Unmanned aerial vehicle interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion

Abstract

The invention discloses an unmanned aerial vehicle interference detection method based on a wide-deep attention residual network and random perturbation enhancement, and belongs to the technical field of unmanned aerial vehicle communication security. The method comprises the following steps: firstly, extracting basic features of unmanned aerial vehicle received signals, and expanding feature dimensions through polynomial interaction; and then feature importance is calculated by using a set uniformity index and dynamic classification is carried out. In a network construction stage, a wide-deep fusion architecture is designed: a deep path establishes an independent SE-ResNet sub-network for each feature level, and high-dimensional nonlinear features are extracted by using residual connection and a channel attention mechanism; a wide-layer path directly processes full-amount characteristics and memorizes a low-order linear rule. And finally, fusing and classifying the two paths of features. In the reasoning stage, an enhancement mechanism during testing is introduced, and environmental accidental errors are eliminated by injecting tiny random noise into real-time signals and performing multi-step average prediction. The method effectively solves the problems of insufficient feature physical association mining and weak model generalization ability in the prior art, and significantly improves the accuracy and robustness of unmanned aerial vehicle interference detection in a complex electromagnetic environment.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) communication security technology, and in particular to a UAV interference detection method based on wide-depth attention residual networks and random perturbation enhancement. This method aims to improve the anti-interference capability of UAV communication systems and ensure the normal operation of UAVs in complex electromagnetic environments. Background Technology

[0002] With the rapid development of drone technology, the security of drone communication systems is receiving increasing attention. During missions, drones may encounter various forms of radio signal interference. This interference not only affects the communication quality of the drone but may also threaten its safe flight and even cause it to lose control. Therefore, effectively detecting and classifying the interference signals encountered by drones has become one of the key issues in ensuring drone communication security.

[0003] Traditional interference detection methods primarily rely on spectrum analysis techniques, which determine the presence of interference by monitoring changes in signal strength within a specific frequency range. However, this method suffers from low detection accuracy and slow response speed, making it difficult to meet the demands of modern UAVs for rapid and accurate interference detection. Furthermore, traditional methods often neglect the correlation between signal features, failing to fully utilize signal feature information for efficient interference classification.

[0004] In recent years, deep learning technology has achieved remarkable success in fields such as image recognition and natural language processing due to its powerful feature extraction and pattern recognition capabilities. Deep learning-based methods have also been attempted for application in UAV interference detection, automatically learning signal features by constructing models such as convolutional neural networks, thus improving the efficiency and accuracy of interference detection. However, most existing deep learning-based interference detection methods employ static feature extraction, failing to fully consider the differences in importance between different features, resulting in insufficiently refined feature extraction and affecting the effectiveness of interference detection and classification.

[0005] To address the aforementioned problems, this invention proposes a UAV interference detection method based on a wide-depth architecture and adaptive hierarchical feature fusion. This method innovatively constructs a wide-depth network model including both wide-side and deep-side pathways. It utilizes a parallel channel attention residual subnetwork on the deep side to perform nonlinear deep mining of dynamically hierarchical features, while simultaneously taking into account the linear memory of global features through the wide-side pathway. Furthermore, a test-time enhancement mechanism is introduced during the inference phase to eliminate prediction randomness, thereby achieving high-precision and robust detection and classification of UAV interference signals in complex electromagnetic environments. Summary of the Invention

[0006] To address the problems of low feature utilization, weak generalization ability, and insufficient detection accuracy of existing UAV interference detection methods in complex electromagnetic environments, this invention proposes a UAV interference detection method and system based on a wide-depth architecture and adaptive hierarchical feature fusion.

[0007] A UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion includes the following steps:

[0008] Step 1: Acquire the received signal of the UAV communication link and extract the original feature set from the received signal;

[0009] Step 2: Perform nonlinear expansion processing on the original feature set to generate an expanded feature set containing the original features and polynomial interaction features;

[0010] Step 3: Calculate the importance score of each feature in the extended feature set based on the set uniformity index, and use a clustering algorithm to divide the features into K feature subsets with different importance levels;

[0011] Step 4: Construct a wide-depth interference detection network model, which includes a wide-side path and a deep-side path;

[0012] Step 5: Input the complete extended feature set into the breadth-side path for global linear feature extraction; at the same time, input the K feature subsets divided in Step 3 into the corresponding K parallel sub-networks in the depth-side path for deep nonlinear feature extraction.

