Voiceprint monitoring method and system integrating bird recognition and noise traceability

Abstract

The invention discloses a voiceprint monitoring method and system integrating bird recognition and noise traceability, and the method comprises the following steps: a position optimization deployment stage: collecting environment parameters of a destination, and carrying out the evaluation of a plurality of candidate positions through a scoring model based on the environment parameters, selecting the candidate position with the highest score as the final deployment position of the microphone array; in the voiceprint data acquisition stage, at the final deployment position, a solar-powered microphone array is used for acquiring mixed sound signals in the environment; a voiceprint analysis processing stage: carrying out parallel processing on the mixed sound signal, the parallel processing process comprising: a bird recognition processing sub-stage: extracting voiceprint features of bird buzzing from the mixed sound signal, and inputting the voiceprint features to a deep learning classification model integrated with a feature enhancement module, and outputting an identification result containing the bird species and the calibrated confidence. According to the invention, high-precision bird monitoring can be realized, and meanwhile, precise noise control can be synchronously carried out based on bird monitoring equipment.

Description

Technical Field

[0001] This invention relates to the field of sound monitoring technology, specifically to a voiceprint monitoring method and system that integrates bird identification and noise source tracing. Background Technology

[0002] Existing voiceprint monitoring technologies suffer from several key shortcomings in integrating bird identification and noise source tracing. These shortcomings are interconnected and collectively constrain overall monitoring performance. Specifically, they manifest in the following ways: First, in the basic deployment stage of signal acquisition, there is a lack of scientifically quantifiable solutions tailored to the dual needs of bird identification and noise source tracing. Microphone array site selection often relies on subjective experience, failing to comprehensively assess factors such as solar irradiance, obstacle obstruction rate, and environmental noise levels, and even neglecting key parameters such as geographic information and human activity frequency. Improper deployment directly leads to poor quality acquisition of the mixed sound signals shared by both types of monitoring, hindering subsequent identification and source tracing. Firstly, the lack of an anti-interference mechanism leads to several issues. Firstly, the absence of a mechanism to prevent interference from natural sources like wind, vegetation noise, and building reverberation creates a cumulative negative impact on both types of monitoring. Secondly, in bird identification, the absence of an attention mechanism that focuses on key frequency bands makes it difficult to extract clear Mel-frequency cepstral coefficient features from the interference, resulting in misidentification or missed identification. Thirdly, in noise source tracing, the lack of a phase transformation weighting function in the generalized cross-correlation time delay estimation method, and the absence of adaptive filtering to compensate for multipath effects, leads to signal distortion between microphone array channels. This makes it impossible to accurately calculate the time difference of noise sources, reducing both positioning accuracy and noise generation. The system suffers from several issues. First, the accuracy of sound type recognition is compromised. Second, the core algorithms are not optimized collaboratively. The raw probabilities output by the deep learning classification model for bird identification are not calibrated using logistic regression fitting on a historical data validation set, leading to either overconfidence or underconfidence, resulting in low reliability of the monitoring data. Third, the latency estimation algorithm for noise source tracing does not share the clean signal after anti-interference processing with the feature extraction stage of bird identification. Instead, the lack of effective noise filtering further interferes with the extraction of bird call features. The isolated algorithms not only fail to mutually enhance each other but also cause accuracy loss. Finally, the integrated system lacks intelligent closed-loop management. The results of bird identification and noise source tracing are not effectively integrated and packaged to provide unified and complete monitoring data to the remote management terminal. Furthermore, no adaptive response mechanism adapted to the bird habitat is designed for the traced human-caused noise. Fixed-volume warnings either fail to provide effective warnings due to insufficient volume or create secondary noise that damages the bird habitat due to excessive volume. This makes it impossible to achieve closed-loop management from data collection to analysis to intervention, failing to meet the dual needs of ecological protection and public noise control. These technical pain points in the integrated scenario are what prompted the proposal of this integrated voiceprint monitoring method and system for bird identification and noise source tracing. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a voiceprint monitoring method integrating bird identification and noise source tracing, comprising the following steps:

[0004] Location optimization and deployment phase: Collect environmental parameters of the target location, evaluate multiple candidate locations based on the environmental parameters using a scoring model, and select the candidate location with the highest score as the final deployment location of the microphone array;

[0005] Voiceprint data acquisition phase: At the final deployment location, a solar-powered microphone array is used to acquire mixed sound signals from the environment;

[0006] Voiceprint analysis and processing stage: The mixed audio signals are processed in parallel;

[0007] The bird identification processing sub-stage extracts the voiceprint features of bird calls from the mixed sound signals and inputs them into a deep learning classification model with an integrated feature enhancement module, outputting identification results that include the bird species and their calibrated confidence scores.

[0008] The noise source tracing sub-stage involves: utilizing the inter-channel signals of the microphone array, employing a sound source localization algorithm based on time delay estimation and supplemented with anti-interference processing, calculating and outputting the direction or location information of the noise source; and identifying the type of the noise signal; the anti-interference processing includes signal preprocessing and algorithm optimization.

[0009] Result output and response stage: The bird identification result and the direction or location information of the noise source are encapsulated into a monitoring result and sent to the remote management terminal; when the identified noise type is a preset man-made noise, the directional acoustic persuasion device is automatically triggered to perform adaptive response linkage control.

