Unmanned aerial vehicle signal intelligent identification method and system based on deep learning, and storage medium

Abstract

The invention discloses an unmanned aerial vehicle signal intelligent identification method and system based on deep learning, and a storage medium. The method comprises the following steps: executing local change rate calculation based on a standardized signal data set, constructing a dynamic adjustment rule, and generating an adaptive window structure; executing a frequency spectrum transformation operation to generate a frequency spectrum drift index; if the frequency spectrum drift index is greater than a preset drift threshold value in a continuous time period, executing window contraction updating; if the frequency spectrum drift index is smaller than a drift threshold value in a continuous time period and the fluctuation frequency in a preset time interval is larger than a preset fluctuation frequency threshold value, the window structure is kept unchanged, and a spectrogram sequence is formed; and inputting the signals into the improved CBAM-CNN model, and outputting a signal category label, a structural feature label and a confidence score. According to the method, the identification accuracy, the response speed and the structure understanding capability of the communication signal of the unmanned aerial vehicle in a strong-interference and multi-state environment are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of wireless signal processing technology, and in particular to a method, system, and storage medium for intelligent identification of UAV signals based on deep learning. Background Technology

[0002] With the frequent occurrence of low-altitude flight activities and the rapid popularization of drone technology, drone identification methods based on communication signal characteristics have gradually become an important research direction for public safety and airspace management. Existing drone signal identification technologies mainly rely on static spectrum analysis or fixed-window feature extraction methods, combined with traditional classification models such as convolutional neural networks, to determine signal categories. However, in complex environments, communication signals exhibit strong non-stationarity, spectral abruptness, and structural diversity. Existing methods generally suffer from the following problems when dealing with dynamic signal structure changes:

[0003] Fixed-window methods cannot adapt to rapid changes in signal strength and structure, easily leading to truncation or redundancy of spectral information, affecting the recognition model's ability to perceive abrupt changes. The neglect of decoupling modeling between frequency and time dimensions during spectrogram construction results in cross-interference in feature representations, hindering the model's ability to distinguish between abrupt and stable structures. Traditional classification models often rely on a single path for signal discrimination, lacking an adaptive path selection mechanism for structural differences, and are unable to effectively handle complex signal morphologies such as structural mixing or spectral drift. Furthermore, the lack of effective adversarial mechanisms in the pseudo-label generation process makes the model's accuracy in identifying signal structure types susceptible to misleading pseudo-labels, reducing overall confidence and classification reliability.

[0004] Therefore, how to provide a method, system, and storage medium for intelligent recognition of drone signals based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based intelligent identification method, system, and storage medium for UAV signals. This invention fully utilizes spectrogram sequences, spectral drift indices, an improved CBAM-CNN model, a structure label prediction channel, and a label-guided fusion decoder. It details the entire process of standardizing UAV communication signals under different environments and flight conditions, calculating variation intensity indices, constructing adaptive window structures, generating spectrograms, extracting spectral features, decoupling frequency and time encoding, outputting structure type labels, and decoding fusion features. It possesses the advantages of high signal recognition accuracy, strong structural classification reliability, and good robustness in intelligent classification.

[0006] A method for intelligent identification of drone signals based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect communication signals from the UAV under different environments and flight conditions, construct a time-domain signal sample sequence, perform amplitude normalization and time length alignment processing, and generate a standardized signal dataset;

[0008] Step 2: Perform local change rate calculation on the standardized signal dataset, generate a change intensity index based on the signal amplitude change rate, construct dynamic adjustment rules according to the change intensity index, determine the starting position and length of the analysis window according to the dynamic adjustment rules, and generate an adaptive window structure.

[0009] Step 3: Perform a spectrum transformation operation on the signal using the adaptive window structure to generate a spectrum sequence, calculate the spectrum center change value of adjacent windows, and generate a spectrum drift index.

[0010] Step 4: If the spectrum drift index is greater than the preset drift threshold within a continuous time period, then perform window shrinkage update; if the spectrum drift index is less than the drift threshold within a continuous time period and the number of fluctuations within the preset time interval is greater than the preset number of fluctuations threshold, then keep the window structure unchanged and form a spectrum sequence.

[0011] Step 5: Input the spectrum sequence into the improved CBAM-CNN model and output the signal category label, structural feature label and confidence score; the improved CBAM-CNN model includes a spectrum feature extraction channel, a frequency-time decoupling coding structure, a structural label prediction channel, a pseudo-label adversarial structure, a recognition path selection structure, a jump recognition path, a stable recognition path and a label-guided fusion decoder.

[0012] Preferably, step one specifically comprises:

[0013] The communication signals of the UAV during the takeoff, cruise, hovering and landing phases are collected, including control signals, image transmission signals and navigation signals.

[0014] The communication signal is continuously sampled according to the set time sampling period to construct a time-domain signal sample sequence with time order, and the amplitude, frequency, phase and time index of each sampling point are recorded.

[0015] The amplitude normalization process is performed on the time-domain signal sample sequence, and a linear mapping operation is performed based on the set upper and lower amplitude limits;

[0016] The time-domain signal sample sequence is subjected to time length alignment processing, and time-domain signal sample sequences of different lengths are interpolated and padded or truncated according to a unified sampling frequency and time reference.

[0017] The processed time-domain signal sample sequences are labeled according to flight status, environment, and signal type labels to generate a standardized signal dataset.

[0018] Preferably, step two specifically includes:

[0019] Each time-domain signal sample sequence in the standardized signal dataset is divided into continuous signal segments;

[0020] For each signal segment, the rate of change of amplitude is calculated, and a rate of change sequence is generated based on the amplitude difference and time difference between adjacent sampling points;

[0021] A sliding window structure is constructed on the rate of change sequence, and the maximum rate of change, average rate of change, and fluctuation amplitude within each sliding window are calculated to generate a change intensity index.

[0022] Multiple adjustment level thresholds are set based on the intensity of change index, and dynamic adjustment rules for the analysis window are established.

[0023] Based on the intensity level of change at the starting position of the analysis window, set the corresponding window length and sliding step size;

[0024] Signal segments are extracted sequentially according to the starting position and corresponding window length of the analysis window, an adaptive window structure is constructed, and a window segmented signal dataset with time positioning information is generated.

