Unmanned aerial vehicle audio classification method based on convolutional neural network

Abstract

The invention discloses an unmanned aerial vehicle audio classification method based on a convolutional neural network. The method comprises the following steps: collecting flight audio, a working condition state and vibration data, and carrying out time alignment; framing time-frequency transformation is carried out to generate a spectrogram, and harmonic alignment normalization is carried out according to the rotating speed of a motor; suppressing low frequency according to wind noise index gating, and performing weight reduction on a frequency band according to a vibration consistency weight to obtain an enhanced spectrogram; constructing a three-channel evidence spectrogram tensor, and generating a working condition vector; using an improved ConvNeXt V2 model to output a category probability, a confidence coefficient and a key evidence set; triggering grading re-judgment according to confidence and wind noise, vibration and self-noise deviation, and adjusting window length and gating intensity to generate a result packet; and caching the low-confidence sample, updating the normalized statistical and gating threshold on line, and continuously outputting a result packet. According to the method, the classification accuracy, robustness and online adaptive capacity under complex working conditions are improved, and the interpretability of key evidences is enhanced.

Description

Technical Field

[0001] This invention relates to the field of deep learning and intelligent signal processing technology, and in particular to a method for classifying drone audio based on convolutional neural networks. Background Technology

[0002] With the large-scale application of drones in scenarios such as inspection, security, and emergency rescue, the demand for environmental sound source identification and abnormal sound alarms based on airborne audio is constantly increasing. Existing drone audio classification technologies mostly employ single-microphone audio acquisition followed by direct spectrogram feature extraction, and then use traditional convolutional neural networks or general audio classification networks to complete category discrimination. Some solutions only suppress propeller noise and wind noise interference through fixed bandpass filtering or simple noise reduction processing. However, the following problems are commonly found in actual flight applications:

[0003] The acquired audio signals are significantly affected by changes in motor speed. The harmonic peaks drift with the speed, causing the spectrum distribution to shift across operating conditions. The model is prone to misclassifying differences in operating conditions as differences in categories. The low-frequency strong noise introduced by wind noise and machine vibration, as well as structural conducted noise, have strong non-stationarity. Fixed threshold noise reduction is difficult to balance suppression effect with effective sound source fidelity, resulting in increased false alarm and false negative rates. The sampling frequencies of audio and multi-source data such as operating conditions and vibration are inconsistent and the temporal alignment is unstable. Existing methods lack frame-level synchronization and consistency weighting mechanisms, leading to discontinuous feature construction. Most methods only output category results and lack key evidence localization and low-confidence re-judgment mechanisms, making it difficult to meet the reliability and replayability requirements of real-time deployment on the edge.

[0004] Therefore, how to provide a drone audio classification method based on convolutional neural networks 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 UAV audio classification method based on convolutional neural networks. This invention fully utilizes spectrogram harmonic alignment normalization, wind noise-gated low-frequency suppression and vibration consistency band weighting enhancement, multi-channel evidence spectrogram tensor construction, and the conditional GRN and harmonic sensitive band gating of the improved ConvNeXt V2, as well as multi-head evidence output. Combined with confidence-triggered hierarchical reasoning and online lightweight adaptive update mechanism, it achieves stable classification of UAV flight audio under complex operating conditions and interpretable output of key evidence. It has the advantages of high classification accuracy, strong robustness against wind noise and vibration, good adaptability to changes in operating conditions, and strong online continuous update capability.

[0006] According to an embodiment of the present invention, a method for classifying drone audio based on a convolutional neural network includes the following steps:

[0007] Step 1: Collect audio data, operational status data, and vibration data during the drone's flight process, align the execution time, and obtain the operational status data and vibration data corresponding to each audio frame;

[0008] Step 2: Perform frame segmentation and time-frequency transformation on the audio data to generate a spectrogram, and perform harmonic alignment and normalization processing on the spectrogram based on the motor speed in the operating condition data;

[0009] Step 3: Based on the wind noise index, perform gated low-frequency suppression processing on the spectrogram, and generate vibration consistency weights based on vibration data to perform weight reduction processing on the corresponding frequency bands of the spectrogram to obtain the enhanced spectrogram;

[0010] Step 4: Construct the main spectrum channel from the enhanced spectrogram, and construct the phase stability evidence channel and the self-noise deviation evidence channel to form a three-channel evidence spectrum tensor, and generate the operating condition vector at the same time;

[0011] Step 5: Input the three-channel evidence spectrum tensor and the operating condition vector into the improved ConvNeXt V2 model. The improved ConvNeXt V2 model includes a harmonic sensitive frequency band gate layer, a ConvNeXt V2 convolution backbone, an operating condition-conditioned GRN module, a classification head, and an evidence head. The multi-head outputs the class probability, confidence level, and key evidence set.

[0012] Step 6: Based on the confidence level, wind noise index, vibration index, and self-noise deviation index, trigger the hierarchical reasoning and re-judgment process, and generate the classification result package by adjusting the window length and gating strength;

[0013] Step 7: Cache the low-confidence samples and their corresponding evidence spectrum tensors, and perform online lightweight adaptive updates on the normalized statistics and gating thresholds to output a continuously updated classification result package.

[0014] Optionally, step one specifically includes:

[0015] Collect audio data, operating status data, and vibration data during the drone's flight, and write a first timestamp for each audio frame;

[0016] The operating condition data includes motor speed, throttle opening and attitude angle data, and a second timestamp is written for each piece of operating condition data.

[0017] The vibration data consists of triaxial acceleration data, and a third timestamp is written for each vibration data point.

[0018] Using the audio frame time index corresponding to the first timestamp as the alignment reference, the second and third timestamps are mapped to the audio frame time index;

[0019] When the operating condition data or vibration data is missing at a certain audio frame time index, it is filled in by keeping the previous valid value, and a missing identifier is written at the filling position to obtain the operating condition features and vibration features corresponding to each audio frame.

[0020] Optionally, step two specifically includes:

[0021] The audio data is divided into frames. After windowing, a fast Fourier transform is performed on each frame. The frames are then stitched together in order to obtain the time-frequency amplitude spectrum. Logarithmic compression is then performed on the amplitude spectrum to generate the initial spectrogram.

[0022] The speed level corresponding to the current audio frame is determined based on the motor speed, and a harmonic position table is generated based on the speed level.

[0023] According to the harmonic location table, the harmonic peak search interval is determined on the frequency axis of the initial spectrum. The local maximum frequency point is extracted in each harmonic peak search interval as the harmonic peak frequency point. The harmonic peak frequency point is aligned and mapped with the target frequency point in the harmonic location table to form the frequency band alignment offset.

[0024] Based on the frequency band alignment offset, the frequency axis of the initial spectrogram is resampled and aligned so that the harmonic peak frequencies of each audio frame are mapped to the same frequency band index position in the resampled spectrogram. Then, amplitude normalization is performed on the resampled spectrogram to obtain the harmonic aligned normalized spectrogram.

