End-to-end unmanned aerial vehicle identification method and system based on original sound waves

Abstract

The invention discloses an end-to-end unmanned aerial vehicle identification method and system based on original sound waves, and the method comprises the steps: setting a plurality of microphone arrays in a target region, and collecting original waveform signals formed by different types of unmanned aerial vehicles in different states and environments; the collected original waveform signals are preprocessed; a 1D-CNN parallel multi-scale one-dimensional convolution feature extraction network is constructed, the same segment of preprocessed waveform is convolved by using short-time, medium-scale and long-time time windows, and local time domain features are generated; constructing a Transform model, and inputting the local time domain feature into a Transform network to generate a global voiceprint feature; and outputting the existence, model, distance and rotating speed information of the unmanned aerial vehicle according to the global voiceprint features. According to the method, a traditional time-frequency feature extraction method does not need to be adopted, the accuracy and the recognition robustness of unmanned aerial vehicle recognition in a strong noise environment are remarkably improved, and the method is suitable for scenes such as security monitoring, anti-unmanned aerial vehicle detection and early warning.

Description

Technical Field

[0001] This invention relates to the field of drone acoustic feature recognition technology, specifically to an end-to-end drone recognition method and system based on raw acoustic waves. Background Technology

[0002] Traditional drone detection technologies mainly include radar detection and optical detection. Radar detection detects targets by emitting electromagnetic waves and receiving their reflected waves, providing information such as the target's distance, speed, and orientation. However, radar systems are expensive, bulky, and difficult to deploy in some complex environments. Optical detection technology uses cameras or infrared devices to capture images of the drone, but this method is greatly affected by weather and lighting conditions, and performs poorly at night and in low-visibility environments.

[0003] Acoustic recognition technology offers a novel solution for identifying the acoustic features of drones. This is achieved by capturing and analyzing the unique sound signals (i.e., "voiceprints" or "audio fingerprints") generated during drone flight to enable detection, localization, and identification. The core of this approach lies in matching the drone's acoustic features with a pre-built database and combining signal processing and artificial intelligence algorithms for judgment. However, existing methods for identifying drones using acoustic features largely rely on methods such as time-frequency transformation (e.g., Mel-spectrogram, MFCC), traditional CNN or DNN classification models, and acoustic array localization combined with feature matching algorithms. These methods also have limitations: spectrum generation loses original time-domain information, reducing the model's performance under low signal-to-noise ratio conditions; CNNs tend to ignore the long-term modulation structure of drone propeller noise; and they require extensive preprocessing steps, resulting in poor adaptability. Summary of the Invention

[0004] The purpose of this invention is to provide an end-to-end UAV identification method and system based on raw sound waves to solve the problems mentioned in the background art.

[0005] The specific technical solution provided by this invention is as follows: An end-to-end UAV identification method based on raw sound waves, comprising the following operational steps:

[0006] Step S1: Set up multiple microphone arrays in the target area to collect the raw waveform signals generated by different types of drones under different states and environments.

[0007] Preferably, the data acquisition process includes: Step S11: The data acquisition equipment synchronizes with the UAV control system or positioning system in time; Step S12: Start the data acquisition device and initialize the sampling parameters, including sampling rate, number of channels, and gain; Step S13: Control the drone to fly in the designated area according to the preset trajectory and operating conditions; Step S14: The data acquisition device synchronously acquires the raw acoustic wave signals emitted by the UAV; Step S15: Label the collected data with tags including: drone model, propeller type, operating condition, flight speed, altitude, and distance, and store them in a metadata file, corresponding one-to-one with each audio segment.

[0008] Step S2: Perform preprocessing on the acquired raw waveform signal, including normalization and filtering.

[0009] Preferably, the preprocessing includes: segmenting the continuously acquired raw acoustic signal into segments according to a fixed window time length and a sliding step size, and calculating the length of each segment; shifting the acoustic signal upward or downward by a constant value to remove noise; applying short-time mean removal to the sliding window, dividing the window into one frame per second, and subtracting the frame mean from the value of each frame; filtering the input samples using a first-order high-pass filter to remove extremely low-frequency drift and device temperature drift; calculating the standard deviation for each signal segment, and then normalizing the signal amplitude.