[0013] Step 6: Concatenate and fuse the outputs of the breadth-side path and the depth-side path, and output the predicted probability distribution of the interference type through the classifier;

[0014] Step 7: In the model inference stage, an enhancement mechanism is introduced during testing to perform multiple micro-perturbation predictions on the input samples and take the average value to obtain the final interference detection result.

[0015] In step 4, the deep side path includes K parallel channel attention residual subnetworks, each of which processes a subset of features at an importance level. The channel attention residual subnetwork consists of several stacked residual blocks and squeezing and excitation modules, which are used to adaptively weight and calibrate the feature channels while extracting deep features.

[0016] The mathematical processing of the extrusion and excitation module includes:

[0017] Squeezing operation: compressing the feature map output by the convolution. Perform global average pooling to generate channel statistics. , where the k-th element The calculation formula is:

[0018]

[0019] Activation operation: Establish the dependency relationship between channels through two fully connected layers to generate a weight vector. :

[0020]

[0021] in, To correct the linear unit function, It is a sigmoid function. To reduce dimensionality weights, The weights are for dimensionality scaling, and r is the scaling parameter.

[0022] Recalibration operation: The generated weights Perform channel multiplication with the original feature map U to obtain the final output. :

[0023]

[0024] In the wide-depth interference detection network model, all convolutional and fully connected layers use the Swish self-gated activation function, the complete mathematical expression of which is:

[0025]

[0026] In this method, hyperparameters The default value is 1. This function has the characteristics of being unbounded at the upper end, bounded at the lower end, smooth and non-monotonic. It can preserve negative information in signal features and improve the gradient flow of deep networks.

[0027] The breadth-side pathway in step S4 is a multilayer perceptron, whose input is a complete extended feature set. The output dimension is designed so that the dimensionality ratio of the breadth-side features and the depth-side features before fusion is within a preset range, balancing memory capability and generalization capability. Compared to existing technologies that often directly concatenate features of different dimensions, leading to one side's features dominating during training, this invention effectively balances the gradient flow of the model by presetting the dimensionality ratio of the breadth and depth features. This design can maintain the "memory capability" of low-order feature combinations through the breadth-side pathway, preventing information loss, and can fully utilize the "generalization capability" of high-order nonlinear features through the depth-side pathway, thereby maximizing the complementary advantages of multi-view feature fusion while ensuring model convergence speed.

[0028] When training a wide-depth interference detection network model, a cross-entropy loss function with label smoothing is employed, combined with a cosine annealing hot restart strategy to dynamically adjust the learning rate, thereby improving the model's convergence speed and generalization accuracy. Traditional step-wise or exponentially decaying learning rate strategies easily cause the model to get stuck in local minima (saddle points), making it difficult to obtain the global optimum. This invention introduces a cosine annealing hot restart strategy, which can periodically increase the learning rate to escape local traps and explore a wider parameter space. At the same time, combined with label smoothing technology, it effectively suppresses the model's overfitting tendency in the later stages of training and reduces the model's sensitivity to noise labels that may exist in the training set, thus significantly enhancing the model's robustness in unknown and complex electromagnetic environments.

[0029] The method for calculating the feature importance score in step 3 is as follows:

[0030] Define set uniformity index The calculation formula is as follows:

[0031]

[0032] in The total number of interference categories, For set The Middle The proportion of samples of each class; for each extended feature The median of this feature sample set Divided into two subsets and Calculate the weighted uniformity after partitioning:

[0033]

[0034] Calculate the importance score of this feature. Gain amount for uniformity index before and after division:

[0035]

[0036] The score The larger the value, the more characteristic it represents. The higher the contribution to distinguishing categories, the better. Compared to existing feature selection methods based on variance (such as Principal Component Analysis, PCA) or linear correlation coefficients, the ensemble evenness index (based on information entropy and Gini impurity principles) used in this invention can effectively capture the complex nonlinear dependencies between features and interfering categories, no longer limited to linear features. Furthermore, this method uses the median as the dividing threshold, exhibiting inherent robustness to potential sudden spikes or outliers in the acquired signal, and is less susceptible to extreme values. This dynamic scoring mechanism based on information gain can accurately select high-discrimination features, ensuring that subsequent deep neural networks can concentrate computational resources on key information, avoiding model performance degradation due to feature redundancy or noise interference.