[0025] Optionally, step three specifically includes:

[0026] Based on the harmonic-aligned normalized acoustic spectrum, the spectral energy of each audio frame in the low-frequency range is statistically analyzed according to the preset low-frequency range, and the spectral energy of the audio frame in the full frequency range is also statistically analyzed. The proportion of low-frequency energy is calculated as the low-frequency energy index. At the same time, the change in the low-frequency energy index between adjacent audio frames is calculated as the low-frequency change index. The low-frequency energy index and the low-frequency change index are combined to generate the wind noise index.

[0027] Based on the comparison between the wind noise index and the preset wind noise threshold, when the wind noise index exceeds the preset wind noise threshold, a low-frequency suppression weight is generated, and the spectral amplitude value of the corresponding low-frequency range in the harmonic aligned normalized sound spectrum is scaled according to the low-frequency suppression weight to obtain the wind noise suppression sound spectrum. When the wind noise index does not exceed the preset wind noise threshold, the low-frequency suppression weight is set to no scaling weight and the wind noise suppression sound spectrum is output.

[0028] Within the time window corresponding to the current audio frame, the mean and peak values ​​of the absolute values ​​of acceleration are calculated for the vibration data along the three axes to form a vibration intensity sequence. Based on the vibration intensity sequence and the preset vibration threshold, a vibration consistency weight is generated.

[0029] The vibration consistency weights are mapped to the frequency axis of the wind noise suppression spectrogram, and the spectral amplitude of the mapped frequency band is reduced and scaled to obtain the enhanced spectrogram.

[0030] Optionally, step four specifically includes:

[0031] The enhanced spectrogram is cropped or scaled according to the preset time window length and preset number of frequency bands to be uniformly of the preset size, and the enhanced spectrogram of the preset size is used as the main spectrogram channel.

[0032] When the drone is configured as an airborne microphone array, time-frequency transformation is performed on the audio frames of each microphone within the same time window to obtain multiple spectrograms. The cross-spectral consistency index of any two spectrograms at the corresponding time-frequency positions is calculated. The cross-spectral consistency index is spliced ​​in the frequency dimension and the time dimension to form a phase stability evidence channel. When the drone is configured as a single microphone, the phase continuity index and spectral stability index are calculated between adjacent audio frames of the enhanced spectrogram. The phase continuity index and spectral stability index are mapped according to the preset fusion rules to form a phase stability evidence channel.

[0033] Based on the harmonic location table, the peak search intervals corresponding to each order harmonic are determined in the enhanced acoustic spectrum. The peak amplitude and peak width features within the peak search interval are extracted, and the peak amplitude and peak width features are compared with the preset self-noise template features to obtain the self-noise deviation index map as a self-noise deviation evidence channel.

[0034] The main spectrum channel, phase stability evidence channel, and self-noise deviation evidence channel are spliced ​​together in the channel dimension to form a three-channel evidence spectrum tensor. The operating condition data, wind noise index, vibration intensity sequence, and self-noise deviation index are then spliced ​​together in sequence to form an operating condition vector.

[0035] Optionally, the improved ConvNeXt V2 model includes a harmonic-sensitive band gating layer, a ConvNeXt V2 convolutional backbone, a conditional GRN module, a classification header, and an evidence header:

[0036] The harmonic sensitive frequency band gating layer receives the three-channel evidence spectrum tensor and the operating condition vector, reads the rotation speed field from the operating condition vector and generates a set of harmonic frequency band positions and corresponding bandwidth parameters, constructs a frequency band weight vector on the frequency axis of the spectrum based on the set of harmonic frequency band positions, and performs interval truncation and neighborhood smoothing on the frequency band weight vector to obtain the gating weight vector, copies the gating weight vector along the time dimension to form a gating weight map, performs element-wise multiplication and weighting on each channel of the three-channel evidence spectrum tensor according to the frequency band, and outputs the gating three-channel evidence spectrum tensor.

[0037] The ConvNeXt V2 convolutional backbone includes a first convolutional embedding layer, first to fourth level feature extraction stages, and a downsampling layer. The first convolutional embedding layer maps the gated three-channel evidence spectrum tensor into a first feature map through convolution. The downsampling layer is set between adjacent level feature extraction stages and sequentially connects adjacent level feature extraction stages. The feature map flows in hierarchical order to the fourth level feature extraction stage for output.

[0038] Each ConvNeXt V2 convolutional block sequentially includes a depthwise separable convolutional layer, a layer normalization layer, a first point convolutional layer, a GELU activation layer, a conditional GRN module, a second point convolutional layer, and a residual connection layer. The depthwise separable convolutional layer is a channel-grouped convolutional layer that performs local time-frequency texture extraction on the feature map. The first point convolutional layer expands the channel dimension. The second point convolutional layer projects the channel dimension back to the number of channels input to the convolutional block. The residual connection layer adds the feature map input to the convolutional block and the feature map output by the second point convolutional layer element-wise to obtain the output feature map of the convolutional block.

[0039] The conditional GRN module is embedded in each ConvNeXt V2 convolutional block of the first-level feature extraction stage and the second-level feature extraction stage. It receives the channel features and conditional vectors output by the GELU activation layer, inputs the conditional vectors into the multilayer perceptron, generates a scaling vector and a bias vector consistent with the number of channels, performs global aggregation on the channel features in the spatial dimension to obtain the global response intensity of each channel, and performs normalization on the global response intensity to obtain the GRN intermediate features. The GRN intermediate features are multiplied element-wise by the scaling vector by channel and added element-wise by the bias vector by channel to obtain the conditional GRN output and pass it to the second point convolutional layer.

[0040] The evidence head and classification head are set in parallel after the output of the fourth-level feature extraction stage. The classification head performs global average pooling on the feature map output by the fourth-level feature extraction stage to obtain a global feature vector, and outputs the class probability through a fully connected layer. At the same time, temperature calibration is performed on the class probability to obtain the confidence level. The evidence head performs convolution on the feature map output by the fourth-level feature extraction stage to obtain an evidence heatmap, performs normalization on the evidence heatmap, and selects the Top-K positions according to the response value from large to small. The Top-K positions are mapped back to the time-frequency block index of the three-channel evidence spectrum tensor before gating, and the corresponding response value is used as the weight to form a key evidence set.

[0041] Optionally, step six specifically includes:

[0042] Based on category probability, confidence level, key evidence set, wind noise index, vibration consistency weight and self-noise deviation index, a hierarchical inference trigger field is generated. The hierarchical inference trigger field includes low confidence trigger identifier, wind noise trigger identifier, vibration trigger identifier and deviation trigger identifier.

[0043] When any trigger flag is set, the review process is initiated. If the review process is not initiated, the audio data is windowed using the first inference window length and a three-channel evidence spectrum tensor is generated. The improved ConvNeXt V2 model is called to output the first category probability, the first confidence level, and the first key evidence set, which are then output as the classification result package.

[0044] When entering the review process, the inference window length is switched from the first inference window length to the second inference window length, and the low-frequency suppression weight is switched to the enhanced suppression weight in the gated low-frequency suppression processing. At the same time, in the harmonic sensitive frequency band gated layer of the improved ConvNeXt V2 model, the suppression weight of the harmonic frequency band is switched to the enhanced suppression weight. Based on the second inference window length, the three-channel evidence spectrum tensor is regenerated, and the improved ConvNeXt V2 model is called to output the second category probability, the second confidence level, and the second key evidence set. The final classification result package is generated with the category corresponding to the second category probability as the final category, the second confidence level as the final confidence level, and the second key evidence set as the final key evidence set.