[0010] Furthermore, in the location optimization deployment phase, the environmental parameters used to evaluate candidate locations include solar irradiance, obstacle obstruction rate, and environmental noise level.

[0011] The scoring model is a weighted scoring model calculated by first normalizing solar irradiance, obstacle obstruction rate, and environmental noise level, and then combining these factors with weighting coefficients. The specific calculation formula is as follows:

[0012] ;

[0013] Where S is the location score. I norm To normalize solar irradiance, O norm To normalize the obstacle occlusion rate, N norm To normalize the environmental noise level, and These are the weighting coefficients, and .

[0014] Furthermore, in the bird recognition processing sub-stage, the feature enhancement module uses an attention mechanism; the attention mechanism achieves selective focusing on key features by calculating an attention weight matrix;

[0015] The extracted voiceprint features are Mel frequency cepstral coefficient features;

[0016] The deep learning classification model is a convolutional neural network with an integrated attention mechanism. The attention weight matrix of the attention mechanism is calculated as follows:

[0017] ;

[0018] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively. Let K be the dimension of the key vector, A be the attention weight matrix used to enhance the model's ability to focus on key frequency band features of bird calls, and T be the standard transpose notation in matrix operations, which, when applied to the key matrix K, yields the transposed key matrix. .

[0019] Furthermore, in the noise source tracing sub-stage, the sound source localization algorithm based on time delay estimation is specifically the generalized cross-correlation time delay estimation method;

[0020] Anti-interference processing is achieved by introducing a specific weighting function into the generalized cross-correlation function of the generalized cross-correlation time delay estimation method;

[0021] The expression for the generalized cross-correlation function is:

[0022] ;

[0023] in, This is the phase transformation weighting function used to achieve anti-interference processing, which is used to suppress environmental reverberation and noise interference; and Let be the Fourier transform of the received signals from the i-th and j-th microphones in the microphone array. The time difference of signal arrival. j It is the imaginary unit.

[0024] Furthermore, in the result output and response phase, the adaptive response of the linkage control specifically includes: dynamically adjusting the output volume of the persuasion device based on the monitored noise sound pressure level, wherein the calculation formula for the playback volume is:

[0025] ;

[0026] in, For output volume, Based on the basic volume, The noise sound pressure level was monitored. is the preset noise threshold, and k is the adjustment coefficient that is adaptively adjusted according to the environment.

[0027] Furthermore, the specific calculation process for the calibrated confidence level is as follows:

[0028] ;

[0029] Where C is the final confidence level after calibration, P is the original probability output by the deep learning classification model, and a and b are calibration parameters obtained by logistic regression fitting on the validation set using historical data.

[0030] Furthermore, in the noise source tracing sub-stage, signal preprocessing includes adaptive filtering of the microphone array signal before time delay estimation to compensate for multipath effects;

[0031] Adaptive filtering is achieved by updating its weight vector, and the formula for updating the weight vector is:

[0032] ;

[0033] in, This represents the filter weight vector at the current moment. To converge the step size, The error between the desired signal and the filter output signal, The input signal vector.

[0034] Furthermore, during the location optimization and deployment phase, the environmental parameters collected for the target site also include geographic information, vegetation coverage, and frequency of human activities. The scoring model uses the analytic hierarchy process (AHP) to construct a judgment matrix to determine the weight of each environmental parameter in the scoring.

[0035] A voiceprint monitoring system integrating bird identification and noise source tracing includes:

[0036] The location optimization deployment module is used to collect environmental parameters and run a scoring model to select the final deployment location of the microphone array;

[0037] The voiceprint data acquisition module, including a solar-powered unit and a microphone array, is used to acquire mixed sound signals in the environment;

[0038] The voiceprint analysis and processing module includes:

[0039] The bird identification unit is used to extract the vocalization features of bird calls and run a deep learning classification model.

[0040] The noise source tracing unit is used to run the sound source localization algorithm;

[0041] The results output and response module is used to output and send monitoring results and control the directional acoustic persuasion equipment.

[0042] The beneficial effects of this invention are reflected in:

[0043] By introducing a location-optimized deployment model and a multi-stage collaborative voiceprint analysis and processing mechanism, the systematic accuracy and reliability of voiceprint monitoring are effectively improved. Optimized microphone array deployment, considering a weighted scoring model that comprehensively takes into account solar irradiance, environmental obstruction, and noise levels, ensures continuous power supply and signal quality for the monitoring equipment. In the voiceprint analysis stage, a feature enhancement module with an integrated attention mechanism focuses on key features of bird calls and outputs high-confidence recognition results through probability calibration. Simultaneously, anti-interference processing combining generalized cross-correlation delay estimation and adaptive filtering enables accurate location and type identification of noise sources. Finally, the system can intelligently trigger adaptive linkage control of directional acoustic persuasion devices based on the identified noise type, achieving closed-loop management from monitoring and analysis to intervention. This not only significantly improves the accuracy of bird monitoring data but also enhances the real-time control capability of man-made noise, expanding the system's practical value in various scenarios such as ecological protection and public noise control. Attached Figure Description

[0044] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0045] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0046] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0047] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0048] like Figure 1 As shown, a voiceprint monitoring method integrating bird identification and noise source tracing includes the following steps:

[0049] Location optimization and deployment phase: Collect environmental parameters of the target location, evaluate multiple candidate locations based on the environmental parameters using a scoring model, and select the candidate location with the highest score as the final deployment location of the microphone array;

[0050] Voiceprint data acquisition phase: At the final deployment location, a solar-powered microphone array is used to acquire mixed sound signals from the environment;

[0051] Voiceprint analysis and processing stage: The mixed audio signals are processed in parallel;

[0052] The bird identification processing sub-stage extracts the voiceprint features of bird calls from the mixed sound signals and inputs them into a deep learning classification model with an integrated feature enhancement module, outputting identification results that include the bird species and their calibrated confidence scores.