[0037] The specific operation of step S2 is as follows:

[0038] By using a polynomial feature generation algorithm, the original feature set is extended by a second-order polynomial. The resulting extended feature set not only contains the original features, but also the product terms and square terms between features, so as to explicitly provide the network with nonlinear correlation information between features.

[0039] The generated extended feature set is standardized to make it follow a distribution with a mean of 0 and a variance of 1.

[0040] The clustering algorithm in step S3 uses the K-means algorithm:

[0041] The importance scores of all features are used as one-dimensional data inputs into the K-means algorithm, and the number of cluster centers is set to 3.

[0042] Based on the value of the cluster centers, features are automatically divided into three levels: high importance, medium importance, and low importance, corresponding to the inputs of three parallel sub-networks in the deep side-channel. Existing technologies typically use manually set fixed thresholds for feature selection, which is difficult to adapt to varying electromagnetic interference environments. This invention employs the K-means clustering algorithm, which can adaptively determine the hierarchical boundaries based on the actual distribution characteristics of feature importance scores, achieving fully data-driven hierarchical structure. By guiding features of different importance to parallel sub-networks, not only is feature processing decoupled, avoiding interference from low-value information in core feature extraction, but the network can also learn differentiated representations for features at different levels, significantly improving the model's analytical depth and classification accuracy when processing mixed interference signals.

[0043] The test-time enhancement mechanism in step S7 specifically includes:

[0044] For the test sample to be detected, N random Gaussian noise vectors are generated, where the noise intensity is controlled by hyperparameters;

[0045] Construct N augmented samples and input these N samples into the trained wide-depth interference detection network model respectively;

[0046] The model outputs N probability distribution vectors, and their arithmetic mean is calculated as the final predicted probability distribution. The category with the highest probability is selected as the final detection result. Existing technologies mostly employ a single forward pass during model deployment, making their detection results highly susceptible to interference from instantaneous environmental noise or random jitter during signal transmission, leading to fluctuations or even misjudgments. This invention introduces a Test-Time Enhancement (TTA) mechanism, which actively injects controlled random Gaussian perturbations during the inference phase and performs multiple Monte Carlo sampling averages, essentially statistically modeling the uncertainty of the signal to be detected. This method effectively eliminates random errors and smooths decision boundaries by utilizing statistical averaging effects without increasing model training costs or changing model parameters, thereby significantly improving the model's anti-interference capability and prediction confidence in dynamic and complex electromagnetic environments.

[0047] The broad-side pathway in step S4 is a multilayer perceptron, whose input is a complete extended feature set. The output dimension is designed so that the dimensionality ratio of the broad-side features to the deep-side features before fusion is within a preset range, in order to balance memory capacity and generalization ability. The classifier structure in step S6 includes:

[0048] Concatenation layer: Concatenates the output vector of the breadth side with the output vectors of all sub-networks of the depth side along the channel dimension;

[0049] Fusion layer: contains fully connected layers, batch normalization layers, and random deactivation layers to prevent overfitting;

[0050] Output layer: A fully connected layer with a size equal to the total number of interference categories, which outputs the classification probability using a normalized exponential function.

[0051] When training the wide and deep interference detection network model, a label-smoothed cross-entropy loss function is used, and a cosine annealing hot restart strategy is combined to dynamically adjust the learning rate in order to improve the convergence speed and generalization accuracy of the model.

[0052] A drone interference detection system includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method steps as described in any one of claims 1 to 11.

[0053] Compared with the prior art, the advantages of the present invention are as follows:

[0054] 1. Complementary Advantages at the Architectural Level (Wide & Deep Fusion Mechanism): This invention innovatively constructs a parallel processing wide & deep fusion network architecture. Compared to traditional "black box" models that rely solely on deep neural networks, this invention directly processes high-dimensional features expanded by polynomials through "broad-side pathways," preserving fleeting low-order co-occurrence patterns in signal features (i.e., "memory ability"); simultaneously, it utilizes "deep-side pathways" to mine deep nonlinear abstract patterns between features (i.e., "generalization ability"). This architectural design effectively solves the problem of single models losing shallow physical laws due to overfitting deep features when facing complex electromagnetic environments, significantly improving the overall discriminative performance of the model.