[0045] Write the inference path identifier and the trigger identifier set into the final classification result package.

[0046] Optionally, step seven specifically includes:

[0047] Read the confidence level, inference path identifier and trigger identifier set from the final classification result package. When the inference path identifier indicates that the review process is to be entered or the confidence level is less than the preset cache threshold, write the three-channel evidence spectrum tensor, working condition vector, category probability, confidence level and key evidence set of the corresponding time window into the cache queue.

[0048] Set a capacity limit for the cache queue. When the cache queue reaches the capacity limit, sort the samples by confidence level from smallest to largest and retain samples with lower confidence level first. When there are samples with the same confidence level, sort them by the number of set flags in the trigger flag set from largest to smallest and retain samples with more set flags first. When there are still samples with the same confidence level, sort them by standard timestamp from newest to oldest and retain newer samples first.

[0049] Online lightweight adaptive updates are triggered at a preset update cycle, and the online lightweight adaptive updates include normalized statistical updates and gating threshold updates;

[0050] The updated normalized statistics and gating thresholds are applied to the spectrogram processing of audio data, the construction of three-channel evidence spectrogram tensors, and the inference process of the improved ConvNeXt V2 model, and the updated classification result package is continuously output.

[0051] The beneficial effects of this invention are:

[0052] This invention establishes a frame-level time alignment between audio data, operating condition data, and vibration data collected during UAV flight, creating a unique operating condition and vibration feature input corresponding to each audio frame. During the spectrogram construction stage, harmonic alignment normalization based on motor speed is introduced to suppress cross-operating condition distribution shifts caused by harmonic peak drift with speed. Furthermore, a gated low-frequency suppression driven by wind noise indices and a frequency-band weighting of vibration consistency are employed to achieve targeted reduction of strong non-stationary wind noise and structurally conducted noise without excessively damaging effective sound source characteristics. Based on this, a three-channel evidence spectrum tensor consisting of the main spectrum, phase stability evidence, and self-noise bias evidence is constructed. Combined with the improved ConvNeXt V2 model inputting the operating condition vector, and through a harmonic sensitive frequency band gating layer, an embedded operating condition-conditioned GRN module, and parallel classification and evidence heads, robust output of class probability and confidence level, as well as traceable location of key evidence sets, are achieved. Furthermore, by combining a hierarchical reasoning and re-judgment mechanism triggered by confidence level, wind noise, vibration, and deviation indicators, and an online lightweight adaptive update mechanism, real-time performance and reliability are balanced under the condition of limited computing power at the edge. This significantly reduces false alarms and missed alarms caused by changes in wind conditions, speed changes, and machine state drift, and enhances the reproducibility and long-term sustainable availability of the results, which has important engineering application value and promotion significance. Attached Figure Description

[0053] 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:

[0054] Figure 1 This is a flowchart of a drone audio classification method based on a convolutional neural network proposed in this invention;

[0055] Figure 2 This is a schematic diagram of a drone audio classification method based on a convolutional neural network proposed in this invention;

[0056] Figure 3 This is a framework diagram of the improved ConvNeXt V2 model in the UAV audio classification method based on convolutional neural networks proposed in this invention. Detailed Implementation

[0057] 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.

[0058] refer to Figure 1-3 A method for drone audio classification based on convolutional neural networks includes the following steps:

[0059] Step 1: Collect audio data, operational status data, and vibration data during the drone's flight process, align the execution time, and obtain the operational status data and vibration data corresponding to each audio frame;

[0060] Step 2: Perform frame segmentation and time-frequency transformation on the audio data to generate a spectrogram, and perform harmonic alignment and normalization processing on the spectrogram based on the motor speed in the operating condition data;

[0061] Step 3: Based on the wind noise index, perform gated low-frequency suppression processing on the spectrogram, and generate vibration consistency weights based on vibration data to perform weight reduction processing on the corresponding frequency bands of the spectrogram to obtain the enhanced spectrogram;

[0062] Step 4: Construct the main spectrum channel from the enhanced spectrogram, and construct the phase stability evidence channel and the self-noise deviation evidence channel to form a three-channel evidence spectrum tensor, and generate the operating condition vector at the same time;

[0063] Step 5: Input the three-channel evidence spectrum tensor and the operating condition vector into the improved ConvNeXt V2 model. The improved ConvNeXt V2 model includes a harmonic sensitive frequency band gate layer, a ConvNeXt V2 convolution backbone, an operating condition-conditioned GRN module, a classification head, and an evidence head. The multi-head outputs the class probability, confidence level, and key evidence set.

[0064] Step 6: Based on the confidence level, wind noise index, vibration index, and self-noise deviation index, trigger the hierarchical reasoning and re-judgment process, and generate the classification result package by adjusting the window length and gating strength;

[0065] Step 7: Cache the low-confidence samples and their corresponding evidence spectrum tensors, and perform online lightweight adaptive updates on the normalized statistics and gating thresholds to output a continuously updated classification result package.

[0066] In this embodiment, step one specifically includes:

[0067] Collect audio data, operating status data, and vibration data during the drone's flight, and write a first timestamp for each audio frame;

[0068] The operating condition data includes motor speed, throttle opening and attitude angle data, and a second timestamp is written for each piece of operating condition data.

[0069] The vibration data consists of triaxial acceleration data, and a third timestamp is written for each vibration data point.

[0070] Using the audio frame time index corresponding to the first timestamp as the alignment benchmark, the second and third timestamps are mapped to the audio frame time index. The working condition data uses the nearest neighbor matching method to select the data with the smallest difference from the current audio frame timestamp as the working condition data corresponding to the audio frame. The vibration data calculates the mean and peak values ​​of the absolute values ​​of acceleration on the three axes within the time window covered by the current audio frame and splices them to form vibration features.

[0071] When the operating condition data or vibration data is missing at a certain audio frame time index, it is filled in by keeping the previous valid value, and a missing identifier is written at the filling position to obtain the operating condition features and vibration features corresponding to each audio frame.

[0072] In this embodiment, step two specifically includes:

[0073] The audio data is divided into frames according to the set frame length and frame shift. After windowing processing, a fast Fourier transform is performed on each frame of audio. The frames are then stitched together in order to obtain the time-frequency amplitude spectrum. The amplitude spectrum is then logarithmically compressed to generate the initial spectrogram.

[0074] The speed gear corresponding to the current audio frame is determined based on the motor speed, and a harmonic position table is generated based on the speed gear. The harmonic position table includes the fundamental frequency position and several harmonic positions. The fundamental frequency position is obtained by converting the speed and combined with the number of blades, and the several harmonic positions are obtained by recursively extrapolating the fundamental frequency position in integer multiples.

