An acoustic scene cognition and multi-source noise reduction unmanned aerial vehicle ecological intelligent perception monitoring system

Abstract

The application provides a sound scene cognition and multi-source noise reduction unmanned aerial vehicle ecological intelligent perception monitoring system, and belongs to the technical field of unmanned aerial vehicle acoustic monitoring, and comprises: a multi-source dynamic noise reduction and sound scene collection module, which is used for constructing a noise field and reducing noise through an unmanned aerial vehicle flight state to obtain an initial ecological sound scene signal; a sparsity evolution sound scene separation module, which is used for performing frequency domain transformation and dynamic sparse separation on the initial ecological sound scene signal to obtain a species voiceprint feature matrix; a sound scene semantic navigation decision module, which is used for fusing unmanned aerial vehicle position data and the species voiceprint feature matrix to realize autonomous route adjustment of the unmanned aerial vehicle; an end cloud collaborative sound scene self-learning module, which is used for realizing continuous iteration through end side recognition and cloud aggregation optimization; and an ecological monitoring result output module, which is used for generating a structured report and pushing the report to a user. The application solves the problems of distorted monitoring data, low efficiency and poor applicability caused by noise interference, low separation precision, rigid navigation mechanism and lack of self-adaption of the model.

Description

Technical Field

[0001] This invention belongs to the field of acoustic monitoring technology for unmanned aerial vehicles (UAVs), and particularly relates to an intelligent ecological perception and monitoring system for UAVs with soundscape cognition and multi-source noise reduction. Background Technology

[0002] Unmanned aerial vehicle (UAV) ecosystem monitoring is a key technology for biodiversity conservation and ecological function assessment. In complex field applications, existing technologies suffer from several challenges at the data acquisition level. The electromechanical noise and wind noise generated by the UAVs themselves during flight highly overlap with the ecological soundscapes such as birdsong and insect calls that need to be collected in the frequency domain. Furthermore, existing noise reduction schemes are mostly based on static parameter settings, making it difficult to adapt to the dynamic flight states of UAVs, such as speed changes and turns. This results in severe distortion of ecological sound signatures, with the effective acoustic data ratio generally below 30%. Although existing technologies have improved this through audio noise reduction and feature extraction, they still struggle to effectively separate dynamically changing flight noise, resulting in limited improvement in the signal-to-noise ratio.

[0003] In terms of soundscape signal processing, current technologies often face challenges due to the simultaneous presence and mutual interference of multiple species' acoustic signatures in the wild. Traditional sound source separation algorithms are not optimized for such ecological scenarios, making it difficult to accurately separate and extract the acoustic features of a single species and hindering in-depth soundscape cognitive analysis. Furthermore, existing drones rely heavily on visual sensors or satellite positioning to plan fixed routes. In adverse weather conditions such as fog, rain, and snow, or in areas with severe signal obstruction such as under forest canopies, navigation reliability is significantly reduced. They cannot intelligently perceive and proactively fly to active areas with high biodiversity, resulting in a passive and inefficient monitoring mode. Moreover, existing monitoring systems lack adaptive capabilities, failing to self-optimize in response to seasonal and regional changes in soundscape characteristics. Frequent manual intervention to update parameters is required, leading to high maintenance costs and difficulty in guaranteeing long-term monitoring accuracy. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a soundscape cognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system. This system solves the problems of distorted ecosystem monitoring data, low efficiency, and poor long-term applicability caused by severe interference from UAV dynamic flight noise, insufficient accuracy in multi-species voiceprint separation, lack of soundscape semantic driving in navigation mechanisms, and lack of adaptive capabilities in monitoring models. To achieve the above objectives, the technical solution adopted by this invention is: a soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system, comprising: The multi-source dynamic noise reduction and sound scene acquisition module is used to acquire the drone flight state feature vector, construct the drone noise field based on the flight state feature vector, and use a dual-microphone array that combines the drone noise field and the U-Net-TemporalConformer deep learning network to acquire and reduce the sound scene signal to obtain the initial ecological sound scene signal. The sparse evolutionary soundscape separation module is used to perform frequency domain transformation on the initial ecological soundscape signal and to perform multi-source separation under dynamic sparsity constraints through an improved convolutional temporal audio separation network to obtain the species voiceprint feature matrix. The soundscape semantic navigation decision module is used to fuse UAV location data with species voiceprint feature matrix to obtain soundscape semantic map, process it based on reinforcement learning and model prediction control algorithm, output UAV flight path command, and obtain abnormal soundscape data and abnormal sound source coordinates based on UAV flight path command through anomaly detection and sound source localization calculation. The edge-cloud collaborative soundscape self-learning module is used to process the species voiceprint feature matrix on the UAV side through a lightweight voiceprint recognition model, output uncertain voiceprint samples, transmit the uncertain voiceprint samples to the cloud for multi-machine data fusion, and optimize the lightweight voiceprint recognition model based on the momentum gradient descent algorithm to obtain the optimized lightweight voiceprint recognition model. The ecological monitoring results output module is used to calculate ecological monitoring indicators based on species voiceprint feature matrix, soundscape semantic map and UAV flight path instructions, and push the generated structured monitoring report to different user terminals.

[0005] Furthermore: the multi-source dynamic noise reduction and sound scene acquisition module includes: The multimodal flight state perception unit is used to collect UAV flight attitude parameters through IMU sensors, UAV dynamic parameters through propeller speed sensors, and UAV airflow disturbance parameters through airflow sensors. It then processes the UAV flight attitude parameters, UAV dynamic parameters, and UAV airflow disturbance parameters through time synchronization, normalization, and weighted calculation to obtain the UAV flight state feature vector. The multi-source noise field modeling unit is used to obtain the dynamic noise field of the UAV in flight state by modeling through frequency domain mapping and introducing dynamic correction of the flight state change rate based on the UAV flight state feature vector; The dual-microphone array noise reduction unit is used to acquire the original sound scene signal through a dual-microphone array, and combine it with the dynamic noise field and U-Net-TemporalConformer deep learning network for noise reduction and temporal reconstruction to obtain the temporal noise reduction scene signal. The initial soundscape acquisition unit is used to quantize and encode the time-domain noise reduction soundscape signal and output the initial ecological soundscape signal.

[0006] The further beneficial effects mentioned above are as follows: This invention transforms frequently changing flight noise into a predictable physical model through multimodal flight state perception and dynamic noise field modeling, solving the problem of noise reduction failure caused by dynamic noise changes in existing technologies; combined with a dual-microphone array and U-Net-TemporalConformer deep learning network for guided deep noise reduction, it achieves accurate extraction of ecological soundscapes under strong dynamic noise background, improving the signal-to-noise ratio and fidelity of the original sound signal.