[0053] The noise source tracing sub-stage involves: utilizing the inter-channel signals of the microphone array, employing a sound source localization algorithm based on time delay estimation and supplemented with anti-interference processing, calculating and outputting the direction or location information of the noise source; and identifying the type of the noise signal; the anti-interference processing includes signal preprocessing and algorithm optimization.

[0054] Result output and response stage: The bird identification result and the direction or location information of the noise source are encapsulated into a monitoring result and sent to the remote management terminal; when the identified noise type is a preset man-made noise, the directional acoustic persuasion device is automatically triggered to perform adaptive response linkage control.

[0055] During the location optimization deployment phase, the environmental parameters used to evaluate candidate locations include solar irradiance, obstacle obstruction rate, and environmental noise level.

[0056] The scoring model is a weighted scoring model calculated by first normalizing solar irradiance, obstacle obstruction rate, and environmental noise level, and then combining these factors with weighting coefficients. The specific calculation formula is as follows:

[0057] ;

[0058] Where S is the location score. I norm To normalize solar irradiance, O norm To normalize the obstacle occlusion rate, N norm To normalize the environmental noise level, and These are the weighting coefficients, and ;

[0059] By clearly defining the core environmental parameters for the optimal deployment phase (solar irradiance, obstacle obstruction rate, and ambient noise level), and employing a weighted scoring model that combines normalization with weighting coefficients, the suitability of different candidate locations can be scientifically quantified, avoiding subjectivity and blind spots in deployment location selection. Specifically, incorporating solar irradiance ensures the continuous and stable operation of the solar-powered microphone array; considering obstacle obstruction rate reduces interference from obstructions on sound signal acquisition; and assessing ambient noise level reduces the impact of background noise on subsequent voiceprint analysis. Ultimately, this ensures that the voiceprint monitoring equipment operates in the optimal environment, effectively improving the quality of voiceprint data acquisition and the overall reliability of the system.

[0060] Suppose that a voiceprint monitoring device needs to be deployed in a certain ecological monitoring area, and three candidate locations, A, B, and C, are selected. First, the solar irradiance, obstacle obstruction rate, and environmental noise level at each location are collected and normalized (after normalization, the parameter values ​​are all within the range [0,1]). Simultaneously, a weighting coefficient is set based on the device's dependence on solar power. =0.4 (Solar Irradiance Weight) =0.3 (obstacle occlusion rate weight) =0.3 (environmental noise level weight), and the specific normalization parameters for each location are as follows:

[0061] Normalized solar irradiance at candidate location A I norm =0.8, Normalized obstacle occlusion rate O norm =0.8, Normalized environmental noise level N norm =0.3;

[0062] Normalized solar irradiance at candidate location B I norm =0.6, Normalized obstacle occlusion rate O norm =0.4, Normalized environmental noise level N norm =0.2; Normalized solar irradiance at candidate location C I norm =0.7, Normalized obstacle occlusion rate O norm =0.3, Normalized environmental noise level N norm =0.4.

[0063] According to the scoring formula:

[0064] Calculate the scores for the three candidate positions respectively:

[0065] Score calculation for candidate position A:

[0066] Score calculation for candidate position B:

[0067]

[0068] Score calculation for candidate position C:

[0069] The scoring comparison shows that candidate location A has the highest score (0.77). Therefore, this location was selected as the final deployment location for the microphone array. This location can better meet the requirements of stable power supply, less signal acquisition interference, and low background noise, providing a high-quality foundation for subsequent voiceprint monitoring.

[0070] In the bird identification processing sub-stage, the feature enhancement module uses an attention mechanism; the attention mechanism achieves selective focusing on key features by calculating an attention weight matrix.

[0071] The extracted voiceprint features are Mel frequency cepstral coefficient features;

[0072] The deep learning classification model is a convolutional neural network with an integrated attention mechanism. The attention weight matrix of the attention mechanism is calculated as follows:

[0073] ;

[0074] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively. Let K be the dimension of the key vector, A be the attention weight matrix used to enhance the model's ability to focus on key frequency band features of bird calls, and T be the standard transpose notation in matrix operations, which, when applied to the key matrix K, yields the transposed key matrix. ;

[0075] The class recognition sub-stage selects Mel-frequency cepstral coefficients (MFCCs) as the voiceprint feature. This feature aligns with the human ear's perception of sound frequencies and can accurately capture the core acoustic information of bird calls (such as characteristic frequencies and harmonic distributions), providing high-quality basic features for subsequent recognition. The feature enhancement module employs an attention mechanism that selectively focuses on key frequency band features of bird calls by calculating an attention weight matrix, effectively suppressing feature interference from other ambient sounds (such as wind, insect chirps, and falling leaves). The convolutional neural network (CNN) integrating this attention mechanism retains the strong ability of CNNs to extract local acoustic features and deeply explore the subtle differences in MFCC features. It also enhances the representation of key bird recognition features through the attention mechanism. The two work together to significantly improve the deep learning classification model's ability to distinguish different bird calls, reduce misidentification and missed identification, and ensure the reliability of the recognition results' confidence, thus ensuring the accuracy and robustness against interference in the bird recognition task.