[0025] Preferably, step three specifically includes:

[0026] Perform a short-time Fourier transform operation on each window segment in the adaptive window structure to generate a two-dimensional spectrogram containing information in both the frequency and time dimensions;

[0027] The two-dimensional spectrograms are stacked sequentially according to the window order to construct a spectrogram sequence, and the time index and window number of each spectrogram are labeled.

[0028] For the spectrum diagrams corresponding to two adjacent windows, extract the main frequency distribution interval, calculate the position difference of the spectrum energy center, and generate the spectrum center change value;

[0029] The change values ​​of the spectral center of each pair of adjacent spectrograms are combined into a time series to form a spectral drift index.

[0030] Record the change value, direction of change and drift magnitude of the spectrum drift index at each time point to construct an indexed spectrum drift index sequence.

[0031] Preferably, step four specifically includes:

[0032] Set the spectrum drift threshold and the spectrum fluctuation number threshold, and define the continuous time period and the preset time interval;

[0033] A threshold comparison operation is performed on the drift amplitude values ​​in the spectrum drift index sequence within a continuous time period. If all drift amplitude values ​​are greater than the spectrum drift threshold, a window shrinking operation is performed on the analysis window within the corresponding time period.

[0034] The window shrinking operation includes shortening the window length, updating the start and end positions, reconstructing the corresponding window number, and synchronously updating the spectrogram sequence.

[0035] If the drift amplitude of the spectrum drift index sequence within the preset time interval is less than the spectrum drift threshold, and the number of times the spectrum drift direction changes within the time interval is greater than the number of times the spectrum fluctuates, then the starting position and length of the current analysis window will remain unchanged.

[0036] The set of windows that are either shrunk or remain unchanged is used as the final set of analysis windows, and the spectrum transformation operation is called sequentially to generate a spectrum sequence.

[0037] Preferably, the spectral feature extraction channel extracts spatial features of the spectrogram sequence based on a convolutional structure and generates a feature tensor. The frequency-time decoupling coding structure performs frequency and time dimension separation calculations on the feature tensor to generate frequency and time vectors, and fuses them through an interactive attention mechanism to form decoupling enhanced features, specifically:

[0038] A spectral feature extraction channel consisting of multi-layer two-dimensional convolutional units, batch normalization units, and nonlinear activation units is constructed. Convolution, feature aggregation, and scaling operations are sequentially performed on the input spectrogram sequence to extract frequency distribution features and temporal evolution features, generating a multidimensional feature tensor.

[0039] In the frequency-time decoupled coding structure, frequency branch coding units and time branch coding units are set up. The frequency branch coding unit performs one-dimensional convolution and global pooling operations along the frequency dimension to extract the local energy distribution of each frequency band and generate a frequency vector. The time branch coding unit performs sliding convolution and attention weight calculation operations along the time dimension to extract the time series change trend and generate a time vector.

[0040] In the interactive attention mechanism, a frequency attention matrix and a time attention matrix are introduced, and the frequency vector and time vector are weighted and adjusted respectively. A bidirectional weighted feature representation is formed by matrix multiplication.

[0041] The weighted frequency vector and time vector are concatenated and fused, and then input into the multi-layer feature mapping unit to generate decoupled enhanced features that contain frequency information and temporal correlation.

[0042] Preferably, the structure label prediction channel receives the decoupled enhancement features and outputs structure type labels by combining the channel attention structure and the spatial attention structure. The structure type labels include jump structures, stable structures, and hybrid structures. The pseudo-label adversarial structure outputs pseudo-structure labels based on the auxiliary generation network. The weight distribution of the structure label prediction channel is then corrected by comparing the structure type label generation adversarial loss with the pseudo-structure labels. Specifically:

[0043] Construct a channel attention structure, perform global average pooling and global max pooling operations on the decoupled enhancement features, generate channel description vectors, perform attention weight calculation, generate channel attention weights, and perform channel-by-channel multiplication operations on the channel attention weights and decoupled enhancement features;

[0044] A spatial attention structure is constructed. Two-dimensional convolution and pooling operations are performed on the channel-weighted feature map to extract the spatial attention feature map and generate a spatial attention weight map. The spatial attention weight map and the channel-weighted feature map are multiplied pixel by pixel to form an attention enhancement feature map.

[0045] The attention-enhanced feature map is input into the structure classifier, and the structure type label is output through a fully connected layer and Softmax. The structure type label includes jumping structure, stable structure and hybrid structure.

[0046] A pseudo-label adversarial structure is constructed, which includes an auxiliary generation network and a discriminant unit. The auxiliary generation network generates pseudo-structure labels by taking random noise and encoded features as input. The discriminant unit simultaneously receives structure type labels and pseudo-structure labels and outputs the discrimination probability.

[0047] Based on the discriminant probability, an adversarial loss function is constructed. The difference between the distribution of real labels and pseudo labels is compared. Gradient backpropagation is performed to update the network parameters in the structural label prediction channel and adjust the weight distribution.

[0048] Preferably, the identification path selection structure activates the jump identification path, stable identification path, or a weighted fusion path of both based on the structure type label, generating a fused feature vector; the label-guided fusion decoder receives the fused feature vector and the structure type label, and outputs the UAV signal identification result, structural interpretation information, and confidence distribution, realizing time-frequency feature decoupling, structural difference identification, and intelligent signal classification, specifically:

[0049] A jump identification path is constructed to identify the frequency changes, amplitude jump patterns, and time discontinuity trends of the abrupt changes contained in the decoupling enhancement features, and to generate jump structure feature vectors.

[0050] Construct a stable recognition path, extract the periodic spectral distribution, continuous time structure and global stable trend contained in the decoupling enhancement features, and generate a stable structure feature vector;

[0051] The jump recognition path, stable recognition path, or weighted fusion path of jump recognition path and stable recognition path is activated according to the structure type label. The weighted fusion path calculates the fusion weight through the structure weight generation function and performs a weighting operation on the jump structure feature vector and the stable structure feature vector to generate a fusion feature vector.

[0052] A label-guided fusion decoder is constructed, which includes a label embedding unit, a feature concatenation unit, and a multi-channel output unit. The label embedding unit encodes the structural type label into an embedding vector. The feature concatenation unit performs a concatenation operation on the fused feature vector and the embedding vector. The multi-channel output unit outputs the signal category label, structural feature label, and confidence score, respectively.