[0007] Furthermore, the expression for the UAV flight state feature vector is as follows:

[0008]

[0009]

[0010]

[0011]

[0012] in, This is the feature vector of the UAV's flight state. For time, Assigning weights to IMU sensors, Assigning weights to the propeller speed sensor, Assigning weights to the airflow sensors, These are the normalized flight attitude parameters of the UAV. The normalized UAV dynamic parameters, These are the normalized airflow disturbance parameters for the UAV. This is the calibration minimum value for the IMU sensor. For the calibration maximum value of the IMU sensor, This is the calibration minimum value for the propeller speed sensor. This is the maximum calibration value for the propeller speed sensor. This is the calibration minimum value for the airflow sensor. This is the calibration maximum value for the airflow sensor. For the flight attitude parameters of the UAV, For roll angle, The pitch angle, Yaw angle For the roll rate, For pitch rate, The yaw rate is... For the power parameters of the drone, For the first propeller speed, These are the airflow disturbance parameters for the UAV. Wind speed in the horizontal direction. Wind speed along the vertical axis. The wind speed is in the vertical direction. Let T be the air pressure and T be the transpose.

[0013] Furthermore, the expression for the dynamic noise field during the flight of the UAV is as follows:

[0014]

[0015]

[0016]

[0017] in, For dynamic noise field, For frequency, For time, This represents the mapping relationship between the flight state of the UAV and the noise power spectral density (PSD). To adjust the coefficient, For the rate of change of flight state, It is the L2 norm. It is the ReLU activation function. To train the weight matrix, For bias terms, Indexes for the neuron dimensions / feature dimensions of the network model. This is the feature vector of the UAV's flight state. For drone noise, For electromechanical noise, It is wind noise.

[0018] Furthermore, the expression for the time-domain noise-reduced scene signal is as follows:

[0019]

[0020]

[0021]

[0022]

[0023] in, For time-domain noise reduction of scene signals, This is the inverse short-time Fourier transform. The output of the U-Net-TemporalConformer deep learning network For the U-Net-TemporalConformer deep learning network, For Mel spectrum conversion, For pre-filtered signals, For optimized dynamic noise field, For frequency, For time, For smoothing coefficients, For threshold coefficient, For indicator functions, For dynamic noise field, It is a phase weighting vector. It is the conjugate transpose. The first signal after the short-time Fourier transform. The second signal is the result of the short-time Fourier transform. This is the first signal in the original soundscape signal. This is the second signal in the original soundscape signal. For the target soundscape signal, and All of these are noise signals.

[0024] Furthermore: the sparsity evolution soundscape separation module includes: The soundscape frequency domain feature transformation unit is used to perform short-time Fourier frequency domain transformation on the initial ecological soundscape signal to obtain a frequency domain complex matrix; The sparsity evolution separation unit is used to obtain the number of species from the frequency domain complex matrix through the spectral peak detection algorithm, obtain the sparsity threshold based on the number of species, combine the sparsity threshold with the convolutional temporal audio separation network to obtain an improved convolutional temporal audio separation network, and perform source separation on the initial ecological sound scene signal to obtain the single species voiceprint signal. The soundscape feature extraction unit is used to extract multi-dimensional features from the voiceprint signal of a single species to obtain the voiceprint feature matrix for each species.

[0025] The further beneficial effects mentioned above are as follows: This invention perceives the number of species in the soundscape in real time by detecting spectral peaks and dynamically adjusts the sparsity constraints accordingly, which solves the problem of inaccurate voiceprint separation caused by unknown or changing number of species in traditional separation algorithms; combined with an improved convolutional temporal audio separation network, it achieves high-precision separation of voiceprints of multiple species and multi-dimensional feature extraction, which significantly improves the precision and robustness of soundscape analysis.

[0026] Furthermore, the expression for the single-species voiceprint signal is as follows:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] in, For a single species' voiceprint signal, it indicates Time of the first Voiceprint signals of each species For the first Species, For time, For transposed convolution, This is the frequency domain species voiceprint signal. For a dynamic sparse layer based on a sparsity threshold, For the first Mask matrix for each species, For input feature representation, The sparsity threshold, For indicator functions, It is the ReLU activation function. For 1D convolution, This is the initial ecological soundscape signal. Based on sparsity, It is a natural exponential function. For adjustment coefficients, For the number of species, It is a complex matrix in the frequency domain. Let f be the amplitude of the frequency domain complex matrix spectrum at time t′. Peak threshold Let be the mathematical expectation (mean) function. This is a short-time Fourier frequency domain transform; The expression for the species voiceprint feature matrix is ​​as follows:

[0034]

[0035]

[0036]

[0037]

[0038] in, This is a species voiceprint feature matrix. For the first The duration of the voiceprint signal of a species For the first The interval period of voiceprint signals for each species For the first The dominant frequency of the voiceprint signal of a species For the first The bandwidth of the voiceprint signal of a species This serves as a dimensional identifier for the species' voiceprint feature matrix. For the first The timestamp of the occurrence and end of the voiceprint signal of a species For the first The timestamp of the start of the voiceprint signal for each species. For the first The voiceprint signal of the species The timestamp of the first occurrence For the first The voiceprint signal of the species The timestamp of the first occurrence To retrieve the function The frequency at which the maximum value is obtained. For Fourier transform, For the first The upper frequency limit of the spectral range of the voiceprint signal of a species. For the first The lower frequency limit of the spectral density of the voiceprint signal of each species.

[0039] Furthermore: the soundscape semantic navigation decision module includes: The soundscape semantic graph construction unit is used to fuse UAV location data with the species voiceprint feature matrix, and construct a soundscape semantic graph by calculating the acoustic diversity index, species sound density and hotspot annotation; The RL-MPC navigation decision unit is used to construct the state space with the soundscape semantic map, the remaining battery power of the UAV and the distance to the hotspot, and to construct the action space with the UAV speed, heading and dwell time adjustment. It performs rolling optimization based on reinforcement learning and model predictive control algorithms, and outputs UAV flight path instructions by solving the future action sequence that maximizes the reward function. The abnormal sound scene response unit is used to calculate the cosine similarity between real-time voiceprint features and normal voiceprint feature templates. When the similarity is determined to be abnormal, abnormal sound scene data is obtained, and sound source localization is performed based on the time delay difference of the dual microphone array to obtain the coordinates of the abnormal sound source.