[0075] According to the harmonic location table, the harmonic peak search interval is determined on the frequency axis of the initial spectrum. The local maximum frequency point is extracted in each harmonic peak search interval as the harmonic peak frequency point. The harmonic peak frequency point is aligned and mapped with the target frequency point in the harmonic location table to form the frequency band alignment offset.

[0076] Based on the frequency band alignment offset, the frequency axis of the initial spectrogram is resampled and aligned so that the harmonic peak frequencies of each audio frame are mapped to the same frequency band index position in the resampled spectrogram. Then, amplitude normalization is performed on the resampled spectrogram to obtain the harmonic aligned normalized spectrogram.

[0077] This implementation method constructs a spectrogram using frame-by-frame windowed FFT, generates a harmonic location table based on the motor speed, takes the local maximum in the search interval, and resamples and aligns it for normalization. This differs from the approach of directly inputting the original spectrogram into a CNN, effectively eliminating harmonic drift caused by changes in motor speed.

[0078] In this embodiment, step three specifically includes:

[0079] Based on the harmonic-aligned normalized acoustic spectrum, the spectral energy of each audio frame in the low-frequency range is statistically analyzed according to the preset low-frequency range, and the spectral energy of the audio frame in the full frequency range is also statistically analyzed. The proportion of low-frequency energy is calculated as the low-frequency energy index. At the same time, the change in the low-frequency energy index between adjacent audio frames is calculated as the low-frequency change index. The low-frequency energy index and the low-frequency change index are combined to generate the wind noise index.

[0080] Based on the comparison between the wind noise index and the preset wind noise threshold, when the wind noise index exceeds the preset wind noise threshold, a low-frequency suppression weight is generated, and the spectral amplitude value of the corresponding low-frequency range in the harmonic aligned normalized sound spectrum is scaled according to the low-frequency suppression weight to obtain the wind noise suppression sound spectrum. When the wind noise index does not exceed the preset wind noise threshold, the low-frequency suppression weight is set to no scaling weight and the wind noise suppression sound spectrum is output.

[0081] Within the time window corresponding to the current audio frame, the mean and peak values ​​of the absolute values ​​of acceleration are calculated for the vibration data along the three axes to form a vibration intensity sequence. Based on the vibration intensity sequence and a preset vibration threshold, a vibration consistency weight is generated. The vibration consistency weight represents the frequency band in the current audio frame's spectrogram that is synchronized with the change in vibration intensity.

[0082] The vibration consistency weights are mapped to the frequency axis of the wind noise suppression spectrogram. The spectral amplitude of the mapped frequency bands is then subjected to weight reduction scaling to obtain an enhanced spectrogram. The weight reduction scaling process first converts the vibration intensity obtained from the vibration data within the current audio frame time window into weight reduction coefficients for each frequency band according to a preset mapping rule. The weight reduction coefficients are limited to a preset minimum and maximum value. Then, the spectral amplitude of each frequency band in each audio frame of the spectrogram is scaled band-by-band according to the corresponding weight reduction coefficient. Frequency bands that are synchronized with the change in vibration intensity are suppressed using a weight reduction coefficient of less than one, while asynchronous frequency bands are kept unchanged with a coefficient equal to one. The weight reduction coefficients are smoothed along the frequency axis in a neighborhood, and the weight reduction coefficients of adjacent audio frames are smoothed by moving average to obtain continuous weight reduction scaling results.

[0083] This implementation generates a wind noise index based on the proportion of low-frequency energy and its corresponding changes, and then gates and suppresses low frequencies. At the same time, it reduces and smooths the synchronous frequency band band by band by band using vibration consistency weights. Unlike fixed filtering or simple spectral reduction schemes, this implementation can still maintain effective information under strong wind and resonance conditions, and significantly reduces false alarms and missed alarms.

[0084] In this embodiment, step four specifically includes:

[0085] The enhanced spectrogram is cropped or scaled according to the preset time window length and preset number of frequency bands to be uniformly of the preset size, and the enhanced spectrogram of the preset size is used as the main spectrogram channel.

[0086] When the UAV is configured with an airborne microphone array, time-frequency transformation is performed on the audio frames of each microphone within the same time window to obtain multiple spectrograms. The cross-spectral consistency index of any two spectrograms at the corresponding time-frequency positions is calculated. The cross-spectral consistency index is spliced ​​in the frequency dimension and the time dimension to form a phase stability evidence channel. When the UAV is configured with a single microphone, the phase continuity index and spectral stability index are calculated between adjacent audio frames for the enhanced spectrogram. The phase continuity index and spectral stability index are mapped according to a preset fusion rule to form a phase stability evidence channel. The preset fusion rule is to truncate the phase continuity index and the spectral stability index and linearly normalize them respectively to obtain a dimensionless index map. The two types of dimensionless index maps after normalization are weighted and summed according to a preset weight coefficient. Neighborhood smoothing is performed on the weighted result to eliminate isolated noise.

[0087] Based on the harmonic location table, the peak search intervals corresponding to each order harmonic are determined in the enhanced acoustic spectrum. The peak amplitude and peak width features within the peak search interval are extracted, and the peak amplitude and peak width features are compared with the preset self-noise template features to obtain the self-noise deviation index map as a self-noise deviation evidence channel. The preset self-noise template features are the historical statistical features of harmonic peak amplitude and peak width collected by the UAV at a preset speed level.

[0088] The main spectrum channel, phase stability evidence channel, and self-noise deviation evidence channel are spliced ​​together in the channel dimension to form a three-channel evidence spectrum tensor. The operating condition data, wind noise index, vibration intensity sequence, and self-noise deviation index are then spliced ​​together in sequence to form an operating condition vector.

[0089] This implementation method constructs a main spectrogram channel, a phase stability evidence channel, and a self-noise deviation evidence channel, and splices them together to generate a working condition vector. This differs from the single-channel spectrogram input scheme, enabling the model to obtain three types of evidence: "external sound, stability, and self-noise drift," thereby improving its ability to distinguish between weak sound sources and changes in the body's state.

[0090] In this embodiment, the improved ConvNeXt V2 model includes a harmonic sensitive band gating layer, a ConvNeXtV2 convolutional backbone, a conditional GRN module, a classification header, and an evidence header.

[0091] The harmonic sensitive frequency band gating layer receives a three-channel evidence spectrum tensor and a working condition vector. It reads the rotational speed field from the working condition vector and generates a set of harmonic frequency band positions and corresponding bandwidth parameters. Based on the set of harmonic frequency band positions, it constructs a frequency band weight vector on the frequency axis of the spectrum. It then performs interval truncation and neighborhood smoothing on the frequency band weight vector to obtain a gating weight vector. The gating weight vector is copied along the time dimension to form a gating weight map. Each channel of the three-channel evidence spectrum tensor is multiplied and weighted by frequency band. Frequency bands that fall into the set of harmonic frequency band positions are multiplied by a suppression weight less than one, while frequency bands that do not fall into the set of harmonic frequency band positions are multiplied by a preservation weight equal to one. The gating three-channel evidence spectrum tensor is then output.

