Bat sound detection system based on voiceprint recognition

Abstract

The invention discloses a bat sound detection system based on voiceprint recognition, and relates to the technical field of bioacoustics and mode recognition, and the method comprises the steps: collecting the original audio of a bat in different development stages, and separating a pure ultrasonic signal from environment noise based on a preset frequency threshold value through multi-channel filtering; the signals are subjected to time marking and classified storage according to development stages, and a staged sample library is constructed. And extracting representative signal segments of each stage from the library, quantitatively analyzing acoustic features to obtain a voiceprint feature parameter set, training a pattern recognition network, and finally generating a voiceprint model capable of distinguishing a development stage and an individual identity. According to the technology, accurate extraction of the bat ultrasonic signals under complex noise is realized, a high-precision recognition model fusing individual growth time sequence characteristics is established, and the accuracy and scientificity of long-term monitoring are improved.

Description

Technical Field

[0001] This invention belongs to the field of bioacoustics and pattern recognition technology, specifically a bat sound detection system based on voiceprint recognition. Background Technology

[0002] In bat ecology and conservation biology research, sound-based non-contact individual identification and monitoring is a key technology. Current techniques typically use general-purpose microphone arrays to collect field audio containing bat calls. For signal processing, fixed-bandwidth bandpass filters or denoising algorithms based on spectral subtraction are commonly used to attempt to extract the bat's ultrasonic signals from the recordings. However, these methods often struggle to accurately distinguish noise from the target sound waves when dealing with highly complex and transient environmental background noise, especially when the noise spectrum partially overlaps with the bat calls. This can easily lead to distortion of the target signal or leave behind significant background interference, affecting the accuracy of subsequent analysis.

[0003] Existing research on bat individual identification based on voiceprints largely relies on constructing feature databases from vocalization samples collected at specific time points. These methods typically treat the sounds emitted by an individual at different stages as homogeneous data or only perform rough age grouping. However, the vocal characteristics of bats, such as frequency, harmonic structure, and pulse duration, change significantly with their growth and development. Ignoring this temporal evolution of features closely related to developmental stages and using static, mixed-developmental-stage data to train identification models leads to unstable model identification capabilities, an inability to accurately track an individual's growth trajectory over time, and limitations on the accuracy of long-term monitoring and individual lifecycle studies. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a bat sound detection system based on voiceprint recognition, comprising: The data acquisition module is used to collect the original audio sequences of bat individuals at different developmental stages in their natural state. The original audio sequences include echolocation sound waves and environmental background noise. The signal processing module is used to perform multi-channel filtering on the original audio sequence, and to separate the echo localization sound wave of the ultrasonic segment from the environmental background noise based on a preset frequency threshold to obtain a pure bat ultrasonic signal. The data storage module is used to time-stamp and classify the purified bat ultrasound signals according to individual developmental stages, and to construct a phased bat sound wave sample library. The feature extraction module is used to extract representative ultrasound signal segments corresponding to the developmental stage from the staged bat acoustic wave sample library, and to perform quantitative analysis on the acoustic features of each segment to obtain the set of voiceprint feature parameters for the corresponding developmental stage. The model training module is used to input the voiceprint feature parameter set into the pattern recognition network for training, and generate a voiceprint recognition model that can distinguish different developmental stages and individual identities.

[0005] Furthermore, the collection of raw audio sequences from bat individuals at different developmental stages in their natural state specifically includes: Deploy an ultrasonic sensor array with wideband response characteristics in a bat habitat, and set the sampling frequency and dynamic range of the ultrasonic sensor array. The sampling frequency must cover the upper limit of the vocal frequency band of the target bat species. The raw analog voltage signal received by the ultrasonic sensor array is continuously recorded, and the raw analog voltage signal is converted into discrete digital waveform data through an analog-to-digital converter; Add timestamp tags to the digital waveform data, and segment the digital waveform data into segments with fixed durations to form the original audio sequence; While acquiring the original audio sequence, the temperature and humidity environmental parameters of the acquisition point are recorded, and the environmental parameters are associated with and stored with the corresponding original audio sequence.

[0006] Furthermore, the multi-channel filtering process performed on the original audio sequence, separating the ultrasonic echo localization sound wave from the environmental background noise based on a preset frequency threshold, specifically involves: The original audio sequence is subjected to a Fast Fourier Transform to convert the time-domain waveform data into frequency-domain spectrum data; In the frequency domain spectrum data, passband frequency boundaries and stopband frequency boundaries are set according to the known vocal frequency range of bat species; Design a set of parallel digital bandpass filters, where the passband and stopband frequency boundaries of each digital bandpass filter can be configured independently to cover noise suppression or signal extraction requirements in different frequency ranges. The original audio sequence is simultaneously fed into the set of parallel digital bandpass filters, and each digital bandpass filter outputs a sub-band audio signal filtered through its specific frequency boundary. Time-domain envelope detection is performed on each sub-band audio signal to calculate the signal energy. Sub-band audio signals with signal energy exceeding the activity detection threshold are identified as valid bat ultrasound signals and are then merged to obtain the pure bat ultrasound signal.

[0007] Furthermore, the step of time-stamping and classifying the purified bat ultrasound signals according to individual developmental stages specifically involves: Based on the physiological characteristics and observation records of individual bats, the developmental stage of each collected individual was determined, including the juvenile stage, sub-adult stage, and adult stage; A unique stage identifier is assigned to each developmental stage, and an individual identifier is assigned to different individuals within the same stage. In the pure bat ultrasound signal, each signal segment is associated with a corresponding individual identifier and a stage identifier based on the timestamp of signal acquisition. A hierarchical directory structure is established according to the stage identifiers. Pure bat ultrasound signals associated with the identifiers are stored in the acoustic sample files under the corresponding hierarchical directory to form the staged bat acoustic sample library. The index file of the phased bat acoustic sample library records the mapping relationship between each acoustic sample file and the individual identifier, phase identifier, and collection timestamp.