[0010] Step S3: Construct a 1D-CNN parallel multi-scale one-dimensional convolutional feature extraction network, and use short-time, medium-scale and long-time time windows to convolve the same preprocessed waveform to generate local temporal features.

[0011] Preferably, the constructed 1D-CNN parallel multi-scale one-dimensional convolutional feature extraction network includes the following architecture: The first layer is a parallel convolutional bank, used for parallel multi-scale convolutional feature extraction. It contains four Conv1D branches with different parameters. The outputs of the four branches are concatenated and merged in the channel dimension to form a comprehensive feature representation. The second layer consists of multi-layer residual dilated convolutional blocks, which are used to stack four Bloc structures with different dilation coefficients on the comprehensive feature representation to expand the receptive field and perform voiceprint modeling for drones at different time scales. The third layer is hierarchical downsampling. Multi-scale features are extracted by gradually reducing the time dimension. Conv1D is used to downsample the feature sequence, halving the time length each time to achieve stride convolution compression. ResBlock is inserted after each downsampling to form four feature layers with different time scales.

[0012] Step S4: Construct a Transformer model by inputting local temporal features into the Transformer network to generate global voiceprint features.

[0013] Preferably, the architecture of the constructed Transformer model includes: Input embedding: After the audio data is converted into a feature vector, it is input into the encoder part of the Transformer, and the audio signal is encoded using a trained convolutional neural network or other feature extraction methods; Position encoding: Position encoding provides positional information to the Transformer model, enabling the Transformer model to understand the temporal dependencies in the signal; Self-attention mechanism: When processing audio data, the correlation between each time step is dynamically calculated based on the self-attention mechanism to capture long-distance dependencies in the signal; Multi-head self-attention mechanism: The Transformer model uses a multi-head self-attention mechanism, where each attention head is responsible for focusing on a different subset of information, and then the outputs of multiple heads are concatenated. Feedforward Neural Network: Each self-attention layer is followed by a feedforward neural network, which is responsible for performing nonlinear transformations on the features to further extract deeper features of the audio signal.

[0014] Step S5: Output information on the presence, model, distance, and rotation speed of the UAV based on the global acoustic signature characteristics. Preferably, the specific implementation includes: The probability of drone presence in each time window is output using the Sigmoid function, and voting / smoothing is performed on multiple windows within a set period. When the probability of drone presence exceeds the set threshold, the next step is initiated; otherwise, no drone is detected. The Softmax function is used to output the model category for each time window. The model with the highest probability among multiple model determinations is taken as the predicted model. If the probability is lower than the set threshold, it is determined to be an unknown model. The Softmax function is used to output the state of each time window. The state with the highest probability is selected in each time window. Voting / smoothing is performed on multiple windows over a period of time to obtain the drone state. The location of the UAV is determined by designing a direct regression angle θ or a multi-task loss function; The distance to the drone was determined using a regression model; Voting and exponential smoothing are performed on the judgment results of multiple consecutive time windows to conduct post-processing in the time dimension.

[0015] On the other hand, the present invention also provides an end-to-end drone identification system based on raw sound waves, which executes an end-to-end drone identification method based on raw sound waves, including: Acoustic signal acquisition module: used to acquire acoustic signals from the drone via a microphone array; Acoustic signal preprocessing module: used to preprocess the acquired UAV acoustic signals; Parallel multi-scale 1D-CNN temporal feature extraction module: used to convolve the same waveform using short-time, medium-scale, and long-time time windows, and multiple outputs are concatenated and merged in the channel dimension; Transformer global feature modeling module: It is used to input data from the parallel multi-scale 1D-CNN temporal feature extraction module and use the trained Transformer model to identify the type or flight status of the drone. Multi-task output module: Used to output inference results including whether there is a drone, drone model, flight status, direction, and distance.

[0016] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention directly uses the original time-domain sound wave as input, extracts local time-domain features through a multi-scale one-dimensional convolutional neural network (1D-CNN), and then uses a Transformer network to capture the long-term dependent patterns of the drone's acoustic signature, ultimately realizing drone presence detection, model classification, flight attitude recognition, direction estimation, and distance assessment. This invention does not require traditional time-frequency feature extraction (such as spectrograms and MFCC), significantly improving the accuracy and robustness of drone identification in noisy environments, and can be applied to scenarios such as security monitoring, anti-drone detection, and early warning. Attached Figure Description

[0017] Figure 1 This is a flowchart of the end-to-end UAV identification method based on raw sound waves provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the multi-scale one-dimensional convolutional neural network (1D-CNN) structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the Transformer model structure provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0019] Example 1: Combination Figures 1-3 As shown in the figure, the end-to-end UAV identification method based on raw sound waves described in this embodiment includes the following steps: Step S1: Acquire the original waveform signal of the sound wave.