[0053] Preferably, a deep learning-based intelligent recognition system for drone signals includes the following modules:

[0054] The signal processing module is used to collect communication signals of the UAV under different environments and flight conditions, construct a time-domain signal sample sequence, perform amplitude normalization and time length alignment processing, and generate a standardized signal dataset.

[0055] The window construction module is used to perform local change rate calculation on the standardized signal dataset, generate a change intensity index based on the signal amplitude change rate, construct dynamic adjustment rules according to the change intensity index, determine the starting position and length of the analysis window according to the dynamic adjustment rules, and generate an adaptive window structure.

[0056] The spectrum generation module is used to perform spectrum transformation operations on the signal applying the adaptive window structure, generate a spectrum sequence, calculate the spectrum center change value of adjacent windows, and generate a spectrum drift index; if the spectrum drift index is greater than a preset drift threshold in a continuous time period, then window shrinkage update is performed; if the spectrum drift index is less than the drift threshold in a continuous time period and the number of fluctuations in a preset time interval is greater than a preset fluctuation number threshold, then the window structure remains unchanged, forming a spectrum sequence.

[0057] The deep recognition module is used to input the spectrogram sequence into the improved CBAM-CNN model and output signal category labels, structural feature labels and confidence scores. The improved CBAM-CNN model includes a spectrogram feature extraction channel, a frequency-time decoupling coding structure, a structural label prediction channel, a pseudo-label adversarial structure, a recognition path selection structure, a jump recognition path, a stable recognition path and a label-guided fusion decoder.

[0058] Preferably, a deep learning-based intelligent identification storage medium for drone signals stores executable instructions that can be read by a computer processor.

[0059] The beneficial effects of this invention are:

[0060] This invention addresses the spectral abruptness, structural hybridity, and non-stationarity issues of UAV communication signals in complex environments by constructing a multi-level structure comprising a signal processing module, a window construction module, a spectrogram generation module, and a deep recognition module. In the signal preprocessing stage, amplitude normalization and time length alignment mechanisms are introduced to unify the temporal structure and amplitude range of multi-state signal samples. An adaptive window structure is constructed by combining local amplitude change rate calculation and dynamic adjustment rules, enabling the analysis window to sensitively locate abrupt signal segments. Based on the adaptive window structure, a short-time Fourier transform is performed to generate a spectrogram sequence and construct a spectral drift index. The window structure is dynamically updated by setting drift and fluctuation thresholds, improving the spectrogram construction's responsiveness to drift patterns. In the deep recognition stage, an improved CBAM-CNN model is used to decouple the frequency and time of the spectrogram sequence, constructing frequency and time vectors. Decoupling enhancement features are generated through an interactive attention mechanism. These features are then passed sequentially through a structure label prediction channel, a pseudo-label adversarial structure, a recognition path selection structure, and a label-guided fusion decoder to accurately distinguish between abrupt and stable structures and dynamically select recognition paths, enhancing the model's adaptability to signal structural differences. The final output includes signal category labels, structural feature labels, and confidence scores, which significantly improve the structural interpretability and classification reliability of signal recognition. It also exhibits strong robustness and generalizability, meeting the practical needs of intelligent UAV signal recognition in dynamic airspace environments. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a flowchart of a deep learning-based intelligent signal recognition method for unmanned aerial vehicles (UAVs) proposed in this invention.

[0063] Figure 2 This is a schematic diagram of the improved CBAM-CNN model proposed in this invention;

[0064] Figure 3 This is a schematic diagram of a deep learning-based intelligent recognition system for drone signals proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1-3 A deep learning-based intelligent recognition method for drone signals includes the following steps:

[0067] Step 1: Collect communication signals from the UAV under different environments and flight conditions, construct a time-domain signal sample sequence, perform amplitude normalization and time length alignment processing, and generate a standardized signal dataset;

[0068] Step 2: Perform local change rate calculation on the standardized signal dataset, generate a change intensity index based on the signal amplitude change rate, construct dynamic adjustment rules according to the change intensity index, determine the starting position and length of the analysis window according to the dynamic adjustment rules, and generate an adaptive window structure.

[0069] Step 3: Perform a spectrum transformation operation on the signal using the adaptive window structure to generate a spectrum sequence, calculate the spectrum center change value of adjacent windows, and generate a spectrum drift index.

[0070] Step 4: If the spectrum drift index is greater than the preset drift threshold within a continuous time period, then perform window shrinkage update; if the spectrum drift index is less than the drift threshold within a continuous time period and the number of fluctuations within the preset time interval is greater than the preset number of fluctuations threshold, then keep the window structure unchanged and form a spectrum sequence.

[0071] Step 5: Input the spectrum sequence into the improved CBAM-CNN model and output the signal category label, structural feature label and confidence score; the improved CBAM-CNN model includes a spectrum feature extraction channel, a frequency-time decoupling coding structure, a structural label prediction channel, a pseudo-label adversarial structure, a recognition path selection structure, a jump recognition path, a stable recognition path and a label-guided fusion decoder.

[0072] This implementation collects communication signals from UAVs under different environments and flight states, constructs a time-domain signal sample sequence, and performs amplitude normalization and time length alignment processing to generate a standardized signal dataset. This solves the problems of inconsistent amplitude distribution and inconsistent sampling lengths in the original communication signals, improving the standardization and comparability of data input. Furthermore, by calculating the local rate of change and constructing a change intensity index, and setting dynamic adjustment rules based on the change intensity index, a window structure with time-adaptive capabilities is generated. This allows the analysis window to accurately locate regions of drastic signal changes, improving the targeting and sensitivity of time-frequency analysis. Further, based on the adaptive window structure, a spectrum transformation is performed to construct a spectrum sequence, and adjacent windows are calculated... The change value of the spectral center of the port generates a spectral drift index, enabling continuous monitoring of the evolution trend of signal frequency distribution. Simultaneously, based on the changing characteristics of the spectral drift index, the system dynamically performs window shrinking updates or maintains the window structure, enhancing its ability to identify abrupt and stationary signals. Finally, the spectrogram sequence is input into an improved CBAM-CNN model, which utilizes the spectral feature extraction channel and frequency-time decoupling coding structure to extract joint time-frequency features. Combined with the structure label prediction channel, pseudo-label adversarial structure, and path selection structure, it achieves differentiated identification of abrupt and stable structures, outputting high-confidence signal category labels and structural feature labels. This improves the accuracy of intelligent identification and structural interpretation capabilities of UAV communication signals in complex scenarios.