[0008] Furthermore, the extraction of representative ultrasound signal segments corresponding to the developmental stage, and the quantitative analysis of the acoustic features of each segment to obtain the set of voiceprint feature parameters for the corresponding developmental stage, specifically involves: Read all acoustic sample files under a specific developmental stage identifier from the segmented bat acoustic sample library; Endpoint detection is performed on each acoustic sample file to locate the start and end points of each independent bat ultrasound pulse signal in the file; Each independent bat ultrasound pulse signal was extracted as a candidate segment. Cluster analysis was performed on the candidate segments of all individuals within the same developmental stage, and the most frequently occurring pulse patterns were selected as representative pulse patterns. From the candidate segments that conform to the representative pulse pattern, a preset number of segments are randomly selected as the representative ultrasound signal segments; For each representative ultrasound signal segment, a set of acoustic characteristic parameters are calculated, including center frequency, bandwidth, duration, frequency modulation slope, and harmonic structure characteristics. The acoustic feature parameters of all representative ultrasound signal segments at the same developmental stage are statistically aggregated, and the mean and variance of each feature parameter are calculated to form a set of acoustic feature parameters representing the specific developmental stage.

[0009] Furthermore, the step of inputting the voiceprint feature parameter set into the pattern recognition network for training specifically involves: A deep convolutional neural network is constructed as the pattern recognition network, the deep convolutional neural network comprising convolutional layers, pooling layers, fully connected layers and a classification output layer; The voiceprint feature parameter set is organized in the form of a feature matrix, with each voiceprint feature parameter set corresponding to a feature matrix, and the developmental stage identifier is used as the category label of the feature matrix. All feature matrices and their class labels are randomly divided into training subsets, validation subsets, and test subsets; The feature matrix in the training subset is input into a deep convolutional neural network. The predicted category is calculated through forward propagation of the network, and the error between the predicted category and the true category label is calculated through a loss function. Using the backpropagation algorithm, the weight parameters of the convolutional and fully connected layers in the deep convolutional neural network are adjusted according to the error. The validation subset is used to evaluate the performance of deep convolutional neural networks during training to prevent overfitting. When the classification accuracy of the deep convolutional neural network on the validation subset reaches the preset standard and tends to stabilize, training is stopped, and the network structure and weight parameters are fixed at this point to form the initial training model.

[0010] Furthermore, the generation of a voiceprint recognition model that distinguishes different developmental stages and individual identities specifically includes: The initial trained model is evaluated using the test subset, and the classification confusion matrix and overall recognition accuracy of the model at different developmental stages are calculated. Based on the confusion matrix, the ability of the initial training model to distinguish easily confused developmental stages is analyzed. For easily confused stages, the corresponding voiceprint feature parameter set is extracted from the staged bat echo sample library. The supplemented set of voiceprint feature parameters is incorporated into the training subset, and the initial training model is incrementally trained to optimize the model weights and improve the model's ability to distinguish between easily confused stages. Before the classification output layer of the deep convolutional neural network, a feature embedding layer is added, and the deep convolutional neural network is trained to output a feature vector with high discriminative power at the feature embedding layer. For individual identity recognition tasks, the feature vector is compared with the pre-registered individual reference feature vector, and the individual identity of the unknown sound sample is determined based on the similarity score. The optimized deep convolutional neural network model, the parameters of the feature embedding layer, and the similarity judgment rules for individual identity comparison are collectively encapsulated into the voiceprint recognition model.

[0011] Furthermore, the endpoint detection for each acoustic sample file, locating the start and end points of each independent bat ultrasound pulse signal in the file, specifically involves: Convert the time-domain waveform data of the acoustic wave sample file into a short-time energy sequence, and calculate the energy mean and standard deviation of the short-time energy sequence; A dynamic threshold based on the energy mean and standard deviation is set, and the portion of the short-time energy sequence that exceeds the dynamic threshold is marked as a high-energy region; For each high-energy region, points where the energy drops to the background noise level are searched forward and backward, and these are defined as candidate positions for the start and end points of the independent bat ultrasound pulse signal, respectively. Calculate the spectral centroid of the signal segment between the candidate start point and the end point. If the spectral centroid is within the typical frequency band of bat ultrasound, then the candidate start point and end point are confirmed as valid start points and end points. Record the time positions of all valid start and end points in the acoustic sample file to form the pulse position index table of the acoustic sample file.

[0012] Furthermore, after encapsulating the optimized deep convolutional neural network model, the parameters of the feature embedding layer, and the similarity judgment rules for individual identity comparison into the voiceprint recognition model, the process also includes online model updates: New raw bat audio sequences were continuously collected and, after multi-channel filtering, new pure bat ultrasound signals were obtained. The new, pure bat ultrasound signals were manually or semi-automatically labeled with developmental stage and individual identity. Extract a new set of voiceprint feature parameters from the newly labeled data and add the new set of voiceprint feature parameters to the original training data pool; The voiceprint recognition model is periodically retrained using an updated training data pool to adjust the model parameters to adapt to the potential slow changes in voiceprint features. The retrained model version is updated synchronously with the individual reference feature vector library.

[0013] Furthermore, the recording and acquisition points' temperature and humidity environmental parameters are linked and stored in association with the corresponding original audio sequences, specifically as follows: An environmental parameter acquisition unit is configured, which includes a temperature sensor and a humidity sensor, and is deployed at the same or adjacent locations as the ultrasonic sensor array to collect the same local environmental conditions. The acquisition clocks of the environmental parameter acquisition unit and the ultrasonic sensor array are synchronized to ensure that the environmental parameter data and the original audio sequence have a consistent time reference. The environmental parameter acquisition unit continuously acquires temperature and humidity values ​​at fixed time intervals and generates a corresponding timestamp for each acquired environmental parameter data point. When storing the original audio sequence, the temperature and humidity values ​​at the same timestamp are used as environmental parameter metadata and written into the file header of the original audio sequence or a separate associated database; When reading the original audio sequence, the corresponding temperature and humidity environmental parameters are retrieved from the environmental parameter metadata or associated database by timestamp, and used for environmental factor compensation analysis in the signal processing module or the model training module.

[0014] Compared with the prior art, the beneficial effects of the present invention are: A multi-channel filtering technique based on preset frequency thresholds was employed. For specific frequency bands in bat echogenic signals, filtering thresholds adapted to the developmental stage were set in multiple parallel channels. This technique can specifically identify and separate bat echo localization sound waves hidden in broadband environmental noise, achieving the extraction of high-fidelity, low signal-to-noise ratio pure ultrasound signals from the original audio sequence under natural conditions. This avoids signal loss or noise residue caused by traditional broadband filtering or single-threshold noise reduction, providing a reliable data foundation for subsequent feature analysis.