[0076] Suppose that in a monitoring scenario within a nature reserve, it is necessary to identify two bird species whose calls are easily confused: the Great Tit and the Yellow-rumped Warbler. The specific implementation process is as follows:

[0077] Voiceprint feature extraction: Collect the call sound signals of Great Tit and Yellow-rumped Warbler (mixed with a small amount of environmental wind noise interference), process the two types of sound signals frame by frame, and extract the 12-dimensional Mel-frequency cepstral coefficient (MFCC) features to obtain the MFCC feature matrix of the two types of bird calls (each row corresponds to the MFCC features of 1 frame of signal).

[0078] Attention weight matrix calculation: based on the attention weight matrix formula Set the model parameters as follows: query matrix Q, key matrix K, and value matrix V are all 1×12 dimensions (simplifying the model dimensions to fit the MFCC feature dimensions; in practical applications, these dimensions can be adjusted according to the feature size), and the key vector dimension is... =12.

[0079] Specifically, regarding the MFCC characteristics of the great tit's call, the following settings were defined: K=Q (simplified setting, actually generated by MFCC feature mapping). .

[0080] Step 1 calculation Since both Q and K are 1×12 dimensions, It has 12×1 dimensions, therefore Given a 1×1 matrix, the calculation yields:

[0081]

[0082] Second step calculation Therefore

[0083] Third step calculation softmax Function: Because the result is a single value, (High feature matching accuracy in simplified scenarios; however, in actual scenarios with multiple feature dimensions, it will be achieved through...) softmax Assign weights);

[0084] Step 4: Calculate attention weight matrix A: During this process, the attention mechanism assigns higher implicit weights to the high-frequency MFCC features of the great tit's call (corresponding to high-value elements in V), while weakening the low-value features corresponding to wind noise interference.

[0085] Model recognition performance verification: The MFCC features enhanced by the attention mechanism were input into the convolutional neural network with integrated attention, and compared with the CNN model without the attention mechanism: Without the attention mechanism, the recognition accuracy of the Great Tit was 81%, the recognition accuracy of the Yellow-rumped Warbler was 79%, and the probability of mutual misclassification between the two types of birds was 16%; After the attention mechanism was applied, the recognition accuracy of the Great Tit improved, the recognition accuracy of the Yellow-rumped Warbler improved, the probability of mutual misclassification decreased, and the confidence of the recognition results of the two types of birds was stable, which fully verified the advantages of this case in bird recognition.

[0086] In the noise source tracing sub-stage, the sound source localization algorithm based on time delay estimation is specifically the generalized cross-correlation time delay estimation method;

[0087] Anti-interference processing is achieved by introducing a specific weighting function into the generalized cross-correlation function of the generalized cross-correlation time delay estimation method;

[0088] The expression for the generalized cross-correlation function is:

[0089] ;

[0090] in, This is the phase transformation weighting function used to achieve anti-interference processing, which is used to suppress environmental reverberation and noise interference; and Let be the Fourier transform of the received signals from the i-th and j-th microphones in the microphone array. The time difference of signal arrival. j π is the imaginary unit, f is the value of pi, and f is the frequency.

[0091] By explicitly adopting the generalized cross-correlation time delay estimation method as the core sound source localization algorithm in the noise source tracing sub-stage, this algorithm can efficiently utilize the time difference information of signals between microphone array channels to achieve sound source localization. At the same time, by introducing a phase transformation weighting function Ψ(f) into the generalized cross-correlation function, anti-interference processing is achieved, which can specifically suppress the interference of environmental reverberation (such as echoes reflected from obstacles and spatial acoustic reflections) and background noise (such as wind noise and natural environmental noise) on time delay estimation, reduce localization errors, and improve the calculation accuracy of noise source direction or location information. In addition, the clear signal after anti-interference processing can provide a more reliable signal basis for noise type identification, avoid misjudgment of type caused by interference signals, ensure the accuracy and reliability of noise source tracing results, and provide accurate location and type data support for the subsequent linkage control of directional acoustic persuasion equipment.

[0092] Suppose we want to monitor human-caused noise (such as loud noise from tourists) in an urban wetland park. We would use an array of two microphones (labeled microphones i and j, spaced 0.6m apart) to trace the noise source. The specific process is as follows:

[0093] Signal Acquisition and Transformation: Microphones i and j synchronously acquire mixed sound signals from the environment (including noise from tourists and wind noise from the reeds). A Fourier transform is performed on the acquired time-domain signals to obtain the Fourier transform of the signal received by microphone i. Fourier transform of the signal received by microphone j ,in for The conjugate of complex numbers.