[0040] The further beneficial effects mentioned above are as follows: This invention spatializes acoustic features by constructing a soundscape semantic graph, thus solving the problem of the separation between navigation systems and ecological perception information; it adopts the RL-MPC algorithm to drive dynamic flight path optimization based on soundscape hotspots, achieving coordinated optimization of monitoring efficiency and energy consumption; through abnormal voiceprint recognition and sound source localization, the UAV can proactively respond to emergencies, enabling the UAV to change from a preset flight path to an intelligent flight path decision-making system with environmental understanding and autonomous decision-making capabilities.

[0041] Furthermore, the expression for the state space is as follows:

[0042] in, For state space, As a hot topic, The remaining battery power of the drone. Distance to the hotspot; The expression for the action space is as follows:

[0043] in, For the action space, For drone speed, For the course, Duration of stay; The expression for the reward function is as follows:

[0044] in, For the reward function, , and All are weights. For drones to consume power, This represents the total battery power of the drone.

[0045] Furthermore: the edge-cloud collaborative soundscape self-learning module includes: The edge-side real-time decision unit, deployed on a drone, is used to run a lightweight voiceprint recognition model to identify the voiceprint feature matrix of a species and to filter out uncertain samples based on a preset confidence threshold. The cloud-based soundscape data aggregation unit, deployed on a remote server, is used to receive uncertain samples uploaded from at least one drone, obtain multi-source data, and fuse and label the multi-source data to obtain an updated training dataset. The soundscape model iteration unit, deployed on a remote server, is used to optimize and train the lightweight voiceprint recognition model using the momentum gradient descent algorithm based on the updated training dataset, and then send the optimized lightweight voiceprint recognition model parameters to the drone to obtain the optimized lightweight voiceprint recognition model.

[0046] The further beneficial effects mentioned above are as follows: by filtering uncertain samples in real time at the edge and uploading them to the cloud, the problem between massive data transmission and local real-time processing is solved; by using the cloud server to fuse multi-source data and correct labels, and by iteratively optimizing the model using the momentum gradient descent algorithm and issuing updates, the UAV has the ability to continuously adapt to the dynamic changes in the ecological environment, reducing manual maintenance costs and ensuring the accuracy of long-term monitoring.

[0047] The beneficial effects of this invention are as follows: This invention dynamically constructs a noise field by fusing multimodal flight state data and combines it with a deep learning network for noise reduction, achieving precise suppression of dynamic flight noise from UAVs. This suppresses more noise, increases the proportion of effective soundscape signals, and ensures the purity and reliability of the data source. This invention employs an improved separation network based on dynamic sparsity constraints for soundscape analysis, which can adapt to changes in the number of species in the wild, achieving high-precision separation of mixed voiceprints from multiple species and providing a precise, structured species voiceprint feature matrix. This invention utilizes a soundscape semantic-driven UAV navigation mechanism, constructing a soundscape semantic map and using the RL-MPC algorithm for real-time path planning. This enables UAVs to actively track and focus on ecological hotspots, overcoming the limitations of traditional UAV navigation in harsh environments and improving the efficiency of acoustic data collection in ecological hotspots. This invention designs an edge-cloud collaborative self-learning architecture. By filtering uncertain samples on the UAV end and aggregating data and iterating the model on the cloud server, continuous optimization of the voiceprint recognition model is achieved. This allows it to adapt to dynamic changes in the ecological environment and reduces the long-term operation and maintenance costs of UAV ecological monitoring. This invention calculates ecological monitoring indicators and generates structured reports for targeted delivery, transforming raw data into standardized information products that can directly support management decisions. This achieves full automation of the ecological monitoring process and improves the timeliness and ease of use of ecological monitoring results. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the structure of an intelligent ecological perception and monitoring system for drones that combines soundscape cognition and multi-source noise reduction. Detailed Implementation

[0049] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0050] Example 1 like Figure 1 The diagram shown is a structural schematic of a soundscape recognition and multi-source noise reduction drone ecological intelligent perception and monitoring system. This invention provides a soundscape recognition and multi-source noise reduction drone ecological intelligent perception and monitoring system, comprising: The multi-source dynamic noise reduction and sound scene acquisition module is used to acquire the drone flight state feature vector, construct the drone noise field based on the flight state feature vector, and use a dual-microphone array that combines the drone noise field and the U-Net-TemporalConformer deep learning network to acquire and reduce the sound scene signal to obtain the initial ecological sound scene signal. The sparse evolutionary soundscape separation module is used to perform frequency domain transformation on the initial ecological soundscape signal and to perform multi-source separation under dynamic sparsity constraints through an improved convolutional temporal audio separation network to obtain the species voiceprint feature matrix. The soundscape semantic navigation decision module is used to fuse UAV location data with species voiceprint feature matrix to obtain soundscape semantic map. Based on reinforcement learning and model prediction control algorithm, it processes the data and outputs UAV flight path instructions. Based on the UAV flight path instructions, it obtains abnormal soundscape data and abnormal sound source coordinates through anomaly detection and sound source localization calculation.

[0051] The edge-cloud collaborative soundscape self-learning module is used to process the species voiceprint feature matrix on the UAV side through a lightweight voiceprint recognition model, output uncertain voiceprint samples, transmit the uncertain voiceprint samples to the cloud for multi-machine data fusion, and optimize the lightweight voiceprint recognition model based on the momentum gradient descent algorithm to obtain the optimized lightweight voiceprint recognition model. The ecological monitoring results output module is used to calculate ecological monitoring indicators based on species voiceprint feature matrix, soundscape semantic map and UAV flight path instructions, and push the generated structured monitoring report to different users.

[0052] For existing drone ecosystem monitoring, drones generate dynamically changing electromechanical and wind noise during flight, making it even more difficult to extract and separate the already weak ecological soundscape. Existing noise reduction methods are difficult to adapt to the real-time changes in drone flight status, resulting in acoustic data distortion. Furthermore, in the wild environment, the voiceprints of multiple species overlap, and traditional voiceprint separation algorithms have not been optimized for this, making it difficult to accurately extract the characteristics of a single species and limiting the depth of soundscape cognition. In addition, conventional drone flights rely on preset routes or visual navigation methods, and cannot actively perceive and focus on soundscape hotspots with high biodiversity, i.e., active areas with high biodiversity.