[0015] By time-stamping and classifying pure ultrasound signals according to individual developmental stages, and extracting representative segments from each stage for feature quantification, a set of voiceprint feature parameters with a time dimension was constructed. This method enables pattern recognition networks to learn and model the dynamic patterns of acoustic features changing with developmental stages. The trained voiceprint recognition model can thus distinguish the differences in vocal features of the same bat at different growth stages, while stably identifying its individual identity. This achieves dynamic and accurate tracking and identification of the growth and development process of individual bats, enhancing the scientific value of long-term monitoring research. Attached Figure Description

[0016] Figure 1 This is a timing diagram of the bat sound detection system based on voiceprint recognition described in this invention; Figure 2 A flowchart for classifying, storing, and constructing a sample library for acoustic signals; Figure 3 Trend chart of total daily audio duration collected from bats at different developmental stages; Figure 4 A layout of the ultrasonic sensor array in bat habitat and an optimal inspection route map for the entire data collection phase; Figure 5 A heatmap showing the correlation of five acoustic features of bat ultrasound signals. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1 The data acquisition module is responsible for collecting raw audio sequences from bats at different developmental stages in their natural state. These raw audio sequences contain both echolocation echoes and environmental background noise. The signal processing module performs multi-channel filtering on the acquired raw audio sequences. This processing, based on a preset frequency threshold, effectively separates the echolocation echoes from the environmental background noise, thus obtaining pure bat ultrasound signals. The data storage module then timestamps and categorizes the separated pure bat ultrasound signals according to individual developmental stages, constructing a structured, staged bat ultrasound sample library. The feature extraction module extracts representative ultrasound signal segments from this staged bat ultrasound sample library for different developmental stages and performs quantitative analysis of the acoustic features of each segment, ultimately obtaining a set of voiceprint feature parameters that characterize each corresponding developmental stage. The model training module inputs these voiceprint feature parameter sets into a preset pattern recognition network for training. By optimizing the network parameters, a voiceprint recognition model capable of distinguishing different developmental stages and individual identities is generated.

[0019] In one embodiment of the present invention, an ultrasonic sensor array with wideband response characteristics is deployed in a bat habitat. The sampling frequency and dynamic range of the ultrasonic sensor array need to be preset, and the sampling frequency setting of the ultrasonic sensor array needs to cover the upper limit of the vocal frequency band of the target bat species. The ultrasonic sensor array continuously records the received raw analog voltage signal, which is converted into discrete digital waveform data by an analog-to-digital converter. A timestamp is added to the digital waveform data, and the digital waveform data is segmented in units of fixed duration to form a raw audio sequence. In practice, while acquiring the original audio sequence, the temperature and humidity environmental parameters at the acquisition points are recorded. An environmental parameter acquisition unit, including a temperature sensor and a humidity sensor, is configured and deployed at the same or adjacent location as the ultrasonic sensor array to acquire the same local environmental conditions. The acquisition clock of the environmental parameter acquisition unit is synchronized with that of the ultrasonic sensor array to ensure that the environmental parameter data and the original audio sequence have a consistent time reference. The environmental parameter acquisition unit continuously acquires temperature and humidity values ​​at fixed time intervals and generates a corresponding timestamp for each acquired environmental parameter data point. When storing the original audio sequence, the temperature and humidity values ​​at the same timestamp are used as environmental parameter metadata and written to the header of the original audio sequence file or a separate associated database. When reading the original audio sequence, the corresponding temperature and humidity environmental parameters are retrieved from the environmental parameter metadata or associated database using the timestamp.

[0020] In some embodiments, the ultrasonic sensor array is deployed in areas with frequent bat activity, such as cave entrances or under forest canopies. The sampling frequency of the ultrasonic sensor array is dynamically adjusted according to the vocalization characteristics of the target bat species. For bat species with a vocalization frequency upper limit of 120kHz, the sampling frequency of the ultrasonic sensor array is set to 250kHz. The dynamic range of the ultrasonic sensor array is set to capture amplitude variations from weak to strong bat calls, for example, a dynamic range of 80dB. In some embodiments, the fixed time interval of the environmental parameter acquisition unit is set to be consistent with the segment duration of the original audio sequence. For example, when the segment duration is 1 second, the environmental parameter acquisition unit acquires temperature and humidity values ​​at a frequency of once per second. When the environmental parameter metadata is stored in the associated database, an index relationship is established between the timestamp and the original audio sequence. The associated database adopts a time-series database structure to achieve efficient querying.

[0021] Optionally, the sampling frequency of the ultrasonic sensor array is set according to the formula: ; in: This indicates the sampling frequency of the ultrasonic sensor array. This indicates the upper limit of the vocal frequency band of the target bat species. This represents an oversampling coefficient greater than 2; for bat species with an upper limit of 100kHz in their vocal frequency range, it is set to... At that time, the sampling frequency of the ultrasonic sensor array is calculated to be 250kHz. Optionally, the temperature and humidity sensors of the environmental parameter acquisition unit are digital output models, and the data from the temperature and humidity sensors are connected to the main controller of the environmental parameter acquisition unit through a serial communication interface; the main controller of the environmental parameter acquisition unit and the data acquisition card of the ultrasonic sensor array share the same clock source to achieve clock synchronization.

[0022] It is understandable that the segment duration of the original audio sequence is set according to the typical duration of bat calls, for example, a segment duration of 5 seconds to cover multiple complete bat ultrasound pulse sequences. It is also understandable that when environmental parameter metadata is written to the file header of the original audio sequence, the file header format adopts an Extensible Markup Language (XML) structure, with temperature and humidity values ​​embedded as tag attributes; the environmental parameter data in the associated database uses the timestamp as the primary key, and the file path of the original audio sequence as a foreign key for association and storage. In specific implementations, the wideband response characteristics of the ultrasonic sensor array cover a frequency range from 20kHz to 200kHz, and each sensor unit of the ultrasonic sensor array has a uniform directional pattern to reduce acquisition deviation; the analog-to-digital converter (ADC) resolution of the original analog voltage signal is set to 16 bits, and the sampling clock of the ADC is locked to the acquisition clock of the ultrasonic sensor array. In specific implementations, the timestamp tags are generated based on the GPS clock or Network Time Protocol (NAT) clock, and the timestamp tag accuracy reaches the millisecond level; the segmentation operation with fixed duration units adopts an overlapping segmentation method, with a 50% overlap rate between adjacent segments to increase data continuity.