[0092] The ConvNeXt V2 convolutional backbone includes a first convolutional embedding layer, first to fourth level feature extraction stages, and a downsampling layer. Each level feature extraction stage is a feature extraction unit composed of several ConvNeXt V2 convolutional blocks connected in series. Adjacent level feature extraction stages are connected by downsampling layers, which reduces the spatial resolution of the feature map and increases the number of channels. The first convolutional embedding layer is a convolutional embedding layer with a stride of 4, which maps the gated three-channel evidence spectrum tensor to a first feature map through convolution. The downsampling layer is set between adjacent level feature extraction stages and connects adjacent level feature extraction stages in sequence. The downsampling layer includes a layer normalization layer and a convolutional layer with a stride of 2, which is used to downsample the output feature map of the previous level feature extraction stage and increase the number of channels. The feature map flows in hierarchical order to the output of the fourth level feature extraction stage.

[0093] Each ConvNeXt V2 convolutional block sequentially includes a depthwise separable convolutional layer, a layer normalization layer, a first point convolutional layer, a GELU activation layer, a conditional GRN module, a second point convolutional layer, and a residual connection layer. The depthwise separable convolutional layer is a 7×7 convolutional layer grouped by channel, which performs local time-frequency texture extraction on the feature map. The layer normalization layer normalizes the channel dimension. The first point convolutional layer is a 1×1 convolutional layer used to expand the channel dimension. The GELU activation layer performs GELU activation on the expanded channel features. The second point convolutional layer is a 1×1 convolutional layer that projects the channel dimension back to the number of channels input to the convolutional block. The residual connection layer adds the feature map input to the convolutional block and the feature map output by the second point convolutional layer element-wise to obtain the output feature map of the convolutional block.

[0094] The conditional GRN module is embedded in each ConvNeXt V2 convolutional block of the first-level feature extraction stage and the second-level feature extraction stage. It receives the channel features and conditional vectors output by the GELU activation layer, inputs the conditional vectors into the multilayer perceptron, and generates a scaling vector and a bias vector consistent with the number of channels. The multilayer perceptron includes a fully connected layer, a SiLU activation layer and a fully connected layer in sequence. It performs global aggregation on the channel features in the spatial dimension to obtain the global response intensity of each channel, and performs normalization on the global response intensity to obtain the intermediate GRN features. The intermediate GRN features are multiplied element-wise by the scaling vector by channel, and added element-wise by the bias vector by channel to obtain the conditional GRN output and pass it to the second point convolutional layer.

[0095] The evidence head and classification head are set in parallel after the output of the fourth-level feature extraction stage. The classification head performs global average pooling on the feature map output by the fourth-level feature extraction stage to obtain a global feature vector, and outputs the class probability through a fully connected layer. At the same time, temperature calibration is performed on the class probability to obtain the confidence level. The evidence head performs 1×1 convolution on the feature map output by the fourth-level feature extraction stage to obtain an evidence heatmap, performs normalization on the evidence heatmap, and selects the Top-K positions according to the response value from large to small. The Top-K positions are mapped back to the time-frequency block index of the three-channel evidence spectrum tensor before gating, and the corresponding response value is used as the weight to form a key evidence set.

[0096] This implementation uses ConvNeXt V2 as the convolutional backbone to achieve efficient feature extraction from UAV spectrograms, balancing real-time performance at the edge and the ability to represent time-frequency textures. Compared to traditional CNNs, it is more suitable for obtaining higher feature quality under limited computing power. Furthermore, this implementation differs from the original approach of "directly inputting spectrograms into a general network and performing only simple filtering and noise reduction." It introduces specialized improvements for flight conditions within the model: a harmonic-sensitive frequency band gating layer is added at the input, locating harmonic frequency bands according to motor speed and suppressing them band-by-band multiplication, preventing the model from being dominated by propeller noise; a conditional GRN module is embedded in the first and second level feature extraction stages, using conditional vectors to generate scaling and bias, and performing channel multiplication-addition modulation on the normalized channel responses to enhance robustness across speed, wind conditions, and vibration variations; a classification head and an evidence head are connected in parallel at the output, outputting not only the class probability and confidence score but also a set of key evidence, improving reproducibility and engineering usability, thereby reducing false positives and false negatives and improving stable recognition capabilities in complex environments.

[0097] In this embodiment, step six specifically includes:

[0098] Based on category probability, confidence level, key evidence set, wind noise index, vibration consistency weight and self-noise deviation index, a hierarchical inference trigger field is generated. The hierarchical inference trigger field includes low confidence trigger identifier, wind noise trigger identifier, vibration trigger identifier and deviation trigger identifier.

[0099] When the confidence level is less than the preset confidence threshold, the low confidence trigger flag is set. When the wind noise index is greater than the preset wind noise threshold, the wind noise trigger flag is set. When the statistic of the vibration consistency weight is greater than the preset vibration threshold, the vibration trigger flag is set. When the statistic of the self-noise deviation index is greater than the preset deviation threshold, the deviation trigger flag is set. When any trigger flag is set, the re-judgment process is initiated.

[0100] Before entering the review process, the audio data is windowed using the first inference window length and a three-channel evidence spectrum tensor is generated. The improved ConvNeXt V2 model is called to output the first category probability, the first confidence level and the first key evidence set, which are then output as the classification result package.

[0101] When entering the review process, the inference window length is switched from the first inference window length to the second inference window length, and the low-frequency suppression weight is switched to the enhanced suppression weight in the gated low-frequency suppression processing. At the same time, in the harmonic sensitive frequency band gated layer of the improved ConvNeXt V2 model, the suppression weight of the harmonic frequency band is switched to the enhanced suppression weight. Based on the second inference window length, the three-channel evidence spectrum tensor is regenerated, and the improved ConvNeXt V2 model is called to output the second category probability, the second confidence level, and the second key evidence set. The final classification result package is generated with the category corresponding to the second category probability as the final category, the second confidence level as the final confidence level, and the second key evidence set as the final key evidence set.

[0102] An inference path identifier is written into the final classification result package. The inference path identifier is used to indicate that the final classification result package comes from the first inference window length inference path that has not entered the review process or the second inference window length review path that has entered the review process, and is written into the trigger identifier set.

[0103] This implementation method triggers hierarchical reasoning and re-judgment in step six using confidence level and wind noise, vibration, and deviation indicators. During re-judgment, the window length is extended and the gating strength is enhanced. Unlike the "one-size-fits-all" approach of single reasoning, this method re-confirms difficult segments without significantly increasing the average time delay, thereby reducing false alarms and missed alarms under strong winds and speed fluctuations.

[0104] In this embodiment, step seven specifically includes:

[0105] The confidence level, inference path identifier, and trigger identifier set are read from the final classification result package. When the inference path identifier indicates that the review process has been initiated or the confidence level is less than the preset cache threshold, the three-channel evidence spectrum tensor, working condition vector, category probability, confidence level, and key evidence set of the corresponding time window are written into the cache queue. The cache queue uses the standard timestamp and flight mission identifier as index fields.

[0106] Set a capacity limit for the cache queue. When the cache queue reaches the capacity limit, sort the samples by confidence level from smallest to largest and retain samples with lower confidence level first. When there are samples with the same confidence level, sort them by the number of set flags in the trigger flag set from largest to smallest and retain samples with more set flags first. When there are still samples with the same confidence level, sort them by standard timestamp from newest to oldest and retain newer samples first.