[0053] Therefore, this invention proposes a soundscape cognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system. By constructing a UAV noise field and combining a dual-microphone array with a U-Net-TemporalConformer deep learning network for adaptive deep noise reduction, the signal-to-noise ratio and quality of the original soundscape signal are improved. A convolutional temporal audio separation network based on dynamic sparsity constraints is adopted to achieve high-precision separation and feature extraction of voiceprints of different species in mixed soundscapes. The extracted species voiceprint features are fused with the UAV coordinates to construct a soundscape semantic map. Reinforcement learning and model predictive control are used to achieve UAV autonomous navigation and ecological hotspot tracking driven by soundscape semantics, transforming passive soundscape acquisition into active cruise perception. Finally, through an edge-cloud collaborative self-learning architecture, real-time identification and uncertain sample screening are performed on the UAV side, while multi-UAV data aggregation and model iterative optimization are performed on the cloud server, enabling the system to continuously evolve and adapt to environmental changes.

[0054] In one embodiment of the present invention, in existing technical solutions, when a drone performs ecological monitoring tasks, the electromechanical noise generated by its rotor and motor, as well as the aerodynamic noise during high-speed flight, have a wide frequency spectrum and high energy, overlapping with target ecological soundscape signals such as bird calls and insect chirps in the frequency domain; moreover, the noise spectrum changes in real time and in complex ways with the drone's flight state, such as acceleration, climb rate, and turn rate. Traditional static noise reduction filters or general noise reduction algorithms use fixed parameters, which are completely unable to adapt to this dynamically changing noise scene, resulting in insufficient noise reduction making it difficult to separate the target ecological soundscape signal, or excessive noise reduction causing the required target ecological soundscape signal to be lost. Therefore, the present invention designs a multi-source dynamic noise reduction and soundscape acquisition module, including: a multi-modal flight state perception unit, a multi-source noise field modeling unit, a dual-microphone array noise reduction unit, and an initial soundscape acquisition unit.

[0055] In a specific embodiment of the present invention, the multimodal flight state perception unit is used to collect the flight attitude parameters of the UAV through an IMU sensor. , For roll angle, The pitch angle, Yaw angle For the roll rate, For pitch rate, The yaw rate is measured in rad and rad / s, respectively, and the UAV's dynamic parameters are collected via a propeller speed sensor. , For the first The propeller rotation speed is expressed in r / min, and T is the transpose. Airflow disturbance parameters of the UAV are collected by an airflow sensor. , Wind speed in the horizontal direction. Wind speed along the vertical axis. The wind speed is in the vertical direction. The airflow pressure is expressed in m / s and Pa, respectively. The UAV flight attitude parameters, UAV dynamic parameters, and UAV airflow disturbance parameters are processed by time synchronization, normalization, and weighted calculation to obtain the UAV flight state feature vector. This invention utilizes a multimodal flight state perception unit to synchronously and precisely acquire flight attitude, dynamic parameters, and airflow disturbance data, enabling real-time perception of the UAV's dynamic flight state and providing accurate input for noise prediction. Specifically, the IMU sensor, propeller speed sensor, and airflow sensor are all synchronized via GPS, ensuring a time synchronization error ≤10ms. Min-max normalization is employed to eliminate dimensional differences and normalize the data. Finally, weights are assigned based on sensor priority and weighted calculations are performed to obtain the UAV flight state feature vector. The expression for the UAV flight state feature vector is as follows:

[0056]

[0057] in, This is the feature vector of the UAV's flight state. For time, Assigning weights to IMU sensors, Assigning weights to the propeller speed sensor, Assigning weights to the airflow sensors, These are the normalized flight attitude parameters of the UAV. The normalized UAV dynamic parameters, These are the normalized airflow disturbance parameters for the UAV. This is the calibration minimum value for the IMU sensor. For the calibration maximum value of the IMU sensor, This is the calibration minimum value for the propeller speed sensor. This is the maximum calibration value for the propeller speed sensor. This is the calibration minimum value for the airflow sensor. This is the maximum calibration value for the airflow sensor.

[0058] In a specific embodiment of the present invention, the noise generated by a UAV during flight is not a stationary random signal, but rather a combination of physical parameters coupled with its flight attitude, power system operating conditions, and external airflow environment. Current noise identification methods mostly estimate noise spectra with fixed parameters based on historical average data, making it difficult to capture and describe the dynamic characteristics of noise evolution in real time with flight status. This results in subsequent noise reduction processing being unable to suppress the dynamically changing noise during actual UAV flight. Therefore, the present invention designs a multi-source noise field modeling unit. This unit is used to obtain the dynamic noise field of the UAV during flight status by modeling it through frequency domain mapping based on the UAV flight status feature vector and introducing dynamic correction based on the flight status change rate. The multi-source noise field modeling unit decomposes noise sources, including UAV noise. Disassembly into electromechanical noise With wind noise ,get ; Then, frequency domain mapping modeling is performed, and the mapping relationship between the UAV flight state and the noise power spectral density (PSD) is established by fitting training samples. ,in, This represents the mapping relationship between the flight state of the UAV and the noise power spectral density (PSD). To train the weight matrix, , For frequency points, It is the ReLU activation function. For bias terms, This represents the feature vector of the UAV's flight state. Introducing the rate of change of flight state Dynamic corrections are performed to adapt to the dynamic flight scenarios of the drone, resulting in the dynamic noise field during drone flight, expressed as follows:

[0059] in, For dynamic noise field, For frequency, For time, This represents the mapping relationship between the flight state of the UAV and the noise power spectral density (PSD). To adjust the coefficient, For the rate of change of flight state, The L2 norm is used. This invention transforms dynamically changing flight noise from a noise interference source into an internal variable for modeling through a multi-source noise field modeling unit. This changes the noise reduction process from separating noise from mixed signals to targeted suppression under known noise characteristics, thus improving the targeting and effectiveness of noise reduction.

[0060] In a specific embodiment of the present invention, a dual-microphone array noise reduction unit is designed. A dual-microphone array with a spacing of 0.1 m is used to acquire ecological soundscape signals, with a sensitivity ≥40dB@1V / Pa. The two microphones synchronously acquire the original soundscape signal, obtaining the first signal from the original soundscape signal. and the second signal in the original soundscape signal , For the target soundscape signal, and All of these are noise signals; the original soundscape signal is then filtered to obtain a pre-filtered signal, the expression of which is as follows:

[0061]

[0062]

[0063] in, For pre-filtered signals, It is a phase weighting vector. It is the conjugate transpose. For time delay difference, For the speed of sound, The angle of incidence of the sound source. The spacing between the two microphone arrays; Based on the pre-filtered signal and the dynamic noise field, the Recursive Least Squares (RLS) algorithm is used for updating to obtain the optimized dynamic noise field, the expression of which is as follows:

[0064] in, For optimized dynamic noise field, For frequency, For time, The smoothing coefficient can be set to 0.98. This is the threshold coefficient, which can be set to 1.2. For indicator functions, This is a dynamic noise field; The optimized dynamic noise field and pre-filtered signal are input into the U-Net-TemporalConformer deep learning network for deep learning noise reduction, outputting a clean sound field spectrum, in which the pre-filtered signal is converted into Mel-spectral features. With 80 Mel filter groups and a time frame length of 100, the expression for the pure soundscape spectrum is as follows:

[0065] in, The output of the U-Net-TemporalConformer deep learning network For the U-Net-TemporalConformer deep learning network, For Mel spectrum conversion, The signal is pre-filtered; the clean soundscape spectrum output by the U-Net-TemporalConformer deep learning network is subjected to inverse STFT transform to obtain the time-domain denoised soundscape signal. Its noise suppression ratio Signal-to-noise ratio .