[0023] In one embodiment of the present invention, multi-channel filtering is performed on the original audio sequence to separate the echo localization sound wave of the ultrasonic band from the environmental background noise based on a preset frequency threshold. A Fast Fourier Transform (FFT) is then performed on the original audio sequence to convert the time-domain waveform data into frequency-domain spectral data. In the frequency-domain spectral data, passband and stopband frequency boundaries are set according to the known vocalization frequency range of bat species. A set of parallel digital bandpass filters is designed, with each digital bandpass filter having its passband and stopband frequency boundaries independently configurable to cover noise suppression or signal extraction requirements across different frequency ranges. The original audio sequence is simultaneously fed into the set of parallel digital bandpass filters, with each digital bandpass filter outputting a sub-band audio signal filtered through its specific frequency boundaries. Time-domain envelope detection is performed on each sub-band audio signal to calculate the signal energy. Sub-band audio signals with signal energy exceeding the activity detection threshold are determined to be valid bat ultrasonic signals. All determined valid bat ultrasonic signals are then merged to obtain a pure bat ultrasonic signal.

[0024] In some embodiments, the known bat species is the horseshoe bat, whose vocalization frequency range is mainly concentrated between 80 kHz and 120 kHz. The passband frequency boundaries are set at 75 kHz and 125 kHz, and the stopband frequency boundaries are set at 70 kHz and 130 kHz in the frequency domain spectral data. In some embodiments, a set of parallel digital bandpass filters includes four independent digital bandpass filters. The first digital bandpass filter has a passband frequency boundary set at 30 kHz to 50 kHz to suppress low-frequency wind noise; the second digital bandpass filter has a passband frequency boundary set at 50 kHz to 80 kHz to capture any possible lower-frequency bat call components; the third digital bandpass filter has a passband frequency boundary set at 80 kHz to 120 kHz to match the horseshoe bat's main vocalization frequency band; and the fourth digital bandpass filter has a passband frequency boundary set at 120 kHz to 150 kHz to monitor higher-frequency harmonics or signals from other species. The stopband attenuation of each digital bandpass filter is set to at least 60 dB.

[0025] Optionally, time-domain envelope detection is achieved by calculating the short-time average amplitude of the sub-band audio signal, and the signal energy. Through the formula: ; in: This represents the average amplitude energy of the m-th sub-band audio signal within a time window. This indicates the number of sample points within the time window. This represents the nth sample value of the mth sub-band audio signal. Indicates the absolute value operation; activity detection threshold. Set as background noise energy estimate Multiples of, i.e. , This is a preset threshold coefficient. Optional, background noise energy estimate. The average amplitude of the subband audio signal was obtained by calculating the audio signal in an audio segment where no bat activity was confirmed.

[0026] It is understandable that the Fast Fourier Transform uses a Hanning window of length 1024 points, and the frame shift is set to 256 points to balance frequency domain resolution and computational efficiency. It is also understandable that the operation of merging all deemed valid bat ultrasound signals involves superimposing all sub-band audio signal segments with energy exceeding the activity detection threshold in the time domain. Before superposition, time alignment compensation is performed based on the center frequency of each sub-band audio signal to eliminate phase shift caused by filter group delay differences. In the specific implementation, the digital bandpass filter is designed using a finite impulse response (FIR) filter structure. The FIR filter order is set to 200 to ensure a sufficiently steep transition between the passband and stopband. The configuration information for the passband and stopband frequency boundaries is stored in an accessible configuration file, allowing different filter parameter sets to be dynamically loaded based on different bat species or environmental noise spectrum characteristics. In practice, the original audio sequence is simultaneously fed into a set of parallel digital bandpass filters through a multi-threaded or parallel computing architecture, with each digital bandpass filter running on an independent processing thread or computing core; time-domain envelope detection of each sub-band audio signal is equivalently implemented using a circuit or algorithm that performs full-wave rectification followed by a first-order low-pass smoothing filter.

[0027] In one embodiment of the present invention, see [reference] Figure 2 Based on the physiological characteristics and observation records of individual bats, the developmental stage of each collected individual was determined, including juvenile, subadult, and adult stages. A unique stage identifier was assigned to each developmental stage, using alphanumeric encoding, for example, "JUV" for juvenile, "SUB" for subadult, and "ADT" for adult. Individual identifiers were also assigned to different individuals within the same stage, using consecutive numerical numbers, such as "IND_001" and "IND_002". In the pure bat ultrasound signal, the individual identifier and stage identifier were associated with each signal segment according to the timestamp of the signal acquisition. The association operation was completed by parsing the acquisition time information embedded in the audio file name and matching it with the timeline of the individual's feeding record or field observation log. A hierarchical directory structure is established based on stage identifiers. The root directory of the hierarchical directory structure is "BatSoundLib". Subdirectories named after stage identifiers are created under the root directory. Pure bat ultrasound signals associated with the identifiers are stored in the corresponding acoustic sample files under the hierarchical directories. The acoustic sample files are named in the format of "individual identifier_timestamp.wav", forming a staged bat acoustic sample library. In the index file of the staged bat acoustic sample library, the index file uses a comma-separated value format to record the mapping relationship between each acoustic sample file and the individual identifier, stage identifier, and acquisition timestamp. The mapping relationship is stored in the form of rows, and each row contains four fields: file name, individual identifier, stage identifier, and UTC timestamp.

[0028] In some embodiments, physiological characteristics and observation records include regular measurements and recordings of individual bat weight, forearm length, pubic symphysis status, and feather or hair growth. The criteria for juvenile stage are: weight less than 60% of average adult weight and incomplete feather coverage; sub-adult stage is: weight between 60% and 95% of average adult weight and flight capability but immature reproductive organs; adult stage is: weight within the average adult weight range and mature reproductive organs. In some embodiments, the allocation of stage identifiers and individual identifiers is managed through a central database. This central database records the unique biometrics and all measurement records for each individual bat. When a pure bat ultrasound signal needs to be labeled, the system queries the central database based on the signal acquisition time, automatically retrieves individuals active at that time and their current developmental stage, and automatically assigns the corresponding identifier.