[0107] Online lightweight adaptive updates are triggered at a preset update cycle. The online lightweight adaptive updates include normalized statistical updates and gating threshold updates. The normalized statistical updates are as follows: without changing the convolutional weights and fully connected weights of the improved ConvNeXt V2 model, the mean and variance statistics of the normalized layer in the ConvNeXt V2 convolutional backbone are updated by moving average, and the updated statistics are written into the parameters of the improved ConvNeXt V2 model. The gating threshold updates are as follows: based on the wind noise index, vibration consistency weight and self-noise deviation index of the samples in the cache queue, the quantile values ​​are calculated respectively, and the preset wind noise threshold, preset vibration threshold and preset deviation threshold are updated accordingly. At the same time, the upper limit of the low frequency suppression weight and the upper limit of the suppression weight of the harmonic sensitive frequency band gating layer are also updated.

[0108] The updated normalized statistics and gating thresholds are applied to the spectrogram processing of audio data, the construction of three-channel evidence spectrogram tensors, and the inference process of the improved ConvNeXt V2 model, and the updated classification result package is continuously output.

[0109] This implementation method caches low-confidence samples and filters them according to confidence level, trigger flag, and time strategy. It periodically slides and updates the normalized statistics and updates the gating threshold according to the quantile value. Unlike offline fixed models that are not adaptable to drift, this method achieves mild adaptation to model differences, load changes, and noise drift during long-term operation, and maintains continuous and stable output.

[0110] Example 1:

[0111] To verify the feasibility of this invention in practice, it was applied to the "environmental sound source identification and anomaly alarm" task of a nighttime security patrol drone in an urban area. The drone is equipped with a single microphone and a flight control data interface, flies at an altitude of 30–80m, and a speed of 6–10m / s. Typical background noise includes propeller harmonics, its own aerodynamic noise, and road traffic noise. Target identification categories are set as six types: human cries for help, vehicle horns, alarms, construction machinery, other drones, and background noise.

[0112] The audio sampling rate was set to 16kHz, 16bit, and frames were shifted between 32ms and 16ms frames. Time-frequency transformation used an STFT window with a length of 1024 and a step size of 256, followed by logarithmic compression to obtain the initial spectrogram. A table of fundamental frequency and 1st–8th order harmonic positions was generated based on the motor speed field. The harmonic peak search interval was set to ±15Hz of the target frequency. Harmonic peaks were extracted based on local maxima, and frequency axis resampling was performed according to the frequency band alignment offset to complete harmonic alignment normalization. Wind noise was defined as the proportion of low-frequency energy in the 20–200Hz range and its inter-frame variation. The wind noise threshold was set to 0.35, with a low-frequency suppression weight of 0.45 when triggered and 1.0 otherwise. Vibration characteristics were constructed by concatenating the mean and peak values ​​of the absolute values ​​of the IMU's three-axis accelerations. The vibration threshold was 2.2g. A weighting factor ranging from 0.6 to 0.95 was applied to the frequency band synchronized with vibration changes, followed by frequency neighborhood smoothing and inter-frame moving average. The three-channel evidence spectrum tensor consists of an enhanced main spectrum, a phase stability evidence channel (in single-microphone scenarios, phase continuity and spectral stability are truncated and normalized, then weighted at 0.6 / 0.4 and smoothed in the neighborhood), and a self-noise bias evidence channel. The operating condition vector is spliced ​​together based on motor speed, throttle opening, attitude angle, wind noise index, vibration intensity sequence statistics, and self-noise bias statistics. The improved ConvNeXt V2 is deployed on an airborne computing board using the Tiny specification. The Top-K evidence is set to 5, the confidence threshold is set to 0.60, and the cache threshold is set to 0.55. The hierarchical inference window length is 1.5s in the normal path and 3.0s in the re-judgment path. At the same time, the low-frequency suppression weight is reduced to 0.35 and the harmonic suppression weight of the harmonic sensitive frequency band gating layer is increased from 0.85 to 0.65 to ensure stable discrimination under strong wind and high speed conditions. The online lightweight adaptive method updates every 60 seconds, only sliding to update the normalized statistics of the convolutional backbone layer (momentum 0.02) and updating the three-class thresholds and two-class suppression upper limits according to the wind noise / vibration / bias quantile values ​​of the cached samples.

[0113] To demonstrate the comparative effect, two common solutions were selected as existing technologies for comparison in the same scenario: Control A used "spectral mapping + fixed bandpass (300–3400Hz) + ResNet18 classification", without introducing rotational harmonic alignment and multi-source synchronization; Control B used "spectral mapping + lightweight CNN (MobileNet V2) + simple spectral subtraction and noise reduction", similarly without utilizing the consistency of operating conditions and vibration for weighting, and without outputting a key evidence set. The three methods were evaluated in 12 consecutive missions, each flight lasting approximately 18–25 minutes, totaling approximately 14.6 hours of effective audio. Approximately 9,800 event segments were manually sampled as statistical samples. Indicators included macro-average F1, false alarm rate (number of false alarms per 10 minutes), false negative rate, end-side single-window latency, and evidence traceability success rate (manual verification of whether the Top-5 evidence covered the main energy region of the event; coverage was considered successful). The results are shown in Table 1 below.

[0114] Table 1 Summary of Audio Classification Comparison and Evaluation for UAVs in Urban Scenarios

[0115] method Macro average F1 Overall accuracy False alarm rate (times / 10min) Missed Reporting Rate (%) Single window delay (ms) Evidence traceability success rate (%) Strong winds (≥6m / s) F1 Speed ​​fluctuation (±18%) F1 Compare with A 0.836 0.872 1.92 11.4 38 0 0.772 0.781 Compare with B 0.859 0.889 1.47 9.6 29 0 0.801 0.804 This invention 0.918 0.934 0.61 5.2 34 92.7 0.883 0.889

[0116] As shown in Table 1, the macro-average F1 of this invention reaches 0.918, an improvement of 0.082 compared to control A and 0.059 compared to control B. The false alarm rate decreased from 1.92 / 1.47 to 0.61 times / 10min, indicating that the alarms are "cleaner" in actual cruise, and maintenance personnel do not need to repeatedly confirm invalid alarms. The false alarm rate decreased from 11.4% and 9.6% to 5.2%, reflecting that weak signal events such as distress calls and sirens are less likely to be drowned out by propeller noise and wind noise. In terms of end-side latency, the single-window latency of this invention is 34ms, which is slightly higher than control B but still meets the real-time requirements. Moreover, through the hierarchical inference mechanism, the re-judgment is limited to a few periods such as low confidence, strong wind, strong vibration, and abnormal deviation, and the actual average power consumption of the whole machine does not fluctuate significantly. More importantly, in the two "most difficult operating conditions" of strong wind and speed fluctuation, the F1 of the comparison A / B dropped to 0.772 / 0.801 and 0.781 / 0.804 respectively, which means that the low-frequency rise of wind noise and harmonic drift were misjudged as alarms or construction machinery. Under the combined effect of harmonic alignment normalization, gating suppression and vibration consistency reduction, the present invention still maintains 0.883 and 0.889, which shows obvious advantages in robustness under operating conditions.