[0073] In this embodiment, step one specifically includes:

[0074] The system collects communication signals from the UAV during takeoff, cruise, hovering, and landing phases. These communication signals include control signals, image transmission signals, and navigation signals. The control signals include remote controller command encoding signals and flight control feedback signals. The image transmission signals include encoded video streams and assisted positioning image streams. The navigation signals include GNSS navigation data, inertial navigation information, and integrated navigation state vectors.

[0075] The communication signal is continuously sampled according to a set time sampling period; the time sampling period is determined based on the communication link rate and signal bandwidth, and a time-domain signal sample sequence is obtained by using an equal-interval resampling strategy; the amplitude, frequency, phase and time index of each sampling point are recorded in the time-domain signal sample sequence.

[0076] Amplitude normalization is performed on the time-domain signal sample sequence. The amplitude normalization process uses a minimum-maximum normalization method and performs a linear mapping operation. Time length alignment is performed on the time-domain signal sample sequence. The time length alignment process is based on a unified sampling frequency and a unified time reference. For signal sequences with insufficient length, linear interpolation is performed to fill in the gaps. For sequences with excessive length, window pruning is performed. The interpolation fill in the gaps uses a first-order linear interpolation function based on the time index. The window pruning selects sampling segments within the target time range using a center-aligned method.

[0077] The processed time-domain signal sample sequences are labeled with multiple tags; flight status tags include four categories: takeoff, cruise, hovering, and landing; environmental tags include five categories: city, mountainous area, sea surface, forest, and open field; signal type tags are three categories: control signal, image transmission signal, and navigation signal.

[0078] The labeled time-domain signal sample sequences are constructed into a standardized signal dataset. The standardized signal dataset structure includes a sample index, time series, normalized amplitude sequence, sampling frequency, original signal type, flight state label, environment label, and unified time reference index, which are used for subsequent feature extraction and model training.

[0079] In this embodiment, step two specifically includes:

[0080] Each time-domain signal sample sequence in the standardized signal dataset is divided into continuous signal segments at equal intervals; the division process is performed based on a uniform sampling frequency and a minimum segment length, and the start time index and end time index of each signal segment are recorded;

[0081] For each signal segment, the amplitude change rate is calculated; the amplitude change rate is obtained by calculating the ratio of the amplitude difference between adjacent sampling points to the time difference, generating a change rate sequence; the change rate sequence is arranged in the same time order as the original sequence, and each change rate value corresponds to a time index;

[0082] A sliding window structure is constructed on the rate of change sequence; the sliding window length and sliding step size are set according to the set local statistical range and signal density; local statistical operations are performed in each sliding window to calculate the maximum rate of change, the average rate of change, and the fluctuation amplitude; the fluctuation amplitude is obtained by weighted calculation of the range of the rate of change within the window, and the maximum rate of change and the average rate of change are obtained by statistical averaging of the sampled values;

[0083] The local statistical indicators within each sliding window are combined to form the change intensity index; the change intensity index represents the degree of drastic change and fluctuation trend of the signal within the current time interval.

[0084] Multiple adjustment level thresholds are set based on the intensity of change index. The adjustment level thresholds are obtained by fitting a stratified interval division based on standard deviation. Five intensity of change levels are set, corresponding to five signal change scenarios from "very weak" to "very strong". Dynamic adjustment rules for the analysis window are constructed. The dynamic adjustment rules define the window length, sliding step size and maximum extension range corresponding to different intensity of change levels.

[0085] Based on the intensity level of the change at the starting position of the analysis window, the corresponding window length and sliding step size are determined by looking up a table; the corresponding signal segments are extracted by combining the starting position index and the set length, forming an analysis window structure with variable scale;

[0086] Extract signal segments from all analysis windows sequentially according to the total start position index and matching window length; construct an adaptive window structure; each window in the adaptive window structure contains a time-domain signal segment, a corresponding change level label, and start and end time indices; generate a window segmented signal dataset with time positioning information.

[0087] In this embodiment, step three specifically includes:

[0088] A short-time Fourier transform operation is performed on each window segment in the adaptive window structure. The short-time Fourier transform is a sliding window transformation based on a set window function and overlap length. Each transformation result generates a two-dimensional spectrum containing information in the frequency and time dimensions. Each pixel in the two-dimensional spectrum represents the amplitude intensity at the corresponding time and frequency positions.

[0089] The two-dimensional spectrograms are stacked sequentially according to the time order in the window structure to construct a spectrogram sequence; each image in the spectrogram sequence is accompanied by a time index, window number and change intensity level;

[0090] For the spectrum graphs corresponding to two adjacent windows, the main distribution interval of spectral energy is extracted; the main distribution interval is truncated by calculating the cumulative energy ratio in the frequency dimension and setting an energy ratio threshold to extract frequency segments;

[0091] Calculate the location of the spectral energy center within the main distribution interval; the location of the spectral energy center is the weighted average frequency, which is obtained by weighting the frequency location with the corresponding spectral intensity, and represents the frequency centroid of the spectrum.

[0092] The difference between the positions of the spectral energy centers corresponding to the spectrograms of two adjacent windows is calculated to obtain the change value of the spectral center; the change value of the spectral center represents the magnitude of the shift of the frequency centroid per unit time.

[0093] All spectral center change values ​​are combined in chronological order to form a spectral drift index sequence; each element in the spectral drift index sequence includes the change value, change direction and drift amplitude, and records the time index and frequency change status;

[0094] The data at each time point in the spectrum drift index sequence can be regarded as a multi-dimensional feature vector, which is then input into the drift structure judgment module for window structure update judgment.

[0095] In this embodiment, step four specifically includes:

[0096] Set a spectral drift threshold and a spectral fluctuation number threshold. The spectral drift threshold is obtained by statistical analysis of historical spectral drift index samples. The upper limit of the drift threshold is determined by a three-standard-deviation interval fitting method. The spectral fluctuation number threshold is determined based on the statistical distribution of the direction of change of the spectral center within a unit of time, and serves as the limit for the number of changes in the spectral stability judgment. Define a continuous time period and a preset time interval. The continuous time period represents the length of time during which the drift index maintains stable growth or decay, and the preset time interval represents the time span of the sliding judgment range.