[0020] In this embodiment, multiple microphone arrays are set up in the target area to collect sound wave signals from different angles, and then the collected raw signals are denoised, filtered and normalized.

[0021] For example, the specific implementation process includes: Data Acquisition Target and Environment Definition: This invention selects open outdoor locations, such as spacious playgrounds or test ranges, away from roads, crowds, and strong noise sources to minimize environmental interference. It collects raw acoustic signals from various drone models, including quadcopters, multi-rotors, and fixed-wing drones, under different operating conditions such as hovering, climbing, descending, constant speed flight, and acceleration / deceleration, while also considering different load conditions. The collected raw acoustic signals undergo targeted filtering to remove wind noise, rain noise, bird calls, and random noise, ensuring high-quality signal acquisition.

[0022] Acoustic signal acquisition hardware configuration: As a key component of the UAV acoustic monitoring system, the acoustic signal acquisition hardware mainly involves microphone selection and parameter configuration. This invention selects condenser microphones or MEMS microphone arrays, as these types of microphones offer high sensitivity and good frequency response. Since UAV propeller noise is generally concentrated in the tens of Hz to several kHz range, but higher harmonics extend to even higher frequencies, this invention sets the frequency response range to 20Hz~20kHz. Simultaneously, for acoustic signals, the sampling rate typically needs to be at least twice the signal frequency (satisfying the Nyquist sampling theorem). Common sampling frequencies may be between 44.1kHz and 192kHz. This invention selects signal-to-noise ratio (SNR), frequency response, and dynamic range as key indicators. Specifically, an SNR ≥ 60dB is preferable; a flat frequency response reduces frequency band distortion; and a large dynamic range can record near-field strong sounds without easily clipping distortion.

[0023] Furthermore, for subsequent sound source localization and beamforming, the array placement methods in this invention include: linear array arrangement, suitable for one-dimensional sound source localization; circular or ring array arrangement, suitable for 360° omnidirectional sound source localization; and 2D matrix arrangement, providing area array arrangement with higher spatial resolution localization capabilities. The microphone spacing is set between 0.1m and 0.5m to avoid spatial overlap, and the microphone is positioned 1-1.5m above the ground to reduce the impact of ground reflection. Simultaneously, the acquisition card / data acquisition device has multi-channel synchronous acquisition capabilities, with key parameters including: sampling rate ≥44.1kHz (commonly 48kHz or 96kHz), and quantization bit depth of 16bit or 24bit.

[0024] Data collection process: 1) The data acquisition equipment synchronizes with the UAV control system or positioning system in time; 2) Start the data acquisition device and initialize the sampling parameters (sampling rate, number of channels, gain, etc.); 3) Control the drone to fly according to the preset trajectory and operating conditions; 4) Data equipment synchronously collects the raw acoustic signals emitted by the UAV; 5) Label the collected data with tags including: drone model, propeller type, operating condition (hovering / acceleration / cruising), flight speed, altitude, distance, etc., and store them in a metadata file (JSON format), corresponding one-to-one with each audio segment.

[0025] Step S2: Preprocess the acquired acoustic signal.

[0026] In this embodiment, the present invention performs preprocessing on the original acoustic signal in the following order: Segmentation: The continuously acquired raw acoustic signal is segmented according to a fixed window time length and sliding step size, and the length of each segment is calculated. The calculation is performed using the following formula:

[0027] in, The sampling frequency (e.g., 16kHz or 48kHz). The segment length of the sound wave is given (e.g., 1 second). To improve time resolution, a sliding window is used between adjacent segments, with a sliding step size of [value missing]. (For example, 0.5s corresponds to 50% overlap), then the corresponding sample step size The calculation formula is as follows:

[0028] Then, the original sound wave signal of length N... , No. Each segmented sample is from the sample point Initially, a continuous signal of length L is extracted to form the k-th segment:

[0029] When the length of the remaining sample signal at the end is less than L, zeros are padded to the end to the signal length L to ensure the consistency of the input length of subsequent Transformer models.