[0094] Generalized cross-correlation function calculation: Based on the expression of the generalized cross-correlation function, the phase transform (PHAT) weighting function is selected as the anti-interference method. ,Right now This function can suppress reverberation and noise interference by normalizing the signal amplitude.

[0095] With frequency f Taking 800Hz as an example, assuming the Fourier transform result is: (phase) ), (phase) );

[0096] First calculate Then calculate:

[0097] (Imaginary parts cancel each other out, real parts add up);

[0098] calculate ,therefore ;

[0099] Substituting the contribution term of the generalized cross-correlation function at this frequency point:

[0100] At this point, the signal amplitude is normalized, and the interference effect is greatly reduced.

[0101] Location effect verification: If not introduced (Right now =1), wind noise interference causes the peak of the generalized cross-correlation function to be blurred, and the calculated direction of tourist noise deviates by 18°, making it impossible to accurately locate the sound source; introducing Afterwards, wind noise interference was effectively suppressed, the peak value of the generalized cross-correlation function was clear, and the calculated noise source direction deviation was only 3°. The noise was successfully and accurately located from the viewing platform 20m east of the park. At the same time, the type of "human-caused noise" was identified by combining the clear signal, providing a location basis for the accurate response of subsequent directional acoustic persuasion equipment.

[0102] In the result output and response phase, the adaptive response of the linkage control specifically includes: dynamically adjusting the output volume of the persuasion device based on the monitored noise sound pressure level; the calculation formula for the playback volume is:

[0103] ;

[0104] in, For output volume, Based on the basic volume, The noise sound pressure level was monitored. is the preset noise threshold, and k is the adjustment coefficient that is adaptively adjusted according to the environment;

[0105] By clarifying the adaptive response logic of the directional acoustic persuasion device in the result output and response stages—that is, dynamically adjusting the output volume according to the monitored noise sound pressure level—the problem of "insufficient volume leading to ineffective persuasion or excessive volume causing secondary noise pollution" that may occur with a fixed volume is avoided. Simultaneously, it automatically triggers linkage control for preset man-made noise, ensuring the timeliness and targeted nature of the management; its volume calculation formula...

[0106] In this system, the adjustment coefficient k can be adaptively adjusted according to the environment, which can adapt to the acoustic needs of different scenarios (such as quiet ecological protection areas and suburban parks with high traffic). It ensures both the persuasion effect and environmental friendliness, and provides a quantifiable and intelligent execution solution for closed-loop management from noise monitoring to intervention, thereby improving the efficiency and flexibility of human noise control.

[0107] Suppose this system is deployed in a visitor area of ​​a nature reserve (the types of human noise to be intervened with are "loud noise from tourists" and "illegal use of portable speakers"), and the basic parameters are set as follows: basic volume.

[0108] Decibels (under normal conditions in this area, this volume level is clearly perceptible without disrupting the quiet ecological environment), preset noise threshold. Decibel (the threshold value at which warnings should be issued for man-made noise in this area), and an adaptive adjustment coefficient k=0.7 is set according to the vegetation coverage of the area (which affects sound propagation) (when the vegetation is dense, the sound decays slightly faster, so a moderate coefficient is used to avoid insufficient volume).

[0109] When two tourists were detected talking loudly at close range, the noise level was... Decibels (slightly above the threshold): Output volume calculated using the formula. The volume is only slightly higher than the baseline value, which can accurately remind tourists to lower the volume without causing additional disturbance to the surrounding birds' habitat.

[0110] When a tourist was detected playing music using a portable speaker, the noise level was [not specified]. Decibels (significantly above the threshold): Calculated The volume level is sufficient to effectively cover speaker noise, ensuring that the warning message "Do not use the sound system without permission" is clearly received by the user, while not exceeding the 68-decibel acoustic control limit set by the protected area.

[0111] When the noise sound pressure level is monitored Decibel (below the threshold, a mixture of natural ambient sound and tourists' quiet conversation): The system determines that the noise is not man-made and does not trigger directional acoustic guidance equipment, thus avoiding unnecessary equipment activation, reducing solar power consumption and ambient sound redundancy, and fully demonstrating the accuracy of adaptive response.

[0112] The specific calculation process for the calibrated confidence level is as follows:

[0113] ;

[0114] Where C is the final confidence level after calibration, P is the original probability output by the deep learning classification model, and a and b are calibration parameters obtained by logistic regression fitting on the validation set using historical data.

[0115] By introducing a confidence calibration mechanism based on logistic regression fitting of historical data validation sets, the deviation between the original bird identification probability P output by the deep learning classification model and the actual identification accuracy can be effectively corrected. This avoids situations where the model has "excessive confidence" (e.g., high original probability but actual identification error) or "excessive confidence" (e.g., low original probability but actual identification correct) due to problems such as training data distribution bias and overfitting. With the calibration parameters a and b obtained from the fitting, the original probability is transformed into a final confidence level C that more closely reflects the actual identification effect. This makes the confidence level of the bird identification results more valuable for reference, providing an accurate basis for the subsequent screening of monitoring data (e.g., retaining high confidence results) and manual review (e.g., verifying low confidence results), significantly improving the credibility of bird monitoring data and the scientific nature of subsequent decisions.

