 128 x 128 pixels. Step 2: Construct an adaptive convolutional gating mechanism. Based on a meta-learning-driven approach, reconstruct the underlying logic of the convolutional neural network feature extraction layer to achieve online dynamic optimization of the convolutional kernel parameters. Specifically, this includes using formulas... To achieve online dynamic optimization of convolution kernel parameters, where Based on the convolutional kernel weights, The offset matrix is ​​dynamically generated; Step 3: Perform dynamic weight generation. Use a lightweight long short-term memory network to analyze the spectral change features in the spectrum time-frequency graph in real time and output the convolution kernel offset matrix. The lightweight long short-term memory network is a single-layer unidirectional structure. The input is a 128-dimensional sequence of the 128×128 spectrum time-frequency graph unfolded by time frame. After normalization, it is fed into LSTM. The number of hidden layer units is 16. Finally, a fully connected layer maps the hidden state of the last time step to a 9-dimensional vector and reshapes it into a 3×3 convolution kernel offset matrix. The processing delay does not exceed 2 milliseconds. Step 4: Implement gradient-aware gating. Construct a binary switch matrix based on the S-shaped growth curve function. When the spectral energy distribution deviates from the training set threshold by ±15%, trigger the update mode; otherwise, maintain the pass-through mode. This specifically includes using an activation function. ,in This represents the relative deviation between the current average energy and the baseline threshold. This is the steepness coefficient. With the offset center, the response time of the binary switch matrix is ​​1.5 milliseconds; Step 5: Run the meta-learning optimization engine. Based on the model-independent meta-learning algorithm, the gradient direction is estimated online periodically according to the cached samples, and the learning rate parameter is dynamically adjusted. Specifically, the cache adopts a circular queue structure with a maximum capacity of 500 samples. Every 50 milliseconds, 4 tasks are randomly sampled from it. Each task consists of 5 support samples and 10 query samples. The learning rate is adjusted using the formula... and The learning rate parameter is dynamically adjusted to a range of 0.0005 to 0.

002. Step 6: Complete spectrum state recognition. The processed spectrum features are classified and judged using an optimized convolutional neural network (CNN), outputting the spectrum occupancy status detection result. The optimized CNN contains 3 convolutional layers and 2 fully connected layers. The feature map output from the third convolutional layer is global average pooled to obtain a 128-dimensional vector. During the inference phase, the 40% of elements with the smallest absolute values ​​are set to zero before being input into the fully connected layer. Finally, the vector is processed by the Softmax function. Output the classification probability of the spectrum occupancy state.

2. The modulation-adaptive deep learning spectrum sensing method according to claim 1, characterized in that: In step 1, the short-time Fourier transform processing uses the Hanning window function with a window length of 256 sampling points, an overlap rate of 50%, a frequency resolution of 15.625 kHz, a time resolution of 64 microseconds, and generates a two-dimensional spectrum time-frequency map with a size of 128 by 128 pixels.

3. The modulation-adaptive deep learning spectrum sensing method according to claim 1, characterized in that: In step 3, the fully connected layers of the lightweight long short-term memory network are jointly trained with the LSTM, and the loss function is... To balance classification accuracy and offset sparsity.

4. The modulation-adaptive deep learning spectrum sensing method according to claim 1, characterized in that: The steepness coefficient of the S-shaped growth curve function in step 4 Offset center .

5. The modulation-adaptive deep learning spectrum sensing method according to claim 1, characterized in that: In step 5, the supported samples contain at least two different modulation types, one of which can be marked as "unknown". If the proportion of "unknown" samples in the cache exceeds 30%, then the tasks containing "unknown" samples will be sampled first.

6. The modulation-adaptive deep learning spectrum sensing method according to claim 5, characterized in that: The meta-learning optimization engine also includes a gradient pruning mechanism, which prunes the gradient when the gradient norm exceeds a set threshold of 2.0, and uses a stochastic gradient descent algorithm with a momentum coefficient of 0.

9.

7. The modulation-adaptive deep learning spectrum sensing method according to claim 1, characterized in that: In step 6, the convolution kernel size is 3 x 3, the stride is 1, the padding method is the same, and the number of neurons in the fully connected layer is 128 and 64 respectively.

8. The modulation-adaptive deep learning spectrum sensing method according to claim 1, characterized in that: It also includes an online performance monitoring module that uses a sliding window to count the results of the last 200 spectrum detections. If the proportion of those detected as "occupied" but subsequently verified as idle exceeds 8%, a recalibration is initiated. The system automatically caches the last 50 such false detection samples and marks their true status through an auxiliary channel. These samples are then used to fine-tune the last two fully connected layers in three rounds, with a fixed learning rate of 0.001 and a batch size of 10 per round.

9. The modulation-adaptive deep learning spectrum sensing method according to claim 8, characterized in that: The incremental learning uses an elastic weight consolidation algorithm, with the importance weight decay coefficient set to 0.

95.

10. The modulation-adaptive deep learning spectrum sensing method according to claim 1, characterized in that: The method is deployed on an edge computing device, using a digital signal processor that supports floating-point operations, with a main frequency of not less than 1.2 GHz and a memory capacity of not less than 2 gigabytes.

Images