[0097] A threshold comparison operation is performed on the drift amplitude values ​​within a continuous time period in the spectral drift index sequence. When all drift amplitude values ​​are greater than the spectral drift threshold, it is determined that the spectral center has a continuous shift trend within that time period. A window shrinkage operation is performed on the analysis window within the corresponding time period. The window shrinkage operation includes shortening the window length, updating the start and end positions, reconstructing the window number, and synchronously updating the spectral sequence. The window length shortening ratio is calculated based on the consistency between the average drift amplitude and the drift direction, and a mapping relationship between the drift amplitude and the shrinkage ratio is established through a linear fitting function.

[0098] If the drift amplitude values ​​of the spectral drift index sequence are all less than the spectral drift threshold within the preset time interval, and the number of changes in the spectral drift direction within the time interval is greater than the spectral fluctuation number threshold, then the signal is judged to be in a local fluctuation state; the starting position and length of the analysis window are kept unchanged to maintain the continuity of the spectral timing.

[0099] The set of windows that are either shrunk or remain unchanged is used as the final set of analysis windows; each window in the final set of analysis windows retains its time index, window number, and status identifier; the spectrum transformation operation is called sequentially to generate an updated spectrum sequence.

[0100] In this embodiment, the spectral feature extraction channel extracts spatial features of the spectrogram sequence based on a convolutional structure and generates a feature tensor. The frequency-time decoupling coding structure performs frequency and time dimension separation calculations on the feature tensor to generate frequency and time vectors, and fuses them through an interactive attention mechanism to form decoupling enhanced features, specifically:

[0101] A spectral feature extraction channel is constructed, which includes a multi-layer structure composed of two-dimensional convolutional units, batch normalization units, and nonlinear activation units. After inputting a sequence of spectrograms, convolutional operations are performed sequentially to extract frequency difference features in local regions of the spectrograms. Normalization units are called to balance the feature distribution, and activation units are called to enhance the nonlinear expressive power. Through multi-layer feature stacking and scale compression operations, a feature tensor containing frequency, time, and channel dimensions is output.

[0102] A frequency-time decoupled coding structure is constructed. In the frequency branch coding unit, a one-dimensional convolution operation is performed along the frequency dimension of the feature tensor to extract the local energy changes of each frequency band. The global pooling structure is called to compress the frequency features and generate a frequency vector. In the time branch coding unit, a sliding convolution is performed along the time dimension to extract the time series change trend. The weight values ​​of each time point are generated by combining the attention scoring module and normalized to form a time vector.

[0103] An interactive attention mechanism is constructed by building frequency attention matrices and time attention matrices using frequency vectors and time vectors, respectively. The frequency attention matrix is ​​called to adjust the response intensity of each dimension in the frequency vector, and the time attention matrix is ​​called to adjust the contribution distribution of each time segment in the time vector. A bidirectional weighting process is executed to achieve cross-fusion between the frequency vector and the time attention matrix, and between the time vector and the frequency attention matrix, through matrix multiplication, to generate bidirectional enhanced feature representations.

[0104] The frequency vector and time vector are concatenated and fused. The feature mapping unit is called to perform dimensional compression and nonlinear transformation on the fused vector, and the output is a decoupled enhanced feature containing frequency structure information and time dynamic information.

[0105] In this embodiment, the structure label prediction channel receives the decoupled enhancement features and outputs structure type labels by combining the channel attention structure and the spatial attention structure. The structure type labels include jump structures, stable structures, and hybrid structures. The pseudo-label adversarial structure outputs pseudo-structure labels based on the auxiliary generation network. The weight distribution of the structure label prediction channel is then corrected by comparing the structure type label generation adversarial loss with the pseudo-structure labels. Specifically:

[0106] A channel attention structure is constructed. After inputting the decoupled enhancement features, global average pooling and global max pooling operations are performed to extract channel statistics, generating two channel description vectors. The channel description vectors are then input into a shared multilayer perceptron structure, where nonlinear transformation and compressed activation operations are performed to output channel attention weight vectors. After expanding the channel attention weight vectors to match the dimension of the decoupled enhancement features, a channel-wise multiplication operation is performed to form a channel-weighted feature map.

[0107] A spatial attention structure is constructed. After inputting the channel-weighted feature map, max pooling and average pooling operations are performed along the channel dimension. Stacked pooling results generate a two-dimensional feature map. The two-dimensional feature map is input into a two-dimensional convolutional unit to extract local spatial structure features and output a spatial attention weight map. The Sigmoid function is called to normalize the weight map and then perform a pixel-wise multiplication operation with the channel-weighted feature map to form an attention enhancement feature map.

[0108] The attention-enhanced feature map is input into the structure classifier, which includes a fully connected layer, a normalization layer, and a softmax output layer. The flatten operation is called to unfold the feature map, and the result is input into the fully connected layer to extract the structural feature representation. The structural feature representation is then input into the softmax classifier to output the structure type label, which includes transition structures, stable structures, and hybrid structures. This label serves as the input for subsequent recognition path selection and label-guided fusion decoder.

[0109] A pseudo-label adversarial structure is constructed, which includes an auxiliary generation network and a discriminant unit. The auxiliary generation network takes a random noise vector obtained by sampling from a Gaussian distribution and a feature vector encoded by the intermediate layer of the structure classifier as input, and outputs pseudo-structure labels through a multi-layer generative mapping structure. The discriminant unit takes the real structure type label and the pseudo-structure label as input, and calculates the structure consistency discrimination probability through a convolutional extraction layer and a discriminant function.

[0110] Construct an adversarial loss function, call the cross-entropy loss function to calculate the distribution difference between the structure type label and the pseudo structure label, and generate the structure adversarial loss value; call the backpropagation algorithm to backpropagate the structure adversarial loss value to the structure label prediction channel, update the channel attention structure, spatial attention structure and trainable parameters in the structure classifier, and complete the dynamic correction of the structure label weight.

[0111] In this embodiment, the identification path selection structure activates the jump identification path, stable identification path, or a weighted fusion path of both based on the structure type label, generating a fused feature vector; the label-guided fusion decoder receives the fused feature vector and the structure type label, and outputs the UAV signal identification result, structural interpretation information, and confidence distribution, realizing time-frequency feature decoupling, structural difference identification, and intelligent signal classification, specifically:

[0112] A jump detection path is constructed, with decoupled enhancement features as input. A local mutation feature extraction unit consisting of a multi-layer convolutional module, a local response normalization module, and an activation module is set in the jump detection path to extract frequency mutation points, amplitude jump edges, and high-frequency change regions in a short period of time, and output a jump response map. The jump response map is input into the global trend modeling module, and the non-stationary time trend is modeled in combination with a gated loop structure to generate a jump structure feature vector.