[0029] Optionally, the association process between a timestamp and an identifier can be achieved using a formula: ; in: A composite identifier that represents the stage and individual information assigned to an audio signal. Indicates from the time of collection Mapping function to identifier, mapping function The implementation is based on querying the observation log database, which stores the developmental stage transition time points corresponding to each individual identifier. ;when Falling on a certain individual identifier When within the interval, The function returns the individual identifier and the corresponding stage identifier. Optionally, the acoustic sample files under the hierarchical directory structure are stored in a way that further organizes them by date. For example, a subdirectory named "2023-10-05" is created under the " / BatSoundLib / ADT" directory to store all adult individual audio files collected on that date.

[0030] It is understandable that the mapping relationship of the index file supports fast queries based on any field. For example, by using the individual identifier "IND_005", a list of all acoustic sample files belonging to that individual, along with their corresponding acquisition time and developmental stage, can be retrieved. It is also understandable that before storing pure bat ultrasound signals in acoustic sample files, their associated individual identifiers and stage identifiers are written as metadata into the "LIST" information block of the WAV file to achieve self-contained descriptive information in the audio file. In specific implementation, the acquisition timestamp comes from the global synchronization clock of the data acquisition module, with a timestamp accuracy of milliseconds, and is consistent with the filename and header information of the original audio sequence. The allocation of individual identifiers remains unchanged for the same individual throughout the long-term monitoring project, even if the individual moves from one developmental stage to the next. In specific implementation, in addition to the basic mapping relationship, the index file of the staged bat acoustic sample library can also be expanded to include additional metadata fields such as file size, audio duration, and average signal-to-noise ratio. These additional metadata fields are automatically generated by analyzing the acoustic sample files and appended to the corresponding lines of the index file.

[0031] See Figure 3 This study presents the daily total audio duration trends of juvenile, sub-adult, and adult bats on different collection dates. Specifically, the audio duration of adult bats fluctuated the most during the collection period, showing significant peaks at 01-07 and 01-14, especially exceeding 65 minutes on 01-14, reflecting strong temporal heterogeneity in the activity of adult bats. The audio duration of juvenile bats reached a peak of approximately 43 minutes on 01-08, exhibiting an overall fluctuating characteristic of "rising first, then falling, and then rising again," which is related to the changes in activity rhythms during their developmental stages. The audio duration of sub-adult bats remained relatively stable (5-25 minutes), with only minor fluctuations on 01-06 and 01-11, reflecting the stability of bat activity patterns at this stage.

[0032] In one embodiment of the present invention, all acoustic sample files under a specific developmental stage identifier are read from a staged bat acoustic sample library. Endpoint detection is performed on each acoustic sample file to locate the start and end points of each independent bat ultrasound pulse signal within the file. Specifically, the endpoint detection process involves: converting the time-domain waveform data of the acoustic sample file into a short-time energy sequence; calculating the energy mean and standard deviation of the short-time energy sequence; and setting a dynamic threshold based on the energy mean and standard deviation. formula: ; in: This represents the energy mean of a short-time energy series. The standard deviation of a short-time energy series is represented by its value. This represents the dynamic threshold coefficient. It is used to select short-time energy sequences that exceed the dynamic threshold. The region is marked as a high-energy region; for each high-energy region, points where the energy drops to the background noise level are searched forward and backward. The background noise level is determined by calculating the short-time energy minimum of the silent portion in the audio segment. These points are defined as candidate locations for the start and end points of the independent bat ultrasound pulse signal, respectively; the spectral centroid of the signal segment between the candidate start and candidate end points is calculated. Through the formula: ; in: This represents the frequency value of the i-th frequency point in the spectrum. This indicates the amplitude at the corresponding frequency point. This represents the total number of spectral points; if the spectral centroid... It falls within the typical frequency band of bat ultrasound, specifically the lower limit of the typical frequency band. Set to 20kHz, the upper limit of the typical frequency band. If the frequency is set to 120kHz, the candidate start point and candidate end point are confirmed as valid start and end points; the time positions of all valid start and end points in the sound wave sample file are recorded to form the pulse position index table of the sound wave sample file.

[0033] In some embodiments, the specific developmental stage identifier read from the phased bat acoustic sample library is "JUV," i.e., juvenile stage. The read operation traverses all WAV format acoustic sample files stored in the " / BatSoundLib / JUV" directory, with a total of 150 acoustic sample files. In some embodiments, the conversion of the short-time energy sequence uses a Hamming window with a window length of 256 sample points and a frame shift of 64 sample points; a dynamic threshold coefficient... The value is set to 2.5; the background noise level is calculated by selecting the first and last 100-millisecond audio segments in the sound wave sample file that are sure there is no bat activity, and calculating the minimum value of the short-time energy sequence of these audio segments.

[0034] Each independent bat echolocation pulse signal was extracted as a candidate segment. Cluster analysis was performed on candidate segments from all individuals within the same developmental stage using a hierarchical clustering algorithm based on dynamic time-warped distance. The most frequently occurring pulse patterns were selected as representative pulse patterns. From the candidate segments conforming to the representative pulse patterns, a predetermined number of segments were randomly selected as representative echolocation pulse segments, with 50 segments selected for each representative pulse pattern. A set of acoustic characteristic parameters was calculated for each representative echolocation pulse segment, including center frequency, bandwidth, duration, frequency modulation slope, and harmonic structure characteristics. The acoustic characteristic parameters of all representative echolocation pulse segments from the same developmental stage were statistically aggregated, and the mean and variance of each characteristic parameter were calculated to form a set of acoustic signature characteristic parameters representing a specific developmental stage (see Table 1).

[0035] Table 1: Representative pulse morphoacoustic characteristics of juvenile bats

[0036] Optionally, the frequency modulation slope is calculated by linearly fitting the instantaneous frequency curve of a representative ultrasound signal segment. The instantaneous frequency is obtained by calculating the derivative of the analytic phase of the signal using Hilbert transform. Optionally, the harmonic structure characteristics are determined by analyzing the power spectrum of a representative ultrasound signal segment. If there is a significant spectral peak in the power spectrum at an integer multiple of the fundamental frequency, it is recorded as the presence of harmonics, and the amplitude ratio of the strongest harmonic to the fundamental peak is recorded.