[0066] In a specific embodiment of the present invention, the U-Net-TemporalConformer deep learning network includes: a downsampling feature extraction module consisting of 3×3 convolutional kernels and 2×2 max pooling, and an upsampling recovery module consisting of 4×4 deconvolution and skip connections; its constrained loss function is expressed as follows:

[0067]

[0068]

[0069] in, The output of the U-Net-TemporalConformer deep learning network For a pure soundscape spectrum. The separated noise spectrum. .

[0070] The initial soundscape acquisition unit is used to quantize and encode the time-domain noise reduction soundscape signal and output the initial ecological soundscape signal; it uses 16-bit quantization encoding, a sampling rate of 44.1kHz, and a storage format of WAV.

[0071] In one embodiment of the present invention, the real ecological soundscape in the wild is a dynamic mixture of calls from various organisms such as birds, insects, and amphibians occurring simultaneously and overlapping each other, and the number of active species changes in real time with time and space. In the prior art, whether it is a traditional clustering method or a more advanced general audio separation network, its separation performance will decrease when faced with a number of dynamically changing species not covered by the training set, and it is easy to produce erroneous soundscape separations such as merging multiple species voiceprints into one or splitting a single species voiceprint into multiple parts. Therefore, the present invention designs a sparsity evolutionary soundscape separation module, which includes: a soundscape frequency domain feature conversion unit, a sparsity evolutionary separation unit, and a soundscape feature extraction unit.

[0072] In a specific embodiment of the present invention, to address the problem that existing technologies lack effective analysis and characterization of the essential time-frequency structure of acoustic signals in the initial stage of soundscape separation, resulting in the inability to perform refined analysis and separation of aliased ecological sound signatures, the present invention designs a soundscape frequency domain feature transformation unit to perform a short-time Fourier frequency domain transform on the initial ecological soundscape signal, obtaining a frequency domain complex matrix, the expression of which is as follows:

[0073] in, It is a complex matrix in the frequency domain. Let f be the amplitude of the frequency domain complex matrix spectrum at time t′. Peak threshold This is the arithmetic mean operation function. For short-time Fourier frequency domain transform, This serves as the initial ecological soundscape signal; In a specific embodiment of the present invention, the number of active species in the wild ecological soundscape is not fixed, but fluctuates in real time with day and night, season and environment. Existing mainstream blind source separation algorithms usually need to preset or assume a fixed number of sound sources during the training phase. When the number of species in the actual scene does not match the assumption, the output result will produce serious errors: if the actual number of species is more than the preset number of species, the voiceprints of multiple species will be incorrectly merged into the same output, causing feature contamination and information loss; if the actual number of species is less than the preset number of species, the output result will produce false separation output or decompose a complete voiceprint into multiple ones. To address this, the present invention designs a sparsity evolution separation unit, which is used to obtain the number of species from the frequency domain complex matrix through the spectral peak detection algorithm, obtain a sparsity threshold based on the number of species, combine the sparsity threshold with the convolutional temporal audio separation network to obtain an improved convolutional temporal audio separation network, and perform source separation on the initial ecological soundscape signal to obtain a single species voiceprint signal including a single species voiceprint signal; The sparsity evolution separation unit first performs real-time analysis of the time-frequency matrix using a spectral peak detection algorithm to estimate the number of species at the current moment, as expressed below:

[0074] in, For the number of species, It is a complex matrix in the frequency domain. Let f be the amplitude of the frequency domain complex matrix spectrum at time t′. Peak threshold This is the arithmetic mean operation function. For time traversal variables, For indicator functions; Based on the number of species, dynamic sparsity is calculated to obtain the sparsity threshold, which is expressed as follows:

[0075] in, The sparsity threshold, Based on sparsity, It is a natural exponential function. This is the adjustment coefficient; By combining a sparsity threshold with a convolutional temporal audio separation network, an improved convolutional temporal audio separation network is obtained, comprising an encoding layer, a dynamic sparse layer, a separation layer, and a decoding layer. The encoding layer maps the initial input ecological soundscape signal into an input feature representation through a one-dimensional convolution with a kernel length of 2 and a stride of 1, as shown in the following expression:

[0076] in, For input feature representation, It is the ReLU activation function. For 1D convolution, This serves as the initial ecological soundscape signal; The expression for the dynamic sparse layer is as follows:

[0077] in, For a dynamic sparse layer based on a sparsity threshold, For indicator functions; The separation layer consists of eight sequentially connected dilated convolutional layers, PReLU activation functions, and residual layers, capable of generating the first... Mask matrix of sound sources ; The expression for the decoding layer is as follows:

[0078]

[0079] in, For a single species' voiceprint signal, it indicates Time of the first Voiceprint signals of each species For the first Species, For time, For transposed convolution, This is the frequency domain species voiceprint signal. For a dynamic sparse layer based on a sparsity threshold, For the first Mask matrix for each species; In a specific embodiment of the present invention, the expression for the loss function of the improved convolutional temporal audio separation network is as follows:

[0080] in, To improve the loss function of convolutional temporal audio separation networks, For the first Species, For the number of species, These are voiceprint signals from real species. To improve the single-species voiceprint signal output by the convolutional temporal audio separation network, this invention's convolutional temporal audio separation network combines a sparsity threshold to form a dynamic sparsity mechanism, enabling the network to adaptively sparsify according to the number of sound source species: when the number of species is large, the dynamic sparsity layer applies stronger sparsity constraints, allowing the network to learn more independent and mutually exclusive sound source representations to distinguish different species; when the number of species is small, the dynamic sparsity layer reduces sparsity constraints, avoiding unnecessary decomposition of simple scenes and ensuring the integrity of the voiceprint signal.