[0116] Suppose that in a bird identification scenario in a nature reserve, confidence calibration parameters are obtained by logistic regression fitting of a validation set (800 valid samples) of historical bird call annotation samples (including 8 common bird species such as great tits, yellow-rumped warblers, and sparrows). (The average deviation between the calibration confidence score and the actual recognition accuracy after fitting is 2.4%, which meets the requirement of "fitting logistic regression on the validation set using historical data" in the document.) For the three sets of original bird recognition probabilities P output by the deep learning classification model, according to the formula... The final confidence level C after calibration is calculated as follows:

[0117] Scenario 1: The model identifies a sound signal as "great tit" and outputs the original probability. (The model's original decision is high confidence)

[0118] Substitute into the formula to calculate: ;

[0119] because Therefore (The actual recognition accuracy of this sample was 0.82. After calibration, the deviation between C and the actual accuracy was only 1.4%, which corrected the slight overconfidence of the model. It can be directly used as high-reliability monitoring data for system entry.)

[0120] Scenario 2: The model identifies a sound signal as "Yellow-rumped Warbler" and outputs the original probability. (The model's original judgment has moderate confidence)

[0121] Substitute into the formula to calculate: ;

[0122] because Therefore (The actual recognition accuracy of this sample was 0.69. After calibration, the deviation between C and the actual accuracy was only 1.3%, which corrected the model's slight lack of confidence in recognizing the Yellow-rumped Warbler. It is suggested that further confirmation should be made in conjunction with the activity period of the Yellow-rumped Warbler in this area to avoid misrecording.)

[0123] Scenario 3: The model identifies a sound signal as "sparrow" and outputs the original probability. (The model's original decision has low confidence)

[0124] Substitute into the formula to calculate: ;

[0125] because Therefore (The actual recognition accuracy of this sample was 0.57. After calibration, the deviation between C and the actual accuracy was only 1.7%, which significantly corrected the fuzzy judgment under the original low probability of the model. This suggests that the reliability of the result is low and it needs to be excluded or manually verified to avoid invalid data affecting the monitoring statistics.)

[0126] In the noise source tracing sub-stage, signal preprocessing includes adaptive filtering of the microphone array signal before time delay estimation to compensate for multipath effects;

[0127] Adaptive filtering is achieved by updating its weight vector, and the formula for updating the weight vector is:

[0128] ;

[0129] in, This represents the filter weight vector at the current moment. To converge the step size, The error between the desired signal and the filter output signal, The input signal vector;

[0130] By introducing adaptive filtering as a signal preprocessing method before time delay estimation in the noise source tracing sub-stage, signal distortion caused by multipath effects (such as sound reflection and refraction caused by obstacles such as trees and buildings in the environment) of the microphone array signal can be specifically compensated, avoiding interference from distorted signals with the accuracy of subsequent generalized cross-correlation time delay estimation. At the same time, its explicit weight vector update formula can flexibly control the adjustment speed and stability of the filter weights through the convergence step size, ensuring that the filtering process can quickly track signal changes to compensate for multipath effects in real time, while avoiding excessive weight fluctuations that lead to unstable filter output. This significantly improves the purity of the noise signal, providing a more reliable signal basis for subsequent noise source localization (time delay estimation) and type identification, and further reducing noise source tracing errors.

[0131] Suppose we are monitoring man-made noise from tourists making loud noises in a visitor area of ​​a nature reserve. The signal received by the microphone array (2 microphones, 0.6m apart) is affected by multipath effects due to reflections from surrounding trees (the direct sound signal and the sound signal reflected from the trees are superimposed, causing signal waveform distortion). This effect needs to be compensated for through adaptive filtering. The specific process is as follows:

[0132] Set initial parameters: Based on the intensity of multipath interference in the environment, set the convergence step size of the adaptive filter. =0.01 (balancing convergence speed and stability), the filter weight vector at the current time n. =[0.12 0.08] (2D vector, corresponding to the filter weights of the two microphone channels), input signal vector =[0.35 0.28] (Sampling values ​​of the signal with multipath interference received by the microphone array at the current moment), the error between the desired signal (the approximately multipath-free signal obtained through the reference channel) and the filter output signal. =0.05 (signal deviation due to multipath effect).

[0133] Weight vector update calculation: based on the weight update formula Calculate the weight vector at the next time step n+1:

[0134] First calculate

[0135] ;

[0136] Recalculate

[0137] .

[0138] Filtering effect verification: Without adaptive filtering, the multipath effect causes two superimposed peaks in the signal waveform received by the microphone array, resulting in a deviation of 18ms in the subsequent generalized cross-correlation delay estimation, corresponding to a noise source localization deviation of approximately 6.12m (the speed of sound in air is approximately 340m / s, 340m / s × 0.018s = 6.12m). After 30 weight iterations (each adjusted according to the above formula), Afterwards, adaptive filtering effectively suppressed the interference of reflected signals, and the signal waveform retained only the single peak of the direct sound signal. The time delay estimation deviation was reduced to 3ms, and the corresponding noise source positioning deviation was reduced to 1.02m. This successfully compensated for the influence of multipath effects and provided a clean signal foundation for the subsequent accurate positioning of the "tourists making loud noises".

[0139] During the location optimization and deployment phase, the environmental parameters collected for the target site also include geographic information, vegetation coverage, and frequency of human activities. The scoring model uses the analytic hierarchy process to construct a judgment matrix to determine the weight of each environmental parameter in the scoring.