[0030] DC removal: Noise in acoustic signals can originate from various sources, including microphone / front-end amplifier bias, ADC digitization offset, power supply noise, and unstable sensor zeros (common in UAV acoustic monitoring nodes). This invention removes noise by shifting the acoustic signal upwards or downwards by a constant value, rather than distributing it symmetrically around zero. Noise removal avoids abnormal convolutional layer activation (DC bias causes excessive ReLU bias), reduces excessively high low-frequency energy in the spectrum, reduces model training difficulty, improves robustness, and allows the signal to vibrate normally around zero. The specific implementation of DC removal involves applying short-time mean removal to a sliding window. The window is divided into frames per second, and for each frame, the frame value is subtracted from the frame mean. The calculation formula is as follows:

[0031] in: The first after DC removal The value of each frame, For the first The value of each frame, The number of frames within the window.

[0032] Single-pass high-pass filtering: The first-order high-pass filter used in this invention is represented as follows:

[0033] in, The filtered output sample. For the input sample, For the filter feedback coefficients, in this invention Use values ​​between 0.95 and 0.999, with 0.995 being commonly used. High-pass filtering can remove extremely low-frequency drift and equipment temperature drift.

[0034] Normalization processing: To avoid interference from amplitude differences caused by sound wave acquisition equipment or scene, the amplitude of the audio signal is normalized to ensure the consistency of the input data. This invention normalizes the amplitude of each signal segment. Calculate its standard deviation:

[0035] in: Let K be the standard deviation of the k-th signal segment, and K be the number of audio samples contained in a signal segment. Then normalization is performed:

[0036] in: The normalized value of the k-th segment of the signal. To prevent division by zero errors on extremely small positive numbers, the normalized signal has a consistent dynamic range across samples, which is beneficial for training subsequent models. The preprocessing process in this invention does not generate any spectrum, which is significantly different from traditional methods.

[0037] Step S3: Parallel multi-scale 1D-CNN temporal feature extraction.

[0038] In this embodiment, the drone's sound waves simultaneously contain: Brief details: the pulse as the propeller blades pass by, the small bursts of gear meshing; Mesoscale structures: blade-pass period of propeller rotation, harmonic clusters; Long-term structure: power changes during takeoff, acceleration, cruise, and deceleration, with slow engine speed drift.

[0039] For example, the present invention designs a 1D-CNN parallel multi-scale one-dimensional convolutional feature extraction network, which uses short-time, medium-scale and long-time time windows to convolve the same waveform. Then, multiple outputs are spliced ​​and merged in the channel dimension to form multi-scale feature fusion. The parallel multi-scale one-dimensional convolutional network includes at least two convolutional layers with different kernel sizes or dilation rates to progressively compress the time dimension while retaining key features.

[0040] Key parameters involved include: kernel size, which determines the receptive field (e.g., kernel=3, 5, 9, 15, etc., where kernel=3 captures fast details and kernel=15 captures temporal ranges spanning tens of milliseconds); dilation, which controls the effective coverage of the convolutional kernel (dilation=1, 2, 4, 8…, where dilation=1 is standard convolution and dilation=8 can cover longer temporal sequences); and layered downsampling stride / pooling, which gradually reduces temporal resolution to balance computational efficiency and feature preservation (multi-layer Conv+stride=2).

[0041] The network structure includes: The first layer: Multi-Scale ConvBank (parallel convolutional bank), used for parallel multi-scale convolutional feature extraction, contains four Conv1D branches with different parameters: Branch 1: Capture quick details; Branch 2: Slightly longer textures; Branch 3: Longer patterns; Branch 4: A rhythm spanning tens of milliseconds; The outputs of the four branches are then concatenated and merged along the channel dimension to form a comprehensive feature representation. :

[0042]

[0043]

[0044]

[0045]

[0046] in, This represents the input signal of different convolutions of the same input. , , , These represent the local temporal features extracted by different convolutional layers. , , , These represent the number of channels corresponding to different convolutional layers. The total number of channels is the sum of the number of channels in each branch, and T is the length of the time dimension.