[0081] In a specific embodiment of the present invention, after obtaining the voiceprint signal of a single species, feature parsing and structured characterization are required to facilitate further in-depth analysis of the species' voiceprints. The present invention employs a soundscape feature extraction unit to extract multi-dimensional features from the voiceprint signal of a single species, obtaining a voiceprint feature matrix for each species. Specifically, regarding the scientific significance of ecological monitoring, the present invention extracts quantitative indicators with clear ecological interpretive significance in both the time and frequency domains from the voiceprint signal of each isolated single species, including the duration reflecting the length of the call. Reflecting the interval period of the chirping rhythm The dominant frequency, which reflects the main pitch of a sound and the bandwidth reflecting the frequency variation range of sound Ultimately, these elements together form an information-rich species voiceprint feature matrix, the expression of which is as follows:

[0082]

[0083]

[0084]

[0085]

[0086] in, This is a species voiceprint feature matrix. For the first The duration of the voiceprint signal of a species For the first The interval period of voiceprint signals for each species For the first The dominant frequency of the voiceprint signal of a species For the first The bandwidth of the voiceprint signal of a species This serves as the dimension identifier for the species' voiceprint feature matrix, indicating that the matrix is ​​a two-dimensional matrix with K rows and 4 columns. For the first The timestamp of the occurrence and end of the voiceprint signal of a species For the first The timestamp of the start of the voiceprint signal for each species. For the first The voiceprint signal of the species The timestamp of the first occurrence For the first The voiceprint signal of the species The timestamp of the first occurrence To retrieve the function The frequency at which the maximum value is obtained. For Fourier transform, For the first The upper frequency limit of the spectral range of the voiceprint signal of a species. For the first The lower frequency limit of the spectral density of the voiceprint signal of each species. The species voiceprint feature matrix of this invention can be used to subsequently construct a soundscape semantic map, mapping sound features into spatialized ecological semantics; at the same time, the species voiceprint feature matrix provides a unified format of feature vectors for voiceprints of different species, which facilitates standardized machine learning model training and model iteration.

[0087] In one embodiment of the present invention, existing drone ecological monitoring navigation mostly relies on pre-set geometric waypoints or visual image positioning and satellite positioning. Therefore, drones mostly fly along fixed routes and may stay in ecologically inactive areas for too long, lacking sampling of soundscape data such as species gathering areas, resulting in low ecological monitoring efficiency. At the same time, in forest canopy or complex ecological environments, visual and satellite navigation often fail, easily leading to interruption of ecological monitoring tasks. Therefore, the present invention designs a soundscape semantic navigation decision module to fuse drone position data with species voiceprint feature matrix to obtain a soundscape semantic map. Based on reinforcement learning and model predictive control algorithms, the module processes the data and outputs drone flight path instructions. Based on the drone flight path instructions, abnormal soundscape data and abnormal sound source coordinates are obtained through anomaly detection and sound source localization calculation. It can adjust the flight path in real time through species voiceprint features to generate a dynamically optimal flight path that balances maximizing ecological information gain, energy consumption constraints, and flight safety. Specifically, it includes: The soundscape semantic map construction unit is used to fuse UAV location data with the species voiceprint feature matrix, and construct a soundscape semantic map by calculating the acoustic diversity index, species sound density, and hotspot annotations; wherein, UAV location data can be obtained using satellite positioning. , Latitude Longitude For height; The expression for the acoustic diversity index is as follows:

[0088] in, As an acoustic diversity index, The effective frequency range is set to 4000Hz. For frequency Number of species in the area Total number of species; The expression for species acoustic density is as follows:

[0089] in, For species sound density, The statistics window can be set to 60 seconds. Hotspot labeling is based on acoustic diversity index and species sound density, and a threshold for acoustic diversity index can be set. and species sound density threshold If the acoustic diversity index and species sound density at the current location of the drone are both greater than the acoustic diversity index threshold... and species sound density threshold The current location will then be marked as a hotspot. ; Finally, a soundscape semantic map was constructed. , For the number of regions, For the first Each region.

[0090] The RL-MPC navigation decision unit uses a soundscape semantic graph, the remaining battery power of the UAV, and the distance to hotspots to form the state space, and the UAV speed, heading, and dwell time adjustment amounts to form the action space. It performs rolling optimization based on reinforcement learning and model predictive control algorithms, and outputs the UAV flight path command by solving for the future action sequence that maximizes the reward function. The expression for the state space is as follows:

[0091] in, For state space, As a hot topic, The remaining battery power of the drone. Distance to the hotspot; The expression for the action space is as follows:

[0092] in, For the action space, For drone speed, For the course, Duration of stay; The expression for the reward function is as follows:

[0093] in, For the reward function, , and All are weights, which can be taken as... , For drones to consume power, The total battery level of the drone; through the reward function, the drone can not only intelligently track sound and scene hotspots, but also autonomously manage its energy to avoid mission failure or equipment damage due to running out of power; Rolling optimization based on reinforcement learning and model predictive control algorithms specifically includes: using future... To predict the time domain, solve the constrained optimization problem. The constraints are , , The drone flight path instructions are: In non-hotspot areas, the drones fly at maximum speed.

[0094] The abnormal sound scene response unit calculates the cosine similarity between real-time voiceprint features and normal voiceprint feature templates. When the similarity is determined to be abnormal, abnormal sound scene data is obtained, and sound source localization is performed based on the time delay difference of the dual-microphone array to obtain the coordinates of the abnormal sound source. Specifically, if the cosine similarity is less than a set similarity threshold, it is determined to be abnormal, and this can be determined by the time delay difference of the dual-microphone array. Calculate azimuth By combining GPS coordinates to obtain the coordinates of the abnormal sound source, the drone can be controlled to fly to the abnormal sound source to collect the acoustic signature signal.

[0095] In one embodiment of the present invention, existing technologies often use pre-trained voiceprint recognition models with fixed parameters for ecological monitoring. Their generalization ability is limited, and the recognition accuracy decreases when faced with soundscape features from new regions, new species, or new seasons not covered by the training data. To address this, the present invention designs a self-learning function: an edge-cloud collaborative soundscape self-learning module. This module processes the species voiceprint feature matrix on the UAV side using a lightweight voiceprint recognition model, outputting uncertain voiceprint samples. These uncertain voiceprint samples are then transmitted to the cloud for multi-machine data fusion. The lightweight voiceprint recognition model is then optimized using a momentum gradient descent algorithm to obtain an optimized lightweight voiceprint recognition model. The lightweight voiceprint recognition model can employ existing voiceprint recognition models such as PyTorch pre-trained models, ResNet-based models, or X-Vector models.