[0140] By adding three environmental parameters—geographic information, vegetation coverage, and frequency of human activity—during the location optimization and deployment phase, the limitations of the original parameters, which only focused on power supply, signal obstruction, and background noise, are overcome. The new parameters can respectively assess equipment installation adaptability (geographic information), sound propagation attenuation (vegetation coverage), and equipment safety and human interference risk (frequency of human activity), making the evaluation dimensions of candidate locations more comprehensive. At the same time, the Analytic Hierarchy Process (AHP) is used to construct a judgment matrix to determine the weight of each parameter. This allows qualitative parameters such as geographic information adaptability and quantitative parameters such as solar irradiance to be uniformly incorporated into the quantitative evaluation system. By comparing the importance of parameters pairwise, subjective weight assignment bias is eliminated, making the location score more in line with the actual monitoring scenario requirements. This ensures that the microphone array deployment location simultaneously meets the requirements of stable power supply, clear signal acquisition, easy equipment installation, and safe and low-interference operation, further improving the scientific nature of deployment decisions and the reliability of subsequent system operation.

[0141] Suppose that voiceprint monitoring equipment needs to be deployed in a mountainous nature reserve. Three candidate locations (A, B, and C) are selected. Environmental parameters are evaluated, including existing parameters (solar irradiance I, obstacle obstruction rate O, and environmental noise level N) and new parameters (geographical information adaptability G: assessing whether the terrain facilitates equipment mounting and cable routing; vegetation coverage V: assessing the impact of vegetation on sound signal propagation, with moderate coverage being optimal; and human activity frequency H: assessing the interference of human activities on equipment safety and voiceprint acquisition). The specific implementation process is as follows:

[0142] AHP judgment matrix construction and weight calculation:

[0143] Using "selecting the optimal deployment location for the microphone array" as the target layer and six types of environmental parameters (I, O, N, G, V, H) as the criterion layer, a 1-9 scale method is used to construct the criterion layer judgment matrix (1 = equally important, 3 = the former is more important than the latter, 5 = the former is significantly more important than the latter, and the reverse is taken). The matrix is ​​as follows (rows / columns correspond to I, O, N, G, V, H respectively):

[0144] ;

[0145] Calculate the largest eigenvalue of a matrix Consistency indicators (n=6 is the number of parameters), substituting, we get ;

[0146] Query random consistency index (RI values ​​corresponding to the 6 parameters), consistency ratio The consistency test is satisfied (CR<0.1 indicates that the weight allocation is reasonable).

[0147] Normalize the eigenvector corresponding to the largest eigenvalue to obtain the weights of each parameter: (Solar irradiance, the core of power supply, has the highest weight) (Obstruction occlusion rate, taken as 1-0 in calculation, less occlusion is better) (Environmental noise level, taken as 1-N in calculations, with lower noise levels being better.) (Geographic information compatibility, installation is key, weight is secondary) (Vegetation coverage, moderate coverage is optimal, use normalized value directly) (Human activity frequency, taken as 1-H in calculation, less activity is better).

[0148] Candidate position parameters are normalized (value range [0,1], the better the value, the closer to 1);

[0149] Location A: I =0.85 (mountain top area, abundant solar energy) O =0.15 (no tall trees obstructing the view, 1- O =0.85), N =0.20 (low background noise, 1- N =0.80), G =0.90 (flat platform, easy to fix equipment) V =0.75 (low shrubs, do not affect sound propagation) H =0.10 (Located far from tourist trails, with little human activity, 1- H =0.90);

[0150] Location B: I =0.60 (mountainside area, solar energy is average) O =0.40 (with some trees providing shade, 1- O =0.60), N =0.15 (extremely low background noise, 1- N =0.85、 G =0.60 (sloping terrain, additional reinforcement equipment required) V =0.90 (Dense trees may attenuate sound signals) H =0.30 (near secondary trails, moderate human activity, 1- H =0.70);

[0151] Position C: I =0.70 (foot of the mountain, moderate solar energy)O =0.30 (shaded by shrubs, 1- O =0.70), N =0.35 (High background noise, 1- N =0.65), G =0.80 (flat ground, easy to install) V =0.60 (Sparse vegetation, good sound transmission but easily disturbed by wind) H =0.50 (Close to the main trail, frequent human activity, 1- H =0.50).

[0152] Calculation of total score for candidate positions (total score = processed value of each parameter × sum of corresponding weights);

[0153] Total score for position A: 0.85×0.30+0.85×0.10+0.80×0.10+0.90×0.20+0.75×0.20+0.90×0.10=0.255+0.085+0.08+0.18+0.15+0.09=0.84;

[0154] Total score for position B: 0.60×0.30+0.60×0.10+0.85×0.10+0.60×0.20+0.90×0.20+0.70×0.10=0.18+0.06+0.085+0.12+0.18+0.07=0.695;

[0155] Total score for position C: 0.70×0.30+0.70×0.10+0.65×0.10+0.80×0.20+0.60×0.20+0.50×0.10=0.21+0.07+0.065+0.16+0.12+0.05=0.675.