[0037] It is understandable that the number of representative pulse morphologies selected in cluster analysis is set empirically, for example, 3 to 5 types to cover the main vocal variation patterns within the developmental stage. It is also understandable that the set of vocal feature parameters formed by statistical aggregation is stored in the form of a data structure or configuration file, such as a JSON file, containing developmental stage identifiers and a list of the average and variance of the feature parameters for each representative pulse morphology. In practice, the determination of background noise levels during endpoint detection employs a dynamic update mechanism, re-estimating the background noise level at regular intervals in long-duration recordings to adapt to slow changes in environmental noise. In practice, the pulse position index table is stored as a separate text file or database table. Each record in the pulse position index table includes the sound sample file name, pulse number, start time, end time, and spectral centroid value, facilitating rapid location and retrieval of audio data during subsequent feature extraction.

[0038] See Figure 4In the full acquisition phase, the planar layout of the ultrasonic sensor array and the planning of the acquisition and inspection routes are presented in two-dimensional coordinates to illustrate the monitoring deployment scheme for bat habitats. The blue rectangle in the figure marks the core area of ​​bat habitats, containing four ultrasonic sensor nodes: S05, S06, S07, and S08. The effective acquisition count ranges from 28 to 31, reflecting that this area is a high-density area of ​​bat activity. Node S08: 28 effective acquisitions, the lowest in the entire figure. This node is located at the edge of the core area and is suitable as a background noise sample acquisition point, providing noise baseline data for multi-channel filtering. S06: 31 effective acquisitions. Its signal is stable and is the main source for acquiring typical pulse patterns (A / B / C) during the juvenile stage, providing core data for constructing the voiceprint feature parameter set. S07: 30 effective acquisitions. Adjacent to S06 and with the same acquisition count, it is suitable for individual identification training, and the discriminative ability of the feature embedding layer can be verified by comparing the signals of the two nodes. S05: 29 effective acquisitions. Due to potential signal overlap, the data from this node can be used to verify the accurate positioning capability of the endpoint detection algorithm in complex scenarios. The red dashed line represents the optimal data collection and inspection route, connecting S01→S05→S07→S08→S06→S02→S04→S03 in sequence, achieving efficient coverage of the core area and peripheral nodes. S01: 42 effective data collections. As an environmental compensation benchmark point, temperature and humidity parameters need to be recorded in detail to study the impact of environmental factors on voiceprint features. S02: 38 effective data collections. As a path planning hub, it balances the spatial distribution of the sample library and provides medium-density samples outside the core area. S04: 36 effective data collections. As a long-distance detection point, it is used to verify the performance of the broadband ultrasonic sensor under low signal-to-noise ratio and the effectiveness of subsequent filtering processing. S03: 45 effective data collections, the highest in the entire map. This is the throat passage for bats entering and exiting the core area, providing a large number of ultrasonic pulses in flight, greatly enriching the feature parameter set. The color gradient bar on the right quantifies the effective number of samples collected at each node using gradient colors, ranging from 28 to 44. Node S03 has the highest number of effective samples (45), indicating that bat activity is most active at this location. Node S08 has the lowest number of effective samples (28), providing data for subsequent sensor layout optimization.

[0039] In one embodiment of the present invention, a deep convolutional neural network is constructed as a pattern recognition network. The deep convolutional neural network includes convolutional layers, pooling layers, fully connected layers, and a classification output layer. The voiceprint feature parameter set is organized in the form of a feature matrix. Each voiceprint feature parameter set corresponds to a feature matrix. The rows of the feature matrix represent a representative ultrasound signal segment, and the columns of the feature matrix represent different acoustic feature parameters, such as center frequency, bandwidth, duration, frequency modulation slope, and harmonic structure feature quantization values. A developmental stage identifier is used as the category label for the feature matrix. All feature matrices and their category labels are randomly divided into a training subset, a validation subset, and a test subset, with the division ratio set to 70% for the training subset, 15% for the validation subset, and 15% for the test subset. The feature matrices in the training subset are input into the deep convolutional neural network. The predicted category is calculated through forward propagation of the network, and the error between the predicted category and the true category label is calculated using a loss function. The cross-entropy loss function is used, and the formula is: ; in: This represents the average loss value for a batch. Indicates batch size, This represents the total number of developmental stage categories. It is an indicator function; its value is 1 when the true class of sample i is c, and 0 otherwise. This represents the probability that the deep convolutional neural network predicts sample i belongs to class c; the weight parameters of the convolutional and fully connected layers in the deep convolutional neural network are adjusted according to the error using the backpropagation algorithm; the performance of the deep convolutional neural network during training is evaluated using a validation subset, with the classification accuracy as the performance evaluation metric; when the classification accuracy of the deep convolutional neural network on the validation subset reaches the preset standard and tends to stabilize, training is stopped, and the network structure and weight parameters at this point are fixed to form the initial training model.

[0040] In some embodiments, the architecture of a deep convolutional neural network includes two convolutional layers. The first convolutional layer uses 64 3x3 kernels, and the second convolutional layer uses 128 3x3 kernels. Each convolutional layer is followed by a 2x2 max-pooling layer. After the pooling layers, two fully connected layers are connected. The first fully connected layer contains 256 neurons, and the second fully connected layer, which is the classification output layer, contains the same number of neurons as the number of categories at the developmental stage. In some embodiments, the preset classification accuracy standard is set to ensure that the accuracy of the verified subset no longer improves within 10 consecutive training epochs, and the accuracy value exceeds 95%. After training stops, the network weight file and structure definition file of the initial trained model are saved.

[0041] The initial trained model was evaluated using a test subset, and the confusion matrix and overall recognition accuracy for different developmental stages were calculated. Based on the confusion matrix, the initial trained model's ability to distinguish easily confused developmental stages was analyzed. For these easily confused stages, corresponding voiceprint feature parameter sets were extracted from a stage-specific bat echo sample database. These supplementary voiceprint feature parameter sets were incorporated into the training subset, and the initial trained model was incrementally trained to optimize model weights. A feature embedding layer (a fully connected layer with 128 neurons) was added before the classification output layer of the deep convolutional neural network. The deep convolutional neural network was trained to output highly discriminative feature vectors at the feature embedding layer. For individual identification tasks, the feature vectors were compared with pre-registered individual reference feature vectors using cosine similarity. The individual identity of the unknown sound sample was determined based on the similarity score, with the individual identity corresponding to the highest-scoring reference feature vector exceeding a preset threshold. The optimized deep convolutional neural network model, the parameters of the feature embedding layer, and the similarity determination rule for individual identity comparison were collectively encapsulated into a voiceprint recognition model.