[0117] To further validate the engineering value of the "evidence header + key evidence set" approach, 120 alarm segments from strong wind night flights were extracted for review. The existing method only provides category results, requiring manual replaying to troubleshoot false alarms. This invention directly outputs a Top-K time-frequency block index, allowing reviewers to pinpoint the cause by first examining the audio segments corresponding to the evidence blocks. Statistics show an evidence traceability success rate of 92.7%, with most false alarms quickly attributable to sudden increases in wind noise or aircraft resonance, facilitating adjustments to the wind noise threshold and low-frequency suppression upper limit. Furthermore, for the "other drones" category, the evidence blocks are concentrated in a continuous narrow band region of mid-to-high frequencies, differentiating them from the broadband transient differences of vehicle horns, thus improving on-site interpretability and review efficiency.

[0118] The comprehensive implementation results show that, without adding complex peripherals, this invention significantly reduces false alarms and missed alarms caused by cross-operating condition distribution drift by achieving frame-level synchronization of audio with operating conditions and vibration, superimposing harmonic alignment normalization driven by motor speed, low-frequency suppression of wind noise gating, and frequency band weighting of vibration consistency. By introducing harmonic sensitive frequency band gating at the input end through the improved ConvNeXt V2 model, embedding operating condition-conditional GRN in a shallow layer, and connecting the classification head and evidence head in parallel, stable classification and traceable evidence output that can be deployed in real time at the end side are achieved. Combined with confidence-triggered hierarchical reasoning and online lightweight adaptive updates, the system can still work stably for a long time under changes in wind conditions, speed fluctuations, and machine state drift, demonstrating outstanding engineering importance and promotional value.

[0119] 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 classifying drone audio based on convolutional neural networks, characterized in that, Includes the following steps: Step 1: Collect audio data, operational status data, and vibration data during the drone's flight process, align the execution time, and obtain the operational status data and vibration data corresponding to each audio frame; Step 2: Perform frame segmentation and time-frequency transformation on the audio data to generate a spectrogram, and perform harmonic alignment and normalization processing on the spectrogram based on the motor speed in the operating condition data; Step 3: Based on the wind noise index, perform gated low-frequency suppression processing on the spectrogram, and generate vibration consistency weights based on vibration data to perform weight reduction processing on the corresponding frequency bands of the spectrogram to obtain the enhanced spectrogram; Step 4: Construct the main spectrum channel from the enhanced spectrogram, and construct the phase stability evidence channel and the self-noise deviation evidence channel to form a three-channel evidence spectrum tensor, and generate the operating condition vector at the same time; Step 5: Input the three-channel evidence spectrum tensor and the operating condition vector into the improved ConvNeXt V2 model. The improved ConvNeXt V2 model includes a harmonic sensitive frequency band gate layer, a ConvNeXt V2 convolution backbone, an operating condition-conditioned GRN module, a classification head, and an evidence head. The multi-head outputs the class probability, confidence level, and key evidence set. Step 6: Based on the confidence level, wind noise index, vibration index, and self-noise deviation index, trigger the hierarchical reasoning and re-judgment process, and generate the classification result package by adjusting the window length and gating strength; Step 7: Cache the low-confidence samples and their corresponding evidence spectrum tensors, and perform online lightweight adaptive updates on the normalized statistics and gating thresholds to output a continuously updated classification result package.

2. The UAV audio classification method based on convolutional neural networks according to claim 1, characterized in that, Step one specifically includes: Collect audio data, operating status data, and vibration data during the drone's flight, and write a first timestamp for each audio frame; The operating condition data includes motor speed, throttle opening and attitude angle data, and a second timestamp is written for each piece of operating condition data. The vibration data consists of triaxial acceleration data, and a third timestamp is written for each vibration data point. Using the audio frame time index corresponding to the first timestamp as the alignment reference, the second and third timestamps are mapped to the audio frame time index; When the operating condition data or vibration data is missing at a certain audio frame time index, it is filled in by keeping the previous valid value, and a missing identifier is written at the filling position to obtain the operating condition features and vibration features corresponding to each audio frame.

3. The UAV audio classification method based on convolutional neural networks according to claim 1, characterized in that, Step two specifically includes: The audio data is divided into frames. After windowing, a fast Fourier transform is performed on each frame. The frames are then stitched together in order to obtain the time-frequency amplitude spectrum. Logarithmic compression is then performed on the amplitude spectrum to generate the initial spectrogram. The speed level corresponding to the current audio frame is determined based on the motor speed, and a harmonic position table is generated based on the speed level. According to the harmonic location table, the harmonic peak search interval is determined on the frequency axis of the initial spectrum. The local maximum frequency point is extracted in each harmonic peak search interval as the harmonic peak frequency point. The harmonic peak frequency point is aligned and mapped with the target frequency point in the harmonic location table to form the frequency band alignment offset. Based on the frequency band alignment offset, the frequency axis of the initial spectrogram is resampled and aligned so that the harmonic peak frequencies of each audio frame are mapped to the same frequency band index position in the resampled spectrogram. Then, amplitude normalization is performed on the resampled spectrogram to obtain the harmonic aligned normalized spectrogram.

4. The UAV audio classification method based on convolutional neural networks according to claim 1, characterized in that, Step three specifically includes: Based on the harmonic-aligned normalized acoustic spectrum, the spectral energy of each audio frame in the low-frequency range is statistically analyzed according to the preset low-frequency range, and the spectral energy of the audio frame in the full frequency range is also statistically analyzed. The proportion of low-frequency energy is calculated as the low-frequency energy index. At the same time, the change in the low-frequency energy index between adjacent audio frames is calculated as the low-frequency change index. The low-frequency energy index and the low-frequency change index are combined to generate the wind noise index. Based on the comparison between the wind noise index and the preset wind noise threshold, when the wind noise index exceeds the preset wind noise threshold, a low-frequency suppression weight is generated, and the spectral amplitude value of the corresponding low-frequency range in the harmonic aligned normalized sound spectrum is scaled according to the low-frequency suppression weight to obtain the wind noise suppression sound spectrum. When the wind noise index does not exceed the preset wind noise threshold, the low-frequency suppression weight is set to no scaling weight and the wind noise suppression sound spectrum is output. Within the time window corresponding to the current audio frame, the mean and peak values ​​of the absolute values ​​of acceleration are calculated for the vibration data along the three axes to form a vibration intensity sequence. Based on the vibration intensity sequence and the preset vibration threshold, a vibration consistency weight is generated. The vibration consistency weights are mapped to the frequency axis of the wind noise suppression spectrogram, and the spectral amplitude of the mapped frequency band is reduced and scaled to obtain the enhanced spectrogram.