[0047] The second layer consists of multi-layer residual dilated convolutional blocks, which are used to stack four Res Blocks with different dilation coefficients on the comprehensive feature representation F0 to expand the receptive field. This is used for voiceprint modeling of UAVs at different time scales, with dilation values ​​of 1, 2, 4, and 8, respectively, representing a gradual transition from short-term fast details to long-term modes. When dilation = 8, the effective receptive field reaches 25 time steps, which is suitable for long-term features such as continuous harmonics and rotational speed changes of UAVs. Specifically, it is represented as follows: Block1: dilation=1 (short-term) Block2: dilation=2 Block3: dilation=4 Block4: dilation=8 (long time) In each Block structure, the following is implemented: Input F0 → Conv1D (k=3, dil=d) → BN → GELU + Residual (shortcut) → Activation (GELU) → Output F1. The original information is preserved through the shortcut connection to avoid gradient vanishing.

[0048] The third layer: Layered downsampling is performed to extract multi-scale features by progressively reducing the time dimension. Conv1D (stride=2) is used to downsample the feature sequence, halving the time length each time to achieve stride convolution compression. ResBlock is inserted after each downsampling to maintain feature extraction capability. The original 16000 sampling points are then progressively compressed to 2000, forming four feature layers with different time scales, as shown below: F0→Conv1D(stride=2)→F1#T:16000→8000 F1 → ResBlock(dil=1) F1→Conv1D(stride=2)→F2#8000→4000 F2 → ResBlock(dil=2) F2→Conv1D(stride=2)→F3#4000→2000 ... Finally, features at different scales are stitched together, while preserving short-term details and long-term patterns, which greatly reduces the amount of data processed by the subsequent Transformer and improves real-time performance. Residual connections are used to avoid information loss and ensure that the compressed features are still discriminative.

[0049] Step S4: Construct the Transformer global feature modeling model.

[0050] In this embodiment, in practical applications, the system can monitor the sound waves emitted by the drone in real time and identify the drone's type or flight status using a trained Transformer model, which can then be used for applications such as security monitoring and flight tracking. Compared to traditional convolutional neural networks (CNNs) and recurrent neural networks (RNNs), the Transformer model is better at processing long-sequence data and can capture global dependencies in signals, making it perform exceptionally well in drone sound wave recognition tasks. The Transformer's self-attention and multi-head attention mechanisms are particularly suitable for processing long-sequence data and can capture complex sound wave features. Through sufficient training and optimization, the Transformer-based model can effectively complete the drone sound wave recognition task, exhibiting strong generalization ability and accuracy.

[0051] For example, the local features of the network output sound wave are extracted using the designed 1D-CNN parallel multi-scale one-dimensional convolutional feature extraction. However, drone sound has characteristics spanning time: continuous harmonic modes, increasing / decreasing rotation speed, time-varying frequency drift, and overlapping variations of multiple propellers. The global features over long time scales are obtained through a Transformer model. The Transformer model mainly consists of an encoder and a decoder. Typically, in sound wave recognition tasks, only the encoder part is used, leveraging its powerful self-attention mechanism to capture long-term dependencies in the audio signal. Therefore, the application architecture of the Transformer model in drone sound wave recognition includes the following five core components: Input Embeddings: After the audio data is converted into feature vectors, it is input into the encoder part of the Transformer. The audio signal can be encoded using a trained convolutional neural network (CNN) or other feature extraction methods (such as MFCC).

[0052] Positional Encoding: Because Transformer lacks the ability to process sequential data, positional encoding is needed to provide positional information to the model, helping the model understand the temporal dependencies in the signal.

[0053] Self-Attention: The core of the Transformer is the self-attention mechanism, which dynamically calculates the correlation between each time step when processing audio data, thereby capturing long-range dependencies in the signal. This is extremely useful for sequence modeling of audio signals, especially for handling the interaction of different frequency components.

[0054] Multi-Head Attention: In order to better capture different feature information, Transformer uses a multi-head self-attention mechanism. Each attention head is responsible for focusing on a different subset of information, and the outputs of multiple heads are concatenated to enhance the expressive power of the model.

[0055] Feedforward Networks: Each self-attention layer is usually followed by a feedforward network, which is responsible for performing nonlinear transformations on the features to further extract deeper features of the audio signal.