[0113] A stable recognition path is constructed, with the input being decoupled enhancement features. A multi-scale convolutional kernel structure and a periodic attention mechanism are set in the stable recognition path to extract the main frequency distribution, mean trend, and stable segment feature map within a long time window. In the global stability analysis module, a spectral trend regression function is constructed to generate a stable structure feature vector by fitting the frequency drift and amplitude fluctuation range of the stable frequency band.

[0114] The logic for selecting the recognition path is controlled by setting the structure type label. When the structure type label is a transitional structure, only the transitional recognition path is activated. When the structure type label is a stable structure, only the stable recognition path is activated. When the structure type label is a hybrid structure, a weighted fusion path of the transitional and stable recognition paths is activated. The weighted fusion path calls the structure weight generation function, calculates the weight vector based on the label embedding vector and feature distribution parameters, multiplies it by the transitional structure feature vector and the stable structure feature vector respectively, performs vector weighting operation, and outputs the fused feature vector.

[0115] A label-guided fusion decoder is constructed, comprising a label embedding unit, a feature concatenation unit, and a multi-channel output unit. The label embedding unit calls an embedding mapping network to convert the structural type label into a low-dimensional embedding vector. The feature concatenation unit performs a tensor concatenation operation on the fused feature vector and the embedding vector to construct an enhanced decoding input tensor. The multi-channel output unit contains three parallel decoding paths, which are used to output signal category labels, structural feature labels, and confidence scores, respectively. The signal category label reflects the functional type of the communication signal, the structural feature label describes the structural morphology of the signal, and the confidence score is calculated using a fully connected layer and a sigmoid function to determine the confidence distribution of the corresponding output.

[0116] In this embodiment, a deep learning-based intelligent recognition system for drone signals includes the following modules:

[0117] The signal processing module is used to collect communication signals of the UAV under different environments and flight conditions, construct a time-domain signal sample sequence, perform amplitude normalization and time length alignment processing, and generate a standardized signal dataset.

[0118] The window construction module is used to perform local change rate calculation on the standardized signal dataset, generate a change intensity index based on the signal amplitude change rate, construct dynamic adjustment rules according to the change intensity index, determine the starting position and length of the analysis window according to the dynamic adjustment rules, and generate an adaptive window structure.

[0119] The spectrum generation module is used to perform spectrum transformation operations on the signal applying the adaptive window structure, generate a spectrum sequence, calculate the spectrum center change value of adjacent windows, and generate a spectrum drift index; if the spectrum drift index is greater than a preset drift threshold in a continuous time period, then window shrinkage update is performed; if the spectrum drift index is less than the drift threshold in a continuous time period and the number of fluctuations in a preset time interval is greater than a preset fluctuation number threshold, then the window structure remains unchanged, forming a spectrum sequence.

[0120] The deep recognition module is used to input the spectrogram sequence into the improved CBAM-CNN model and output signal category labels, structural feature labels and confidence scores. The improved CBAM-CNN model includes a spectrogram feature extraction channel, a frequency-time decoupling coding structure, a structural label prediction channel, a pseudo-label adversarial structure, a recognition path selection structure, a jump recognition path, a stable recognition path and a label-guided fusion decoder.

[0121] In this embodiment, a deep learning-based intelligent identification storage medium for drone signals is provided, wherein the storage medium stores executable instructions that are read by a computer processor.

[0122] Example 1:

[0123] To verify the feasibility of this invention in practice, it was applied to the intelligent monitoring system for unmanned aerial vehicles (UAVs) in a power inspection unit. The focus of the test was on the system's accuracy, response speed, and anti-interference capability in identifying different types of UAV signals.

[0124] The scenario selected is a complex mountainous power transmission corridor, characterized by intricate terrain, drastic climate changes, and significant signal interference. Previously, the organization deployed six multi-rotor drones in this area for routine power cable inspections. These drones belonged to three different suppliers and varied in model and communication protocol. Previously, traditional signal analysis systems were unable to accurately identify certain low-power or rapidly changing control signals, frequently resulting in signal identification failures, misclassifications, or status misjudgments, impacting scheduling efficiency and safety assurance capabilities.

[0125] Before applying the method of this invention, the system performed signal discrimination based on the traditional FFT + CNN method with a static window width. After deploying the improved CBAM-CNN model proposed in this invention, adaptive windows, structure type label prediction, jump recognition paths, and label-guided fusion decoders are introduced, which effectively improves the signal structure understanding and complex signal state recognition capabilities.

[0126] To evaluate the performance of the two schemes, a three-month comparative analysis was conducted on the signal acquisition and discrimination results of UAVs in typical cruise missions. A total of 12,820 signal samples were collected, including control signals, navigation signals, and image transmission signals. The test indicators included signal recognition accuracy, average classification confidence, false alarm rate, and average response delay. The statistical data are shown in the table below.

[0127] Table 1. Performance Comparison of Traditional Methods and the Method of the Present Invention in UAV Signal Recognition

[0128] Test metrics Traditional FFT+CNN method Method of the present invention Average recognition accuracy 91.7% 98.3% Average classification confidence 0.73 0.91 Signal false alarm rate 6.8% 1.4% Average response time (ms) 223 135

[0129] As shown in Table 1, in actual operation, the method of this invention significantly outperforms traditional methods in signal recognition accuracy, improving it by 6.6 percentage points. The classification confidence score increased from 0.73 to 0.91, indicating that the model's judgment of signals is more stable and confident. In mission-critical scenarios, the false alarm rate decreased from 6.8% to 1.4%, greatly reducing the risk of misoperation caused by misidentification in the scheduling system. Furthermore, the average response latency was shortened from 223 milliseconds to 135 milliseconds, providing stronger real-time performance in complex environments with simultaneous multi-target identification, meeting engineering deployment requirements.

[0130] To further verify the system's recognition performance on complex transition signals, 3900 transition signal samples were collected and labeled, including abrupt control signals, short-term interference navigation signals, and frequency drift image signals. Within the same three-month period, indicators such as false positive rate, structural recognition consistency, transition stability detection capability, and number of repeated false positives were recorded. The results are shown in the table below.