[0096] The edge-cloud collaborative soundscape self-learning module includes: The edge-side real-time decision-making unit, deployed on a drone, is used to run a lightweight voiceprint recognition model to identify the voiceprint feature matrix of a species and to filter out uncertain samples based on a preset confidence threshold. The confidence threshold can be set to 0.8. When the confidence of the voiceprint recognition result is less than the confidence threshold, it is filtered as an uncertain sample. The cloud-based soundscape data aggregation unit, deployed on a remote server, receives uncertain samples uploaded from at least one drone to obtain multi-source data. It then fuses and corrects the labels of the multi-source data to obtain an updated training dataset. Specifically, a weighted average method is used to aggregate uncertain samples uploaded by multiple drones to obtain multi-source data. The soundscape model iteration unit, deployed on a remote server, is used to optimize and train the lightweight voiceprint recognition model using the momentum gradient descent algorithm based on the updated training dataset. The optimized lightweight voiceprint recognition model parameters are then sent to the drone to obtain the optimized lightweight voiceprint recognition model. This enables the lightweight voiceprint recognition model deployed on different drones to continuously and automatically adapt to soundscape changes in different environments and at different times, maintaining high accuracy in voiceprint recognition.

[0097] In one embodiment of the present invention, to ensure the real-time nature of ecological monitoring and the timeliness of decision-making, the present invention designs an ecological monitoring result output module, which integrates data such as voiceprint feature matrix, semantic graph, and flight route instructions into ecological monitoring indicators, and constructs a structured report in a unified format for easy viewing and analysis by users; specifically including: Calculate ecological monitoring indicators and generate structured reports, including species richness. , Total number of species, hotspot coverage This refers to the ratio of hotspot areas to the total number of areas, representing the incidence rate of abnormal events. This indicates the number of abnormal events per hour. The structured report includes integrated indicator data, voiceprint samples, GPS coordinates, and hotspot distribution maps, and generates a JSON format report. The report is compressed using the LZ77 algorithm and pushed to the APP via 5G or Beidou communication. The full version of the report and visualization charts are pushed to the Web terminal, and the encrypted raw data is pushed to the management terminal to support historical data retrieval.

[0098] The beneficial effects of this invention are as follows: This invention dynamically constructs a noise field by fusing multimodal flight state data and combines it with a deep learning network for noise reduction, achieving precise suppression of dynamic flight noise from UAVs. This suppresses more noise, increases the proportion of effective soundscape signals, and ensures the purity and reliability of the data source. This invention employs an improved separation network based on dynamic sparsity constraints for soundscape analysis, which can adapt to changes in the number of species in the wild, achieving high-precision separation of mixed voiceprints from multiple species and providing a precise, structured species voiceprint feature matrix. This invention utilizes a soundscape semantic-driven UAV navigation mechanism. By constructing a soundscape semantic map and using the RL-MPC algorithm for real-time path planning, this invention enables UAVs to actively track and focus on ecological hotspots, overcoming the limitations of traditional UAV navigation in harsh environments and improving the efficiency of acoustic data collection in ecological hotspots. This invention designs an edge-cloud collaborative self-learning architecture. By filtering uncertain samples on the UAV end and aggregating data and iterating the model on the cloud server, this invention achieves continuous optimization of the voiceprint recognition model, adapting to dynamic changes in the ecological environment and reducing the long-term operation and maintenance costs of UAV ecological monitoring. This invention calculates ecological monitoring indicators and generates structured reports for targeted delivery, transforming raw data into standardized information products that can directly support management decisions. This achieves full automation of the ecological monitoring process and improves the timeliness and ease of use of ecological monitoring results.

Claims

1. An acoustic scene cognition and multi-source noise reduction unmanned aerial vehicle ecological intelligent perception monitoring system, characterized in that, include: The multi-source dynamic noise reduction and sound scene acquisition module is used to acquire the drone flight state feature vector, construct the drone noise field based on the flight state feature vector, and use a dual-microphone array that combines the drone noise field and the U-Net-TemporalConformer deep learning network to acquire and reduce the sound scene signal to obtain the initial ecological sound scene signal. The sparse evolutionary soundscape separation module is used to perform frequency domain transformation on the initial ecological soundscape signal and to perform multi-source separation under dynamic sparsity constraints through an improved convolutional temporal audio separation network to obtain the species voiceprint feature matrix. The soundscape semantic navigation decision module is used to fuse UAV location data with species voiceprint feature matrix to obtain soundscape semantic map, process it based on reinforcement learning and model prediction control algorithm, output UAV flight path command, and obtain abnormal soundscape data and abnormal sound source coordinates based on UAV flight path command through anomaly detection and sound source localization calculation. The edge-cloud collaborative soundscape self-learning module is used to process the species voiceprint feature matrix on the UAV side through a lightweight voiceprint recognition model, output uncertain voiceprint samples, transmit the uncertain voiceprint samples to the cloud for multi-machine data fusion, and optimize the lightweight voiceprint recognition model based on the momentum gradient descent algorithm to obtain the optimized lightweight voiceprint recognition model. The ecological monitoring results output module is used to calculate ecological monitoring indicators based on species voiceprint feature matrix, soundscape semantic map and UAV flight path instructions, and push the generated structured monitoring report to different user terminals.

2. The soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system according to claim 1, characterized in that, The multi-source dynamic noise reduction and sound scene acquisition module includes: The multimodal flight state perception unit is used to collect UAV flight attitude parameters through IMU sensors, UAV dynamic parameters through propeller speed sensors, and UAV airflow disturbance parameters through airflow sensors. It then processes the UAV flight attitude parameters, UAV dynamic parameters, and UAV airflow disturbance parameters through time synchronization, normalization, and weighted calculation to obtain the UAV flight state feature vector. The multi-source noise field modeling unit is used to obtain the dynamic noise field of the UAV in flight state by modeling through frequency domain mapping and introducing dynamic correction of the flight state change rate based on the UAV flight state feature vector; The dual-microphone array noise reduction unit is used to acquire the original sound scene signal through a dual-microphone array, and combine it with the dynamic noise field and U-Net-TemporalConformer deep learning network for noise reduction and temporal reconstruction to obtain the temporal noise reduction scene signal. The initial soundscape acquisition unit is used to quantize and encode the time-domain noise reduction soundscape signal and output the initial ecological soundscape signal.