[0055] 2. Adaptability and Robustness of Feature Engineering (Dynamic Hierarchical Strategy): This invention abandons the static methods of feature selection in existing technologies that rely on manual experience or simple variance thresholds. Instead, it proposes an adaptive feature hierarchical mechanism based on set evenness index (based on the principle of information entropy) and K-Means clustering. This mechanism can accurately quantify the nonlinear dependency between features and interference categories from a statistical perspective, and dynamically divide features into different importance levels accordingly. This not only achieves on-demand allocation of computing resources, but more importantly, it effectively blocks the propagation of environmental noise in deep networks by physically isolating low-value features from core features, thereby improving anti-interference capabilities from the source.

[0056] 3. Refinement of Deep Feature Extraction (Channel Attention Mechanism): In the deep side pathway, this invention integrates a squeezing and activation (SE) module. Unlike ordinary convolutional networks that treat all channels equally, the SE module can adaptively learn and recalibrate the weight coefficients of each feature channel through global information aggregation (i.e., "recalibration operation"). This mechanism enables the model to automatically suppress background noise channels that are useless for judgment and enhance feature channels containing key interference patterns (such as pulse edges, specific frequency points), thereby maintaining extremely high feature sensitivity even in low signal-to-noise ratio environments. Attached Figure Description

[0057] Figure 1 A flowchart illustrating the overall process of a UAV interference detection method based on a wide-depth architecture and adaptive hierarchical feature fusion, provided for embodiments of the present invention.

[0058] Figure 2 This is a schematic diagram of the overall architecture of the wide-depth interference detection network model constructed in an embodiment of the present invention. Detailed Implementation

[0059] To enhance understanding of the present invention, the embodiments will be described in detail below with reference to the accompanying drawings.

[0060] Example: To address the problems of low feature utilization, weak generalization ability, and insufficient detection accuracy of existing UAV interference detection methods in complex electromagnetic environments, this invention proposes a UAV interference detection method based on a wide-depth architecture and adaptive hierarchical feature fusion.

[0061] A UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion includes the following steps:

[0062] Step 1: Acquire the received signal of the UAV communication link and extract the original feature set from the received signal;

[0063] Step 2: Perform nonlinear expansion processing on the original feature set to generate an expanded feature set containing the original features and polynomial interaction features;

[0064] Step 3: Calculate the importance score of each feature in the extended feature set based on the set uniformity index, and use a clustering algorithm to divide the features into K feature subsets with different importance levels;

[0065] Step 4: Construct a wide-depth interference detection network model, which includes a wide-side path and a deep-side path;

[0066] Step 5: Input the complete extended feature set into the breadth-side path for global linear feature extraction; at the same time, input the K feature subsets divided in Step 3 into the corresponding K parallel sub-networks in the depth-side path for deep nonlinear feature extraction.

[0067] Step 6: Concatenate and fuse the outputs of the breadth-side path and the depth-side path, and output the predicted probability distribution of the interference type through the classifier;

[0068] Step 7: In the model inference stage, an enhancement mechanism is introduced during testing to perform multiple micro-perturbation predictions on the input samples and take the average value to obtain the final interference detection result.

[0069] The deep side path in step 4 includes K parallel channel attention residual subnetworks, each of which processes a subset of features at an importance level. The channel attention residual subnetwork consists of several stacked residual blocks and squeezing and excitation modules, which are used to adaptively weight and calibrate the feature channels while extracting deep features.

[0070] The mathematical processing of the extrusion and excitation module includes:

[0071] Squeezing operation: compressing the feature map output by the convolution. Perform global average pooling to generate channel statistics. , where the k-th element The calculation formula is:

[0072]

[0073] Activation operation: Establish the dependency relationship between channels through two fully connected layers to generate a weight vector. :

[0074]

[0075] in, To correct the linear unit function, It is a sigmoid function. To reduce dimensionality weights, The weights are for dimensionality scaling, and r is the scaling parameter.