[0056] For example, the Transformer model structure includes the following parts: Multi-HeadSelf-Attention: By computing multiple self-attention heads in parallel, it captures the dependencies between different subspaces in the sequence, enhancing the model's ability to learn long-distance associations; Feed Forward: A feedforward neural network that performs a non-linear transformation on the output of the attention layer to further extract and integrate feature information; Layer Normalization: This stabilizes the training process and accelerates convergence by standardizing the feature distribution of each sample to reduce internal covariate bias. Residual connections pass the input directly to the next layer, alleviating the vanishing gradient problem and allowing the network to learn residual mappings instead of complete mappings.

[0057] Output:

[0058] in, The output of the model is the sequence of global acoustic features of the UAV. Let L be the real number field, where L represents the length of the feature sequence, which is equal to the number of time segments of the sound wave; and D represents the feature dimension corresponding to each time position.

[0059] For example, multi-head self-attention is a core component of the Transformer model. By computing multiple self-attention heads in parallel, it enhances the model's ability to capture complex dependencies in a sequence. The self-attention mechanism generates a weighted representation for each element by calculating attention weights with other elements, without relying on external information. First, multi-head expansion is performed, splitting the input sequence into multiple subspaces (e.g., 8 heads). Each head independently calculates its attention weights, and the outputs of all heads are concatenated and fused through a linear layer to obtain the final output.

[0060] When training a Transformer model, you will also need to use: Loss Function: For sound wave recognition tasks, the commonly used loss function is cross-entropy loss, which is used in multi-class classification tasks to measure the difference between the class predicted by the model and the actual class.

[0061] Optimizer: Common optimizers include Adam or AdamW, which can effectively adjust the learning rate during training to help the model converge.

[0062] The training process involves training the model using labeled sound wave data. During training, data augmentation (such as adding noise or changing the audio speed) can be used to increase the model's robustness.

[0063] In this embodiment, after training, the evaluation and optimization of the Transformer model includes: Define evaluation metrics: Use the test dataset to evaluate the model. Evaluation metrics include accuracy, precision, recall, and F1 score.

[0064] Evaluation and optimization: When the model's evaluation results are unsatisfactory, optimization can be performed using the following methods: Increase the amount of data: More training data helps the model learn more features and improves performance.

[0065] Model hyperparameter tuning: Adjusting the model's hyperparameters (such as the number of layers, head number, number of hidden units, etc.) to improve performance.

[0066] Transfer learning: You can try using a Transformer model pre-trained on a large-scale audio dataset (such as a sound classification dataset) and then fine-tuning it, which can speed up training and improve the model's performance.

[0067] After training and optimizing the model, it is deployed to the drone monitoring system. The deployment method can be edge computing or cloud computing, depending on the application scenario.

[0068] Step S5: Perform multi-task output.

[0069] In this embodiment, multi-task output is more effective for drone acoustic recognition. Multi-task in this invention includes: Is it a drone? (Detection required); What type of drone is it? (Model / Category); What flight state is it in? (Takeoff / Hovering / Cruising / Landing) In what direction is the sound source approximately located? (Azimuth, DOA) How far away is the sound source? (Rough estimate of distance / range); What is the sound quality like? (SNR, confidence level, etc.) These tasks are interconnected, and using them simultaneously is more accurate and stable than using a single main task (such as "drone classification"), which is the value of multi-task output.

[0070] In this embodiment, the reasoning and decision-making process applied to the actual system is as follows: Prioritize inference of drone presence / absence: Use the Sigmoid function to output the probability of drone presence in each time window, and perform voting / smoothing on multiple windows over a period of time. If the probability of having a drone exceeds the set threshold, proceed to the next step; otherwise, determine that there are no drones.

[0071] Drone model classification: Softmax is used to output the model category for each time window. The model with the highest probability among multiple model classifications is taken as the predicted model. If the probability is lower than the set threshold, it is classified as an unknown model.

[0072] Drone state determination: Softmax is used to output the state of each time window. The state with the highest probability is selected in each time window, and voting / smoothing is performed on multiple windows over a period of time to obtain the state.

[0073] Drone location determination: can be directly used Returning, or returning Two values ​​are used, and then the angle is restored using atan2 to avoid the discontinuity between 0° and 360°. Alternatively, a multi-task loss function design can be employed.

[0074] Drone distance determination: By regressing the head output distance d, the distance can be roughly determined as far / medium / near.