[0042] Optionally, incremental training uses a smaller learning rate, such as one-tenth of the initial training learning rate, to prevent overfitting to newly added data and forgetting existing knowledge. Optionally, the preset threshold for individual identity comparison is determined by calculating the statistical distribution of similarity scores between registered individual feature vectors; for example, the threshold is set as the average similarity between registered individuals minus twice the standard deviation. It can be understood that the training of the feature embedding layer is driven by a joint loss function combining classification loss and triplet loss, to encourage feature vectors of individuals at the same developmental stage or with the same identity to cluster in the embedding space, while feature vectors of different categories are kept apart. It can be understood that the encapsulated voiceprint recognition model includes an executable file and a model parameter file, and provides an application programming interface (API). The API receives a new, clean bat ultrasound signal feature matrix as input and returns developmental stage classification results and individual identity recognition results.

[0043] See Figure 5The figure presents the pairwise Pearson correlation coefficients of five core acoustic features: center frequency, bandwidth, duration, frequency modulation slope, and harmonic structure characteristics. As can be seen from the figure, the correlation coefficients between all features are in the range of 0.77 to 0.86, showing a significant positive correlation. The highest correlation coefficient is between harmonic structure characteristics and frequency modulation slope (0.86), while the lowest is between center frequency and bandwidth (0.77). This high positive correlation indicates that the various acoustic features of bat ultrasound signals do not change independently but are synergistically correlated; for example, changes in frequency modulation slope are usually accompanied by synchronous changes in harmonic structure characteristics. This result provides a crucial basis for the subsequent construction of the voiceprint feature parameter set: during the model training phase, the collinearity problem between features needs to be considered. Dimensionality reduction methods such as principal component analysis (PCA) can be used to reduce redundant information, thereby improving the efficiency and robustness of the pattern recognition network.

[0044] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A bat sound detection system based on voiceprint recognition, characterized in that, include: The data acquisition module is used to collect the original audio sequences of bat individuals at different developmental stages in their natural state. The original audio sequences include echolocation sound waves and environmental background noise. The signal processing module is used to perform multi-channel filtering on the original audio sequence, and to separate the echo localization sound wave of the ultrasonic segment from the environmental background noise based on a preset frequency threshold to obtain a pure bat ultrasonic signal. The data storage module is used to time-stamp and classify the purified bat ultrasound signals according to individual developmental stages, and to construct a phased bat sound wave sample library. The feature extraction module is used to extract representative ultrasound signal segments corresponding to the developmental stage from the staged bat acoustic wave sample library, and to perform quantitative analysis on the acoustic features of each segment to obtain the set of voiceprint feature parameters for the corresponding developmental stage. The model training module is used to input the voiceprint feature parameter set into the pattern recognition network for training, and generate a voiceprint recognition model that can distinguish different developmental stages and individual identities.

2. The bat sound detection system based on voiceprint recognition according to claim 1, characterized in that, The original audio sequences collected from bat individuals at different developmental stages in their natural state are specifically as follows: Deploy an ultrasonic sensor array with wideband response characteristics in a bat habitat, and set the sampling frequency and dynamic range of the ultrasonic sensor array. The sampling frequency must cover the upper limit of the vocal frequency band of the target bat species. The raw analog voltage signal received by the ultrasonic sensor array is continuously recorded, and the raw analog voltage signal is converted into discrete digital waveform data through an analog-to-digital converter; Add timestamp tags to the digital waveform data, and segment the digital waveform data into segments with fixed durations to form the original audio sequence; While acquiring the original audio sequence, the temperature and humidity environmental parameters of the acquisition point are recorded, and the environmental parameters are associated with and stored with the corresponding original audio sequence.

3. The bat sound detection system based on voiceprint recognition according to claim 1, characterized in that, The process of performing multi-channel filtering on the original audio sequence, separating the ultrasonic echo localization sound wave from the environmental background noise based on a preset frequency threshold, specifically involves: The original audio sequence is subjected to a Fast Fourier Transform to convert the time-domain waveform data into frequency-domain spectrum data; In the frequency domain spectrum data, passband frequency boundaries and stopband frequency boundaries are set according to the known vocal frequency range of bat species; Design a set of parallel digital bandpass filters, where the passband and stopband frequency boundaries of each digital bandpass filter can be configured independently to cover noise suppression or signal extraction requirements in different frequency ranges. The original audio sequence is simultaneously fed into the set of parallel digital bandpass filters, and each digital bandpass filter outputs a sub-band audio signal filtered through its specific frequency boundary. Time-domain envelope detection is performed on each sub-band audio signal to calculate the signal energy. Sub-band audio signals with signal energy exceeding the activity detection threshold are identified as valid bat ultrasound signals and are then merged to obtain the pure bat ultrasound signal.

4. The bat sound detection system based on voiceprint recognition according to claim 1, characterized in that, The process of time-stamping and classifying the purified bat ultrasound signals according to individual developmental stages specifically involves: Based on the physiological characteristics and observation records of individual bats, the developmental stage of each collected individual was determined, including the juvenile stage, sub-adult stage, and adult stage; A unique stage identifier is assigned to each developmental stage, and an individual identifier is assigned to different individuals within the same stage. In the pure bat ultrasound signal, each signal segment is associated with a corresponding individual identifier and a stage identifier based on the timestamp of signal acquisition. A hierarchical directory structure is established according to the stage identifiers. Pure bat ultrasound signals associated with the identifiers are stored in the acoustic sample files under the corresponding hierarchical directory to form the staged bat acoustic sample library. The index file of the phased bat acoustic sample library records the mapping relationship between each acoustic sample file and the individual identifier, phase identifier, and collection timestamp.