[0076] Recalibration operation: The generated weights Perform channel multiplication with the original feature map U to obtain the final output. :

[0077]

[0078] In the wide-depth interference detection network model, all convolutional and fully connected layers use the Swish self-gated activation function, the complete mathematical expression of which is:

[0079]

[0080] In this method, hyperparameters The default value is 1. This function has the characteristics of being unbounded at the upper end, bounded at the lower end, smooth and non-monotonic. It can preserve negative information in signal features and improve the gradient flow of deep networks.

[0081] The method for calculating the feature importance score in step 3 is as follows:

[0082] Define set uniformity index The calculation formula is as follows:

[0083]

[0084] in The total number of interference categories, For set The Middle The proportion of samples of each class; for each extended feature The median of this feature sample set Divided into two subsets and Calculate the weighted uniformity after partitioning:

[0085]

[0086] Calculate the importance score of this feature. Gain amount for uniformity index before and after division:

[0087]

[0088] The score The larger the value, the more characteristic it represents. The higher the contribution to distinguishing categories.

[0089] The specific operation of step S2 is as follows:

[0090] By using a polynomial feature generation algorithm, the original feature set is extended by a second-order polynomial. The resulting extended feature set not only contains the original features, but also the product terms and square terms between features, so as to explicitly provide the network with nonlinear correlation information between features.

[0091] The generated extended feature set is standardized to make it follow a distribution with a mean of 0 and a variance of 1.

[0092] The clustering algorithm in step S3 uses the K-means algorithm:

[0093] The importance scores of all features are used as one-dimensional data inputs into the K-means algorithm, and the number of cluster centers is set to 3.

[0094] Based on the value of the cluster centers, the features are automatically divided into three levels: high importance, medium importance, and low importance, which correspond to the inputs of the three parallel sub-networks of the deep side path, respectively.

[0095] The breadth-side pathway in step S4 is a multilayer perceptron, whose input is a complete extended feature set. The output dimension is designed so that the ratio of the breadth-side features to the depth-side features before fusion is within a preset range, so as to balance memory ability and generalization ability.

[0096] The classifier structure in step S6 includes:

[0097] Concatenation layer: Concatenates the output vector of the breadth side with the output vectors of all sub-networks of the depth side along the channel dimension;

[0098] Fusion layer: contains fully connected layers, batch normalization layers, and random deactivation layers to prevent overfitting;

[0099] Output layer: A fully connected layer with a size equal to the total number of interference categories, which outputs the classification probability using a normalized exponential function.

[0100] When training the wide and deep interference detection network model, a label-smoothed cross-entropy loss function is used, and a cosine annealing hot restart strategy is combined to dynamically adjust the learning rate in order to improve the convergence speed and generalization accuracy of the model.

[0101] The test-time enhancement mechanism in step S7 specifically includes:

[0102] For the test sample to be detected, N random Gaussian noise vectors are generated, where the noise intensity is controlled by hyperparameters;

[0103] Construct N augmented samples and input these N samples into the trained wide-depth interference detection network model respectively;

[0104] Obtain the N probability distribution vectors output by the model, calculate their arithmetic mean as the final predicted probability distribution, and take the category with the highest probability as the final detection result.

[0105] To verify the effectiveness of the proposed UAV interference detection method based on wide-depth attention residual network and random perturbation enhancement, simulation experiments were conducted on a real UAV communication interference dataset, and comparative analysis was performed with various benchmark models and ablation variants.

[0106] Hardware environment:

[0107] Cloud server: A deep learning server based on Ubuntu 22.04.4 LTS, equipped with an NVIDIA GeForce RTX 4080 GPU, an Intel Core i9-14900KF CPU, and 32 GB of system memory.

[0108] Local workstation: Used for code development, remote server connections, and visualization analysis of experimental results.

[0109] Software environment:

[0110] Operating system: Ubuntu 22.04.4 LTS (server); Windows 11 (local).

[0111] Deep learning frameworks: PyTorch 1.12.1, Python 3.10.14.