[0075] Post-processing in the time dimension: Voting and exponential smoothing are performed on the judgment results of multiple consecutive time windows to make the prediction results more stable.

[0076] Example 2 This embodiment describes an end-to-end drone identification system based on raw sound waves, which includes the following functional modules: Acoustic signal acquisition module: used to acquire acoustic signals from the drone via a microphone array; Acoustic signal preprocessing module: used to preprocess the collected UAV acoustic signals, such as segmentation, filtering, noise reduction, and normalization; Parallel multi-scale 1D-CNN temporal feature extraction module: used to convolve the same waveform using short-time, medium-scale, and long-time time windows, and then concatenate and merge multiple outputs in the channel dimension; Transformer global feature modeling module: Used to input data from the parallel multi-scale 1D-CNN temporal feature extraction module and use the trained Transformer model to identify the type or flight status of the drone. Multi-task output module: Used to output inference results including: whether there is a drone, drone model, flight status, direction, and distance.

[0077] This invention directly uses the original time-domain sound wave as input, extracts local time-domain features through a multi-scale one-dimensional convolutional neural network (1D-CNN), and then captures the long-term dependent patterns of the drone's acoustic signature using a Transformer network. This enables drone presence detection, model classification, flight attitude recognition, direction estimation, and distance assessment. This invention eliminates the need for traditional time-frequency feature extraction (such as spectrograms and MFCC), significantly improving the robustness of recognition in noisy environments. It is widely applicable to scenarios such as security monitoring, anti-drone detection, and early warning.

[0078] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0079] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An end-to-end UAV identification method based on raw sound waves, characterized in that: The following steps are included: Step S1: Set up multiple microphone arrays in the target area to collect the raw waveform signals generated by different types of drones under different states and environments; Step S2: Perform preprocessing on the acquired raw waveform signal, including normalization and filtering; Step S3: Construct a 1D-CNN parallel multi-scale one-dimensional convolutional feature extraction network, and use short-time, medium-scale and long-time time windows to convolve the same preprocessed waveform to generate local temporal features; Step S4: Construct a Transformer model by inputting local temporal features into the Transformer network to generate global voiceprint features; Step S5: Output information on the presence, model, distance, and rotation speed of the UAV based on the global acoustic signature characteristics.

2. The end-to-end UAV identification method based on raw sound waves according to claim 1, characterized in that: In step S1: The types of drones include: quadcopters, multi-rotors, and fixed-wing drones; The states of a drone include: hovering, climbing, descending, constant speed flight, and acceleration / deceleration; Microphone array setup methods include: linear arrangement, circular or ring arrangement, and 2D matrix arrangement.

3. The end-to-end UAV identification method based on raw sound waves according to claim 2, characterized in that: The data acquisition process in step S1 includes: Step S11: The data acquisition equipment synchronizes with the UAV control system or positioning system in time; Step S12: Start the data acquisition device and initialize the sampling parameters, including sampling rate, number of channels, and gain; Step S13: Control the drone to fly in the designated area according to the preset trajectory and operating conditions; Step S14: The data acquisition device synchronously acquires the raw acoustic signal emitted by the UAV; Step S15: Label the collected data with tags including: drone model, propeller type, operating condition, flight speed, altitude, and distance, and store them in a metadata file, corresponding one-to-one with each audio segment.

4. The end-to-end UAV identification method based on raw sound waves according to claim 3, characterized in that: The preprocessing in step S2 includes: The raw acoustic wave signals obtained by continuous acquisition are segmented according to a fixed window time length and a sliding step size, and the length of each segment is calculated. The sound wave signal is shifted upwards or downwards by a constant value to remove noise. By applying short-time mean removal to the sliding window, the window is divided into frames per second, and for each frame, the frame value is subtracted from the frame mean. Then, a first-order high-pass filter is used to filter the input samples to remove extremely low-frequency drift and equipment temperature drift. Calculate the standard deviation for each signal segment, and then normalize the signal amplitude.