5. A bat sound detection system based on voiceprint recognition according to claim 4, characterized in that, The process involves extracting representative ultrasound signal segments corresponding to the developmental stage, quantitatively analyzing the acoustic features of each segment, and obtaining the set of voiceprint feature parameters for the corresponding developmental stage. Specifically: Read all acoustic sample files under a specific developmental stage identifier from the segmented bat acoustic sample library; Endpoint detection is performed on each acoustic sample file to locate the start and end points of each independent bat ultrasound pulse signal in the file; Each independent bat ultrasound pulse signal was extracted as a candidate segment. Cluster analysis was performed on the candidate segments of all individuals within the same developmental stage, and the most frequently occurring pulse patterns were selected as representative pulse patterns. From the candidate segments that conform to the representative pulse pattern, a preset number of segments are randomly selected as the representative ultrasound signal segments; For each representative ultrasound signal segment, a set of acoustic characteristic parameters are calculated, including center frequency, bandwidth, duration, frequency modulation slope, and harmonic structure characteristics. The acoustic feature parameters of all representative ultrasound signal segments at the same developmental stage are statistically aggregated, and the mean and variance of each feature parameter are calculated to form a set of acoustic feature parameters representing the specific developmental stage.

6. The bat sound detection system based on voiceprint recognition according to claim 1, characterized in that, The step of inputting the voiceprint feature parameter set into the pattern recognition network for training specifically involves: A deep convolutional neural network is constructed as the pattern recognition network, the deep convolutional neural network comprising convolutional layers, pooling layers, fully connected layers and a classification output layer; The voiceprint feature parameter set is organized in the form of a feature matrix, with each voiceprint feature parameter set corresponding to a feature matrix, and the developmental stage identifier is used as the category label of the feature matrix. All feature matrices and their class labels are randomly divided into training subsets, validation subsets, and test subsets; The feature matrix in the training subset is input into a deep convolutional neural network. The predicted category is calculated through forward propagation of the network, and the error between the predicted category and the true category label is calculated through a loss function. Using the backpropagation algorithm, the weight parameters of the convolutional and fully connected layers in the deep convolutional neural network are adjusted according to the error. The validation subset is used to evaluate the performance of deep convolutional neural networks during training to prevent overfitting. When the classification accuracy of the deep convolutional neural network on the validation subset reaches the preset standard and tends to stabilize, training is stopped, and the network structure and weight parameters are fixed at this point to form the initial training model.

7. A bat sound detection system based on voiceprint recognition according to claim 6, characterized in that, The generation of a voiceprint recognition model that distinguishes different developmental stages and individual identities specifically includes: The initial trained model is evaluated using the test subset, and the classification confusion matrix and overall recognition accuracy of the model at different developmental stages are calculated. Based on the confusion matrix, the ability of the initial training model to distinguish easily confused developmental stages is analyzed. For easily confused stages, the corresponding voiceprint feature parameter set is extracted from the staged bat echo sample library. The supplemented set of voiceprint feature parameters is incorporated into the training subset, and the initial training model is incrementally trained to optimize the model weights and improve the model's ability to distinguish between easily confused stages. Before the classification output layer of the deep convolutional neural network, a feature embedding layer is added, and the deep convolutional neural network is trained to output a feature vector with high discriminative power at the feature embedding layer. For individual identity recognition tasks, the feature vector is compared with the pre-registered individual reference feature vector, and the individual identity of the unknown sound sample is determined based on the similarity score. The optimized deep convolutional neural network model, the parameters of the feature embedding layer, and the similarity judgment rules for individual identity comparison are collectively encapsulated into the voiceprint recognition model.

8. A bat sound detection system based on voiceprint recognition according to claim 5, characterized in that, The endpoint detection for each acoustic sample file, locating the start and end points of each independent bat ultrasound pulse signal in the file, specifically involves: Convert the time-domain waveform data of the acoustic wave sample file into a short-time energy sequence, and calculate the energy mean and standard deviation of the short-time energy sequence; A dynamic threshold based on the energy mean and standard deviation is set, and the portion of the short-time energy sequence that exceeds the dynamic threshold is marked as a high-energy region; For each high-energy region, points where the energy drops to the background noise level are searched forward and backward, and these are defined as candidate positions for the start and end points of the independent bat ultrasound pulse signal, respectively. Calculate the spectral centroid of the signal segment between the candidate start point and the end point. If the spectral centroid is within the typical frequency band of bat ultrasound, then the candidate start point and end point are confirmed as valid start points and end points. Record the time positions of all valid start and end points in the acoustic sample file to form the pulse position index table of the acoustic sample file.

9. A bat sound detection system based on voiceprint recognition according to claim 7, characterized in that, After encapsulating the optimized deep convolutional neural network model, the parameters of the feature embedding layer, and the similarity judgment rules for individual identity comparison into the voiceprint recognition model, the process also includes online model updates. New raw bat audio sequences were continuously collected and, after multi-channel filtering, new pure bat ultrasound signals were obtained. The new, pure bat ultrasound signals were manually or semi-automatically labeled with developmental stage and individual identity. Extract a new set of voiceprint feature parameters from the newly labeled data and add the new set of voiceprint feature parameters to the original training data pool; The voiceprint recognition model is periodically retrained using an updated training data pool to adjust the model parameters to adapt to the potential slow changes in voiceprint features. The retrained model version is updated synchronously with the individual reference feature vector library.

10. A bat sound detection system based on voiceprint recognition according to claim 2, characterized in that, The temperature and humidity environmental parameters of the recorded acquisition points are then associated and stored with the corresponding original audio sequences. Specifically: An environmental parameter acquisition unit is configured, which includes a temperature sensor and a humidity sensor, and is deployed at the same or adjacent locations as the ultrasonic sensor array to collect the same local environmental conditions. The acquisition clocks of the environmental parameter acquisition unit and the ultrasonic sensor array are synchronized to ensure that the environmental parameter data and the original audio sequence have a consistent time reference. The environmental parameter acquisition unit continuously acquires temperature and humidity values ​​at fixed time intervals and generates a corresponding timestamp for each acquired environmental parameter data point. When storing the original audio sequence, the temperature and humidity values ​​at the same timestamp are used as environmental parameter metadata and written into the file header of the original audio sequence or a separate associated database; When reading the original audio sequence, the corresponding temperature and humidity environmental parameters are retrieved from the environmental parameter metadata or associated database by timestamp, and used for environmental factor compensation analysis in the signal processing module or the model training module.

Images