[0112] Key libraries required: data processing and machine learning libraries such as numpy, pandas, and scikit-learn, as well as the Python standard library (os, warnings, random, etc.).

[0113] The experimental data were derived from electromagnetic signal data simulated and acquired using an SDR device. After preprocessing, the dataset contained various typical interference types (such as single-tone interference, frequency sweep interference, and noise-frequency modulation interference). The dataset was divided using a 5-fold stratified cross-validation method to ensure the statistical significance and reliability of the test results.

[0114] Parameter settings

[0115] The key hyperparameter settings for model training and inference are as follows:

[0116] Feature engineering: The polynomial extension order is set to 2 (including interaction terms), and the standardization method is StandardScaler.

[0117] Network structure: The number of deep subnetworks is K=3 (corresponding to the three importance levels of high, medium and low), the hidden layer dimension of the Deep part is 128, the direct connection dimension of the Wide part is 128, and the dropout rate is 0.5.

[0118] Optimizer: The AdamW optimizer is used, with an initial learning rate of 0.001 and a weight decay factor of [value missing]. .

[0119] Learning rate scheduling: Cosine annealing warm restarts are used, with a cycle of... .

[0120] Training strategy: Batch size is set to 64, total number of epochs is 25, and the loss function is cross-entropy loss with a label smoothing coefficient of 0.1.

[0121] Test-Time Enhancement (TTA): Standard Deviation of Noise Injection During Inference Phase The average number of predictions is N=5.

[0122] To highlight the advanced nature of the technical solution of this invention, the classic one-dimensional deep residual network (StandardResNet-1D) was selected as the baseline model.

[0123] The baseline model uses only the original feature inputs, does not include multinomial expansion, does not include wide-side pathways, and does not introduce SE attention mechanisms and test-time enhancement strategies, representing the current level of general deep learning detection.

[0124] The method of this invention adopts a complete wide-depth architecture, integrating polynomial nonlinear expansion, dynamic feature hierarchical structure, SE channel attention mechanism, and test-time enhanced inference mechanism.

[0125] Experimental Results and Analysis

[0126] The average accuracy on the test set was used as the core evaluation metric for the experiment. The comparison results are shown in Table 1.

[0127] Table 1 Comparison of detection accuracy under different interference types

[0128]

[0129] Results Analysis: As can be seen from Table 1, the model of this invention outperforms the benchmark model in all types of interference.

[0130] For simple signals (Clean, SP): Although the baseline model can also achieve a high recognition rate, this invention achieves 100% perfect classification through a wide-depth architecture, completely eliminating false alarms.

[0131] For narrowband signals (Tone): Compared to the baseline model's accuracy of 83.40%, this invention improves accuracy by 8.60%, demonstrating the effective memorization of frequency domain peak characteristics by the wide path.

[0132] For Barrage (PA) signals: The model in this invention achieves the greatest performance gain (more than 7%~8%) on broadband noise and transient impulse interference, which are difficult for the benchmark model to distinguish (accuracy of only about 70%). This is mainly attributed to the adaptive enhancement of weak feature channels by the SE channel attention mechanism, and the smoothing effect of the enhancement mechanism on random noise fluctuations during testing, which significantly improves the robustness of the model in harsh electromagnetic environments.

[0133] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion, characterized in that, Includes the following steps: Step 1: Acquire the received signal of the UAV communication link and extract the original feature set from the received signal; Step 2: Perform nonlinear expansion processing on the original feature set to generate an expanded feature set containing the original features and multinomial interaction features; Step 3: Calculate the importance score of each feature in the extended feature set based on the set uniformity index, and use a clustering algorithm to divide the features into K feature subsets with different importance levels; Step 4: Construct a wide-depth interference detection network model, which includes wide-side and deep-side paths; Step 5: Input the complete extended feature set into the breadth-side pathway for global linear feature extraction; at the same time, input the K feature subsets divided in Step 3 into the corresponding K parallel sub-networks in the depth-side pathway for deep nonlinear feature extraction. Step 6: Concatenate and fuse the outputs of the breadth-side path and the depth-side path, and output the predicted probability distribution of the interference type through the classifier; Step 7: In the model inference stage, an enhancement mechanism is introduced during testing to perform multiple micro-perturbation predictions on the input samples and take the average value to obtain the final interference detection result.