[0131] Table 2 Comparison of Signal Recognition Performance of Jump-Type UAVs

[0132] Test metrics Traditional FFT+CNN method Method of the present invention Jump signal misjudgment rate 11.2% 2.6% Structural identification consistency 82.5% 94.9% Jump signal detection stability 74.3% 91.7% Number of repeated false positives 9.7 1.2

[0133] Table 2 shows that the method of this invention also exhibits significant advantages in the task of identifying transition signals. Traditional methods suffer from a high false positive rate of 11.2% in the identification of transition signals due to fixed windows and poor model adaptation to unstable structures. This invention, however, reduces the false positive rate to 2.6% through structural label prediction and path selection mechanisms. The consistency of structural identification is improved by 12.4%, demonstrating a stronger ability to understand signal structures. The stability rate of transition detection reaches 91.7%, exhibiting good continuity and robustness in continuous anomaly detection scenarios. The number of repeated false detections decreases from an average of 9.7 per 100 signals to 1.2, greatly reducing false alarm interference in the monitoring system.

[0134] This embodiment significantly improves upon the shortcomings of traditional methods in complex signal recognition in real-world scenarios. By introducing five major technical structures—an improved CBAM-CNN model, an adaptive window mechanism, a structural label prediction channel, a jump / stable recognition path, and a label-guided fusion decoder—it achieves comprehensive improvements in accuracy, robustness, and timeliness. In particular, it demonstrates more stable and reliable recognition performance in signal environments with complex structures or severe jumps, making it highly valuable for engineering applications.

[0135] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent recognition of drone signals based on deep learning, characterized in that, Includes the following steps: Step 1: Collect communication signals from the UAV under different environments and flight conditions, construct a time-domain signal sample sequence, perform amplitude normalization and time length alignment processing, and generate a standardized signal dataset; Step 2: Perform local change rate calculation on the standardized signal dataset, generate a change intensity index based on the signal amplitude change rate, construct dynamic adjustment rules according to the change intensity index, determine the starting position and length of the analysis window according to the dynamic adjustment rules, and generate an adaptive window structure. Step 3: Perform a spectrum transformation operation on the signal using the adaptive window structure to generate a spectrum sequence, calculate the spectrum center change value of adjacent windows, and generate a spectrum drift index. Step 4: If the spectrum drift index is greater than the preset drift threshold within a continuous time period, then perform window shrinkage update; if the spectrum drift index is less than the drift threshold within a continuous time period and the number of fluctuations within the preset time interval is greater than the preset number of fluctuations threshold, then keep the window structure unchanged and form a spectrum sequence. Step 5: Input the spectrum sequence into the improved CBAM-CNN model and output the signal category label, structural feature label and confidence score; the improved CBAM-CNN model includes a spectrum feature extraction channel, a frequency-time decoupling coding structure, a structural label prediction channel, a pseudo-label adversarial structure, a recognition path selection structure, a jump recognition path, a stable recognition path and a label-guided fusion decoder.

2. The method for intelligent recognition of UAV signals based on deep learning according to claim 1, characterized in that, Step one specifically involves: The communication signals of the UAV during the takeoff, cruise, hovering and landing phases are collected, including control signals, image transmission signals and navigation signals. The communication signal is continuously sampled according to the set time sampling period to construct a time-domain signal sample sequence with time order, and the amplitude, frequency, phase and time index of each sampling point are recorded. The amplitude normalization process is performed on the time-domain signal sample sequence, and a linear mapping operation is performed based on the set upper and lower amplitude limits; The time-domain signal sample sequence is subjected to time length alignment processing, and time-domain signal sample sequences of different lengths are interpolated and padded or truncated according to a unified sampling frequency and time reference. The processed time-domain signal sample sequences are labeled according to flight status, environment, and signal type labels to generate a standardized signal dataset.

3. The method for intelligent recognition of UAV signals based on deep learning according to claim 2, characterized in that, Step two specifically involves: Each time-domain signal sample sequence in the standardized signal dataset is divided into continuous signal segments; For each signal segment, the rate of change of amplitude is calculated, and a rate of change sequence is generated based on the amplitude difference and time difference between adjacent sampling points; A sliding window structure is constructed on the rate of change sequence, and the maximum rate of change, average rate of change, and fluctuation amplitude within each sliding window are calculated to generate a change intensity index. Multiple adjustment level thresholds are set based on the intensity of change index, and dynamic adjustment rules for the analysis window are established. Based on the intensity level of change at the starting position of the analysis window, set the corresponding window length and sliding step size; Signal segments are extracted sequentially according to the starting position and corresponding window length of the analysis window, an adaptive window structure is constructed, and a window segmented signal dataset with time positioning information is generated.

4. The method for intelligent recognition of UAV signals based on deep learning according to claim 3, characterized in that, Step three specifically involves: Perform a short-time Fourier transform operation on each window segment in the adaptive window structure to generate a two-dimensional spectrogram containing information in both the frequency and time dimensions; The two-dimensional spectrograms are stacked sequentially according to the window order to construct a spectrogram sequence, and the time index and window number of each spectrogram are labeled. For the spectrum diagrams corresponding to two adjacent windows, extract the main frequency distribution interval, calculate the position difference of the spectrum energy center, and generate the spectrum center change value; The change values ​​of the spectral center of each pair of adjacent spectrograms are combined into a time series to form a spectral drift index. Record the change value, direction of change and drift magnitude of the spectrum drift index at each time point to construct an indexed spectrum drift index sequence.

5. The method for intelligent recognition of UAV signals based on deep learning according to claim 4, characterized in that, Step four specifically involves: Set the spectrum drift threshold and the spectrum fluctuation number threshold, and define the continuous time period and the preset time interval; A threshold comparison operation is performed on the drift amplitude values ​​in the spectrum drift index sequence within a continuous time period. If all drift amplitude values ​​are greater than the spectrum drift threshold, a window shrinking operation is performed on the analysis window within the corresponding time period. The window shrinking operation includes shortening the window length, updating the start and end positions, reconstructing the corresponding window number, and synchronously updating the spectrogram sequence. If the drift amplitude of the spectrum drift index sequence within the preset time interval is less than the spectrum drift threshold, and the number of times the spectrum drift direction changes within the time interval is greater than the number of times the spectrum fluctuates, then the starting position and length of the current analysis window will remain unchanged. The set of windows that are either shrunk or remain unchanged is used as the final set of analysis windows, and the spectrum transformation operation is called sequentially to generate a spectrum sequence.