5. The UAV audio classification method based on convolutional neural networks according to claim 1, characterized in that, Step four specifically includes: The enhanced spectrogram is cropped or scaled according to the preset time window length and preset number of frequency bands to be uniformly of the preset size, and the enhanced spectrogram of the preset size is used as the main spectrogram channel. When the drone is configured as an airborne microphone array, time-frequency transformation is performed on the audio frames of each microphone within the same time window to obtain multiple spectrograms. The cross-spectral consistency index of any two spectrograms at the corresponding time-frequency positions is calculated. The cross-spectral consistency index is spliced ​​in the frequency dimension and the time dimension to form a phase stability evidence channel. When the drone is configured as a single microphone, the phase continuity index and spectral stability index are calculated between adjacent audio frames of the enhanced spectrogram. The phase continuity index and spectral stability index are mapped according to the preset fusion rules to form a phase stability evidence channel. Based on the harmonic location table, the peak search intervals corresponding to each order harmonic are determined in the enhanced acoustic spectrum. The peak amplitude and peak width features within the peak search interval are extracted, and the peak amplitude and peak width features are compared with the preset self-noise template features to obtain the self-noise deviation index map as a self-noise deviation evidence channel. The main spectrum channel, phase stability evidence channel, and self-noise deviation evidence channel are spliced ​​together in the channel dimension to form a three-channel evidence spectrum tensor. The operating condition data, wind noise index, vibration intensity sequence, and self-noise deviation index are then spliced ​​together in sequence to form an operating condition vector.

6. The UAV audio classification method based on convolutional neural networks according to claim 1, characterized in that, The improved ConvNeXt V2 model includes a harmonic-sensitive band gated layer, a ConvNeXt V2 convolutional backbone, a conditional GRN module, a classification header, and an evidence header. The harmonic sensitive frequency band gating layer receives the three-channel evidence spectrum tensor and the operating condition vector, reads the rotation speed field from the operating condition vector and generates a set of harmonic frequency band positions and corresponding bandwidth parameters, constructs a frequency band weight vector on the frequency axis of the spectrum based on the set of harmonic frequency band positions, and performs interval truncation and neighborhood smoothing on the frequency band weight vector to obtain the gating weight vector, copies the gating weight vector along the time dimension to form a gating weight map, performs element-wise multiplication and weighting on each channel of the three-channel evidence spectrum tensor according to the frequency band, and outputs the gating three-channel evidence spectrum tensor. The ConvNeXt V2 convolutional backbone includes a first convolutional embedding layer, first to fourth level feature extraction stages, and a downsampling layer. The first convolutional embedding layer maps the gated three-channel evidence spectrum tensor into a first feature map through convolution. The downsampling layer is set between adjacent level feature extraction stages and sequentially connects adjacent level feature extraction stages. The feature map flows in hierarchical order to the fourth level feature extraction stage for output. Each ConvNeXt V2 convolutional block sequentially includes a depthwise separable convolutional layer, a layer normalization layer, a first point convolutional layer, a GELU activation layer, a conditional GRN module, a second point convolutional layer, and a residual connection layer. The depthwise separable convolutional layer is a channel-grouped convolutional layer that performs local time-frequency texture extraction on the feature map. The first point convolutional layer expands the channel dimension. The second point convolutional layer projects the channel dimension back to the number of channels input to the convolutional block. The residual connection layer adds the feature map input to the convolutional block and the feature map output by the second point convolutional layer element-wise to obtain the output feature map of the convolutional block. The conditional GRN module is embedded in each ConvNeXt V2 convolutional block of the first-level feature extraction stage and the second-level feature extraction stage. It receives the channel features and conditional vectors output by the GELU activation layer, inputs the conditional vectors into the multilayer perceptron, generates a scaling vector and a bias vector consistent with the number of channels, performs global aggregation on the channel features in the spatial dimension to obtain the global response intensity of each channel, and performs normalization on the global response intensity to obtain the GRN intermediate features. The GRN intermediate features are multiplied element-wise by the scaling vector by channel and added element-wise by the bias vector by channel to obtain the conditional GRN output and pass it to the second point convolutional layer. The evidence head and classification head are set in parallel after the output of the fourth-level feature extraction stage. The classification head performs global average pooling on the feature map output by the fourth-level feature extraction stage to obtain a global feature vector, and outputs the class probability through a fully connected layer. At the same time, temperature calibration is performed on the class probability to obtain the confidence level. The evidence head performs convolution on the feature map output by the fourth-level feature extraction stage to obtain an evidence heatmap, performs normalization on the evidence heatmap, and selects the Top-K positions according to the response value from large to small. The Top-K positions are mapped back to the time-frequency block index of the three-channel evidence spectrum tensor before gating, and the corresponding response value is used as the weight to form a key evidence set.

7. The UAV audio classification method based on convolutional neural networks according to claim 1, characterized in that, Step six specifically includes: Based on category probability, confidence level, key evidence set, wind noise index, vibration consistency weight and self-noise deviation index, a hierarchical inference trigger field is generated. The hierarchical inference trigger field includes low confidence trigger identifier, wind noise trigger identifier, vibration trigger identifier and deviation trigger identifier. When any trigger flag is set, the review process is initiated. If the review process is not initiated, the audio data is windowed using the first inference window length and a three-channel evidence spectrum tensor is generated. The improved ConvNeXt V2 model is called to output the first category probability, the first confidence level, and the first key evidence set, which are then output as the classification result package. When entering the review process, the inference window length is switched from the first inference window length to the second inference window length, and the low-frequency suppression weight is switched to the enhanced suppression weight in the gated low-frequency suppression processing. At the same time, in the harmonic sensitive frequency band gated layer of the improved ConvNeXt V2 model, the suppression weight of the harmonic frequency band is switched to the enhanced suppression weight. Based on the second inference window length, the three-channel evidence spectrum tensor is regenerated, and the improved ConvNeXt V2 model is called to output the second category probability, the second confidence level, and the second key evidence set. The final classification result package is generated with the category corresponding to the second category probability as the final category, the second confidence level as the final confidence level, and the second key evidence set as the final key evidence set. Write the inference path identifier and the trigger identifier set into the final classification result package.

8. The UAV audio classification method based on convolutional neural networks according to claim 1, characterized in that, Step seven specifically includes: Read the confidence level, inference path identifier and trigger identifier set from the final classification result package. When the inference path identifier indicates that the review process is to be entered or the confidence level is less than the preset cache threshold, write the three-channel evidence spectrum tensor, working condition vector, category probability, confidence level and key evidence set of the corresponding time window into the cache queue. Set a capacity limit for the cache queue. When the cache queue reaches the capacity limit, sort the samples by confidence level from smallest to largest and retain samples with lower confidence level first. When there are samples with the same confidence level, sort them by the number of set flags in the trigger flag set from largest to smallest and retain samples with more set flags first. When there are still samples with the same confidence level, sort them by standard timestamp from newest to oldest and retain newer samples first. Online lightweight adaptive updates are triggered at a preset update cycle, and the online lightweight adaptive updates include normalized statistical updates and gating threshold updates; The updated normalized statistics and gating thresholds are applied to the spectrogram processing of audio data, the construction of three-channel evidence spectrogram tensors, and the inference process of the improved ConvNeXt V2 model, and the updated classification result package is continuously output.

Images