[0156] Location A received the highest total score (0.84) and was selected as the final deployment location. This location perfectly met the multi-dimensional evaluation requirements due to its abundant solar energy (ensuring equipment power supply), high geographical adaptability (facilitating installation), low human activity (reducing equipment safety risks and data collection interference), and moderate vegetation cover (reducing sound attenuation), thus verifying the effectiveness of this case in location optimization deployment.

[0157] A voiceprint monitoring system integrating bird identification and noise source tracing includes:

[0158] The location optimization deployment module is used to collect environmental parameters and run a scoring model to select the final deployment location of the microphone array;

[0159] The voiceprint data acquisition module, including a solar-powered unit and a microphone array, is used to acquire mixed sound signals in the environment;

[0160] The voiceprint analysis and processing module includes:

[0161] The bird identification unit is used to extract the vocalization features of bird calls and run a deep learning classification model.

[0162] The noise source tracing unit is used to run the sound source localization algorithm;

[0163] The results output and response module is used to output and send monitoring results and control the directional acoustic persuasion equipment.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A voiceprint monitoring method integrating bird identification and noise source tracing, characterized in that: Includes the following steps: Location optimization and deployment phase: Collect environmental parameters of the target location, evaluate multiple candidate locations based on the environmental parameters using a scoring model, and select the candidate location with the highest score as the final deployment location of the microphone array; Voiceprint data acquisition phase: At the final deployment location, a solar-powered microphone array is used to acquire mixed sound signals from the environment; Voiceprint analysis and processing stage: Parallel processing of the mixed audio signals, the parallel processing process include: The bird identification processing sub-stage extracts the voiceprint features of bird calls from the mixed sound signals and inputs them into a deep learning classification model with an integrated feature enhancement module, outputting identification results that include the bird species and their calibrated confidence scores. The noise source tracing sub-stage involves: utilizing the inter-channel signals of the microphone array, employing a time-delay-based sound source localization algorithm supplemented with anti-interference processing, calculating and outputting the direction or location information of the noise source; identifying the type of noise signal; and anti-interference processing including signal preprocessing and algorithm optimization. Results Output and Response Stage: The bird identification results and the direction or location information of the noise source are encapsulated into monitoring results and sent to the remote management terminal; when the identified noise type is a preset man-made noise, the directional acoustic persuasion device is automatically triggered to perform adaptive response linkage control.

2. The voiceprint monitoring method integrating bird identification and noise source tracing according to claim 1, characterized in that: During the location optimization and deployment phase, environmental parameters used to evaluate candidate locations include solar irradiance, obstacle obstruction rate, and ambient noise level. The scoring model is a weighted scoring model that is calculated by first normalizing the solar irradiance, obstacle shading rate and environmental noise level, and then combining the weighting coefficients.

3. The voiceprint monitoring method integrating bird identification and noise source tracing according to claim 1, characterized in that: In the bird identification processing sub-stage, the feature enhancement module uses an attention mechanism; the attention mechanism achieves selective focusing on key features by calculating an attention weight matrix. The extracted voiceprint features are Mel frequency cepstral coefficient features; The deep learning classification model is a convolutional neural network with an integrated attention mechanism.

4. The voiceprint monitoring method integrating bird identification and noise source tracing according to claim 1, characterized in that: In the noise source tracing sub-stage, the sound source localization algorithm based on time delay estimation is specifically the generalized cross-correlation time delay estimation method; Anti-interference processing is achieved by introducing a specific weighting function into the generalized cross-correlation function of the generalized cross-correlation time delay estimation method.

5. The voiceprint monitoring method integrating bird identification and noise source tracing according to claim 1, characterized in that: In the result output and response phase, the adaptive response of the linkage control specifically includes: dynamically adjusting the output volume of the persuasion device based on the monitored noise sound pressure level.

6. The voiceprint monitoring method integrating bird identification and noise source tracing according to claim 3, characterized in that: The specific calculation process for the calibrated confidence level is as follows: Where C is the final confidence level after calibration, P is the original probability output by the deep learning classification model, and a and b are calibration parameters obtained by logistic regression fitting on the validation set using historical data.

7. The voiceprint monitoring method integrating bird identification and noise source tracing according to claim 4, characterized in that: In the noise source tracing sub-stage, signal preprocessing includes adaptive filtering of the microphone array signal before time delay estimation to compensate for multipath effects; Adaptive filtering is achieved by updating its weight vector.

8. The voiceprint monitoring method integrating bird identification and noise source tracing according to claim 2, characterized in that: During the location optimization and deployment phase, the environmental parameters collected for the target site also include geographic information, vegetation coverage, and frequency of human activities. The scoring model uses the analytic hierarchy process (AHP) to construct a judgment matrix to determine the weight of each environmental parameter in the scoring.

9. A voiceprint monitoring system integrating bird identification and noise source tracing, wherein the system is based on the method described in any one of claims 1-8, characterized in that: include: The location optimization deployment module is used to collect environmental parameters and run a scoring model to select the final deployment location of the microphone array; The voiceprint data acquisition module, including a solar-powered unit and a microphone array, is used to acquire mixed sound signals in the environment; The voiceprint analysis and processing module includes: The bird identification unit is used to extract the vocalization features of bird calls and run a deep learning classification model. The noise source tracing unit is used to run the sound source localization algorithm; The results output and response module is used to output and send monitoring results and control the directional acoustic persuasion equipment.

Images