5. The end-to-end UAV identification method based on raw sound waves according to claim 4, characterized in that: The 1D-CNN parallel multi-scale one-dimensional convolutional feature extraction network constructed in step S3 includes the following architecture: The first layer is a parallel convolutional bank, used for parallel multi-scale convolutional feature extraction. It contains four Conv1D branches with different parameters. The outputs of the four branches are concatenated and merged in the channel dimension to form a comprehensive feature representation. The second layer consists of multi-layer residual dilated convolutional blocks, which are used to stack four Bloc structures with different dilation coefficients on the comprehensive feature representation to expand the receptive field and perform voiceprint modeling for drones at different time scales. The third layer is hierarchical downsampling. Multi-scale features are extracted by gradually reducing the time dimension. Conv1D is used to downsample the feature sequence, halving the time length each time to achieve stride convolution compression. ResBlock is inserted after each downsampling to form four feature layers with different time scales.

6. The end-to-end UAV identification method based on raw sound waves according to claim 5, characterized in that: The architecture of the Transformer model constructed in step S4 includes: Input embedding: After the audio data is converted into a feature vector, it is input into the encoder part of the Transformer, and the audio signal is encoded using a trained convolutional neural network or other feature extraction methods; Position encoding: Position encoding provides positional information to the Transformer model, enabling the Transformer model to understand the temporal dependencies in the signal; Self-attention mechanism: When processing audio data, the correlation between each time step is dynamically calculated based on the self-attention mechanism to capture long-distance dependencies in the signal; Multi-head self-attention mechanism: The Transformer model uses a multi-head self-attention mechanism, where each attention head is responsible for focusing on a different subset of information, and then the outputs of multiple heads are concatenated. Feedforward Neural Network: Each self-attention layer is followed by a feedforward neural network, which is responsible for performing nonlinear transformations on the features to further extract deeper features of the audio signal.

7. The end-to-end UAV identification method based on raw sound waves according to claim 6, characterized in that: Step S4 also includes: By computing multiple self-attention heads in parallel, the dependencies between different subspaces in the sequence are captured, enhancing the model's ability to learn long-distance associations. Layer normalization is performed to reduce internal covariate bias by standardizing the feature distribution of each sample; By passing the input directly to the next layer, the network is allowed to learn residual mappings instead of complete mappings. Finally, the global acoustic feature sequence of the UAV is output based on the Transformer model.

8. The end-to-end UAV identification method based on raw sound waves according to claim 7, characterized in that: Step S4 also includes training the Transformer model using the specified loss function and optimizer with labeled acoustic data. After training, the Transformer model is evaluated using a test dataset, with evaluation metrics including accuracy, precision, recall, and F1-score. Based on the evaluation results, the Transformer model is optimized by increasing the amount of data, adjusting the model's hyperparameters, and performing transfer learning. The optimized Transformer model is then deployed to the drone monitoring system.

9. The end-to-end UAV identification method based on raw sound waves according to claim 8, characterized in that: Step S5 includes: The probability of drone presence in each time window is output using the Sigmoid function, and voting / smoothing is performed on multiple windows within a set period. When the probability of drone presence exceeds the set threshold, the next step is initiated; otherwise, no drone is detected. The Softmax function is used to output the model category for each time window. The model with the highest probability among multiple model determinations is taken as the predicted model. If the probability is lower than the set threshold, it is determined to be an unknown model. The Softmax function is used to output the state of each time window. The state with the highest probability is selected in each time window. Voting / smoothing is performed on multiple windows over a period of time to obtain the drone state. The location of the UAV is determined by designing a direct regression angle θ or a multi-task loss function; The distance to the drone was determined using a regression model; Voting and exponential smoothing are performed on the judgment results of multiple consecutive time windows to conduct post-processing in the time dimension.

10. An end-to-end UAV identification system based on raw sound waves, executing the end-to-end UAV identification method based on raw sound waves as described in any one of claims 1-9, characterized in that: include: Acoustic signal acquisition module: used to acquire acoustic signals from the drone via a microphone array; Acoustic signal preprocessing module: used to preprocess the acquired UAV acoustic signals; Parallel multi-scale 1D-CNN temporal feature extraction module: used to convolve the same waveform using short-time, medium-time, and long-time time windows, and multiple outputs are concatenated and merged in the channel dimension; Transformer global feature modeling module: It is used to input data from the parallel multi-scale 1D-CNN temporal feature extraction module and use the trained Transformer model to identify the type or flight status of the drone. Multi-task output module: Used to output inference results including whether there is a drone, drone model, flight status, direction, and distance.

Images