6. The method for intelligent recognition of UAV signals based on deep learning according to claim 5, characterized in that, The spectral feature extraction channel extracts spatial features from the spectrogram sequence based on a convolutional structure and generates a feature tensor. The frequency-time decoupling coding structure performs frequency and time dimension separation calculations on the feature tensor to generate frequency and time vectors, which are then fused through an interactive attention mechanism to form decoupling enhanced features. Specifically: A spectral feature extraction channel consisting of multi-layer two-dimensional convolutional units, batch normalization units, and nonlinear activation units is constructed. Convolution, feature aggregation, and scaling operations are sequentially performed on the input spectrogram sequence to extract frequency distribution features and temporal evolution features, generating a multidimensional feature tensor. In the frequency-time decoupled coding structure, frequency branch coding units and time branch coding units are set up. The frequency branch coding unit performs one-dimensional convolution and global pooling operations along the frequency dimension to extract the local energy distribution of each frequency band and generate a frequency vector. The time branch coding unit performs sliding convolution and attention weight calculation operations along the time dimension to extract the time series change trend and generate a time vector. In the interactive attention mechanism, a frequency attention matrix and a time attention matrix are introduced, and the frequency vector and time vector are weighted and adjusted respectively. A bidirectional weighted feature representation is formed by matrix multiplication. The weighted frequency vector and time vector are concatenated and fused, and then input into the multi-layer feature mapping unit to generate decoupled enhanced features that contain frequency information and temporal correlation.

7. The method for intelligent recognition of UAV signals based on deep learning according to claim 6, characterized in that, The structure label prediction channel receives the decoupled enhancement features and outputs structure type labels by combining the channel attention structure and the spatial attention structure. These structure type labels include jump structures, stable structures, and hybrid structures. The pseudo-label adversarial structure outputs pseudo-structure labels based on an auxiliary generation network. The weight distribution of the structure label prediction channel is then corrected by comparing the structure type label generation adversarial loss with the output pseudo-structure labels. Specifically: Construct a channel attention structure, perform global average pooling and global max pooling operations on the decoupled enhancement features, generate channel description vectors, perform attention weight calculation, generate channel attention weights, and perform channel-by-channel multiplication operations on the channel attention weights and decoupled enhancement features; A spatial attention structure is constructed. Two-dimensional convolution and pooling operations are performed on the channel-weighted feature map to extract the spatial attention feature map and generate a spatial attention weight map. The spatial attention weight map and the channel-weighted feature map are multiplied pixel by pixel to form an attention enhancement feature map. The attention-enhanced feature map is input into the structure classifier, and the structure type label is output through a fully connected layer and Softmax. The structure type label includes jumping structure, stable structure and hybrid structure. A pseudo-label adversarial structure is constructed, which includes an auxiliary generation network and a discriminant unit. The auxiliary generation network generates pseudo-structure labels by taking random noise and encoded features as input. The discriminant unit simultaneously receives structure type labels and pseudo-structure labels and outputs the discrimination probability. Based on the discriminant probability, an adversarial loss function is constructed. The difference between the distribution of real labels and pseudo labels is compared. Gradient backpropagation is performed to update the network parameters in the structural label prediction channel and adjust the weight distribution.

8. The method for intelligent recognition of UAV signals based on deep learning according to claim 7, characterized in that, The identification path selection structure activates the jump identification path, stable identification path, or a weighted fusion path of both based on the structure type label, generating a fused feature vector. The label-guided fusion decoder receives the fused feature vector and the structure type label, and outputs the UAV signal identification result, structural interpretation information, and confidence distribution, achieving time-frequency feature decoupling, structural difference identification, and intelligent signal classification. Specifically: A jump identification path is constructed to identify the frequency changes, amplitude jump patterns, and time discontinuity trends of the abrupt changes contained in the decoupling enhancement features, and to generate jump structure feature vectors. Construct a stable recognition path, extract the periodic spectral distribution, continuous time structure and global stable trend contained in the decoupling enhancement features, and generate a stable structure feature vector; The jump recognition path, stable recognition path, or weighted fusion path of jump recognition path and stable recognition path is activated according to the structure type label. The weighted fusion path calculates the fusion weight through the structure weight generation function and performs a weighting operation on the jump structure feature vector and the stable structure feature vector to generate a fusion feature vector. A label-guided fusion decoder is constructed, which includes a label embedding unit, a feature concatenation unit, and a multi-channel output unit. The label embedding unit encodes the structural type label into an embedding vector. The feature concatenation unit performs a concatenation operation on the fused feature vector and the embedding vector. The multi-channel output unit outputs the signal category label, structural feature label, and confidence score, respectively.

9. A deep learning-based intelligent drone signal recognition system, comprising executing the deep learning-based intelligent drone signal recognition method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The signal processing module is used to collect communication signals of the UAV under different environments and flight conditions, construct a time-domain signal sample sequence, perform amplitude normalization and time length alignment processing, and generate a standardized signal dataset. The window construction module is used to perform local change rate calculation on the standardized signal dataset, generate a change intensity index based on the signal amplitude change rate, construct dynamic adjustment rules according to the change intensity index, determine the starting position and length of the analysis window according to the dynamic adjustment rules, and generate an adaptive window structure. The spectrum generation module is used to perform spectrum transformation operations on the signal applying the adaptive window structure, generate a spectrum sequence, calculate the spectrum center change value of adjacent windows, and generate a spectrum drift index; if the spectrum drift index is greater than a preset drift threshold in a continuous time period, then window shrinkage update is performed; if the spectrum drift index is less than the drift threshold in a continuous time period and the number of fluctuations in a preset time interval is greater than a preset fluctuation number threshold, then the window structure remains unchanged, forming a spectrum sequence. The deep recognition module is used to input the spectrogram sequence into the improved CBAM-CNN model and output signal category labels, structural feature labels and confidence scores. The improved CBAM-CNN model includes a spectrogram feature extraction channel, a frequency-time decoupling coding structure, a structural label prediction channel, a pseudo-label adversarial structure, a recognition path selection structure, a jump recognition path, a stable recognition path and a label-guided fusion decoder.

10. A storage medium for intelligent identification of drone signals based on deep learning, characterized in that, The storage medium stores executable instructions, which are read by a computer processor to execute the deep learning-based intelligent recognition method for drone signals as described in any one of claims 1 to 8.

Images