3. The soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system according to claim 2, characterized in that, The expression for the UAV flight state feature vector is as follows: in, This is the feature vector of the UAV's flight state. For time, Assigning weights to IMU sensors, Assigning weights to the propeller speed sensor, Assigning weights to the airflow sensors, These are the normalized flight attitude parameters of the UAV. The normalized UAV dynamic parameters, These are the normalized airflow disturbance parameters for the UAV. This is the calibration minimum value for the IMU sensor. For the calibration maximum value of the IMU sensor, This is the calibration minimum value for the propeller speed sensor. This is the maximum calibration value for the propeller speed sensor. This is the calibration minimum value for the airflow sensor. This is the calibration maximum value for the airflow sensor. For the flight attitude parameters of the UAV, For roll angle, The pitch angle, Yaw angle For the roll rate, For pitch rate, The yaw rate is... For the power parameters of the drone, For the first propeller speed, These are the airflow disturbance parameters for the UAV. Wind speed in the horizontal direction. Wind speed along the vertical axis. The wind speed is in the vertical direction. Let T be the air pressure and T be the transpose.

4. The soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system according to claim 2, characterized in that, The expression for the dynamic noise field during the flight of the UAV is as follows: in, For dynamic noise field, For frequency, For time, This represents the mapping relationship between the flight state of the UAV and the noise power spectral density (PSD). To adjust the coefficient, For the rate of change of flight state, It is the L2 norm. It is the ReLU activation function. To train the weight matrix, For bias terms, For the feature dimension index of the noise power spectral density mapping model, This is the feature vector of the UAV's flight state. For drone noise, For electromechanical noise, It is wind noise.

5. The soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system according to claim 2, characterized in that, The expression for the time-domain noise-reduced scene signal is as follows: in, For time-domain noise reduction of scene signals, This is the inverse short-time Fourier transform. The output of the U-Net-TemporalConformer deep learning network For the U-Net-TemporalConformer deep learning network, For Mel spectrum conversion, For pre-filtered signals, For optimized dynamic noise field, For frequency, For time, For smoothing coefficients, For threshold coefficient, For indicator functions, For dynamic noise field, It is a phase weighting vector. It is the conjugate transpose. The first signal after the short-time Fourier transform. The second signal is the result of the short-time Fourier transform. This is the first signal in the original soundscape signal. This is the second signal in the original soundscape signal. For the target soundscape signal, and All of these are noise signals.

6. The soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system according to claim 1, characterized in that, The sparsity evolution soundscape separation module includes: The soundscape frequency domain feature transformation unit is used to perform short-time Fourier frequency domain transformation on the initial ecological soundscape signal to obtain a frequency domain complex matrix; The sparsity evolution separation unit is used to obtain the number of species from the frequency domain complex matrix through the spectral peak detection algorithm, obtain the sparsity threshold based on the number of species, combine the sparsity threshold with the convolutional temporal audio separation network to obtain an improved convolutional temporal audio separation network, and perform source separation on the initial ecological sound scene signal to obtain the single species voiceprint signal. The soundscape feature extraction unit is used to extract multi-dimensional features from the voiceprint signal of a single species to obtain the voiceprint feature matrix for each species.

7. The soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system according to claim 6, characterized in that, The expression for the voiceprint signal of a single species is as follows: in, For a single species' voiceprint signal, it indicates Time of the first Voiceprint signals of each species For the first Species, For time, For transposed convolution, This is the frequency domain species voiceprint signal. For a dynamic sparse layer based on a sparsity threshold, For the first Mask matrix for each species, For input feature representation, The sparsity threshold, For indicator functions, It is the ReLU activation function. For 1D convolution, This is the initial ecological soundscape signal. Based on sparsity, It is a natural exponential function. For adjustment coefficients, For the number of species, It is a complex matrix in the frequency domain. Let f be the amplitude of the frequency domain complex matrix spectrum at time t′. Peak threshold Let be the mathematical expectation operation function. This is a short-time Fourier frequency domain transform; The expression for the species voiceprint feature matrix is ​​as follows: in, This is a species voiceprint feature matrix. For the first The duration of the voiceprint signal of a species For the first The interval period of voiceprint signals for each species For the first The dominant frequency of the voiceprint signal of a species For the first The bandwidth of the voiceprint signal of a species This serves as a dimensional identifier for the species' voiceprint feature matrix. For the first The timestamp of the occurrence and end of the voiceprint signal of a species For the first The timestamp of the start of the voiceprint signal for each species. For the first The voiceprint signal of the species The timestamp of the first occurrence For the first The voiceprint signal of the species The timestamp of the first occurrence To retrieve the function The frequency at which the maximum value is obtained. For Fourier transform, For the first The upper frequency limit of the spectral range of the voiceprint signal of a species. For the first The lower frequency limit of the spectral density of the voiceprint signal of each species.

8. The soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system according to claim 1, characterized in that, The soundscape semantic navigation decision module includes: The soundscape semantic graph construction unit is used to fuse UAV location data with the species voiceprint feature matrix, and construct a soundscape semantic graph by calculating the acoustic diversity index, species sound density and hotspot annotation; The RL-MPC navigation decision unit is used to construct the state space with the soundscape semantic map, the remaining battery power of the UAV and the distance to the hotspot, and to construct the action space with the UAV speed, heading and dwell time adjustment. It performs rolling optimization based on reinforcement learning and model predictive control algorithms, and outputs UAV flight path instructions by solving the future action sequence that maximizes the reward function. The abnormal sound scene response unit is used to calculate the cosine similarity between real-time voiceprint features and normal voiceprint feature templates. When the similarity is determined to be abnormal, abnormal sound scene data is obtained, and sound source localization is performed based on the time delay difference of the dual microphone array to obtain the coordinates of the abnormal sound source.

9. The soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system according to claim 8, characterized in that, The expression for the state space is as follows: in, For state space, As a hot topic, The remaining battery power of the drone. Distance to the hotspot; The expression for the action space is as follows: in, For the action space, For drone speed, For the course, Duration of stay; The expression for the reward function is as follows: in, For the reward function, , and All are weights. For drones to consume power, This represents the total battery power of the drone.

10. The soundscape recognition and multi-source noise reduction UAV ecological intelligent perception and monitoring system according to claim 1, characterized in that, The edge-cloud collaborative soundscape self-learning module includes: The edge-side real-time decision unit, deployed on a drone, is used to run a lightweight voiceprint recognition model to identify the voiceprint feature matrix of a species and to filter out uncertain samples based on a preset confidence threshold. The cloud-based soundscape data aggregation unit, deployed on a remote server, is used to receive uncertain samples uploaded from at least one drone, obtain multi-source data, and fuse and label the multi-source data to obtain an updated training dataset. The soundscape model iteration unit, deployed on a remote server, is used to optimize and train the lightweight voiceprint recognition model using the momentum gradient descent algorithm based on the updated training dataset, and then send the optimized lightweight voiceprint recognition model parameters to the drone to obtain the optimized lightweight voiceprint recognition model.

Images