2. The UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion according to claim 1, characterized in that, The deep side path in step 4 contains K parallel channel attention residual subnetworks, each of which processes a subset of features at an importance level. The channel attention residual subnetwork consists of several stacked residual blocks and squeezing and excitation modules, which are used to adaptively weight and calibrate the feature channels while extracting deep features. The mathematical processing of the extrusion and excitation module includes: 1) Squeezing operation: compressing the feature map output by convolution. Perform global average pooling to generate channel statistics. The k-th element The calculation formula is: in, Represents the squeezing function; Representation of feature map The One feature channel; Indicates the spatial coordinate index of the feature map; Indicates the first Each feature channel in spatial location Output value at; The range of values ​​is , 2) Activation Operation: Establish the dependency relationship between channels through two fully connected layers to generate a weight vector. : in, Represents the activation function; To correct the linear unit function, It is a sigmoid function. To reduce dimensionality weights, The weights are for dimensionality scaling, and r is the scaling parameter. 3) Recalibration operation: The generated weights... Perform channel multiplication with the original feature map U to obtain the final output. : , in, Indicates the recalibrated first Each feature channel outputs; This represents a channel-wise weighted function; Represents the weight vector The first in The element corresponding to the element The weights of each feature channel; For the original feature map U, the first Data from each feature channel.

3. The UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion according to claim 1, characterized in that, In step 4, all convolutional and fully connected layers in the wide-depth interference detection network model use the Swish self-gated activation function, as follows: In this method, hyperparameters The default value is 1. This function has the characteristics of being unbounded at the upper end, bounded at the lower end, smooth and non-monotonic. It can preserve negative information in signal features and improve the gradient flow of deep networks.

4. The UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion according to claim 1, characterized in that, The specific method for calculating feature importance scores in step 3 is as follows: Define set uniformity index The calculation formula is as follows: in The total number of interference categories, For set The Middle The proportion of samples of each class; for each extended feature The median of this feature sample set Divided into two subsets and Calculate the weighted uniformity after partitioning: Calculate the importance score of this feature. Gain amount for uniformity index before and after division: Score The larger the value, the more characteristic it represents. The higher the contribution to distinguishing categories.

5. The UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion according to claim 1, characterized in that, The specific operation of step S2 is as follows: 1) Using a polynomial feature generation algorithm, the original feature set is extended by a second-order polynomial. The resulting extended feature set not only contains the original features, but also the product terms and square terms between features, so as to explicitly provide the network with nonlinear correlation information between features. 2) Standardize the generated extended feature set so that it follows a distribution with a mean of 0 and a variance of 1.

6. The UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion according to claim 1, characterized in that, The clustering algorithm in step S3 uses the K-means algorithm: 1) Input the importance scores of all features as one-dimensional data into the K-means algorithm, and set the number of cluster centers to 3; 2) Based on the value of the cluster center, the features are automatically divided into three levels: high importance, medium importance, and low importance, which correspond to the inputs of the three parallel sub-networks of the deep side path, respectively.

7. The UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion according to claim 1, characterized in that, The test-time enhancement mechanism in step S7 specifically includes: 1) For the test sample to be detected, generate N random Gaussian noise vectors, where the noise intensity is controlled by hyperparameters; 2) Construct N augmented samples and input these N samples into the trained wide and deep interference detection network model respectively; 3) Obtain the N probability distribution vectors output by the model, calculate their arithmetic mean as the final predicted probability distribution, and take the category with the highest probability as the final detection result.

8. The UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion according to claim 1, characterized in that, The classifier structure in step S6 includes: 1) Concatenation layer: Concatenates the output vector of the breadth side with the output vectors of all sub-networks of the depth side along the channel dimension; 2) Fusion layer: This layer includes a fully connected layer, a batch normalization layer, and a random deactivation layer to prevent overfitting; 3) Output layer: A fully connected layer with a size equal to the total number of interference categories, which outputs the classification probability in conjunction with a normalized exponential function.

9. A drone interference detection system, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed by the processor, they implement the UAV interference detection method based on wide-depth architecture and adaptive hierarchical feature fusion as described in any one of claims 1-8.

Images