Bird activity monitoring and early warning system based on voiceprint analysis

Abstract

The invention discloses a bird activity monitoring and early warning system based on voiceprint analysis, and relates to the technical field of voiceprint analysis, and the system comprises a signal collection module, a signal training model, a judgment feedback module and an early warning analysis module. The technical key points are that voiceprint signals in a target range are collected based on a monitoring mechanism; pre-processing the voiceprint signal in the target range, extracting logarithmic Mel spectrum features, importing a pre-processed result and the logarithmic Mel spectrum features into a pre-established energy calculation model, and training to obtain an energy index and a bird activity scene category label corresponding to the target range; judging the bird activity state based on the energy index of the target range, and updating the monitoring mechanism according to the bird activity state; extracting warning activities based on the bird activity scene category labels, establishing a rule engine, and generating a first early warning event and a second early warning event; according to the invention, the probability of bird death errors is reduced, and the effectiveness of early warning management is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of voiceprint analysis, and in particular to a bird activity monitoring and early warning system based on voiceprint analysis. Background Art

[0002] Voiceprint recognition is a biometric authentication technology with high application value. Currently, voiceprint recognition technology in the field of bird conservation has achieved basic sound signal collection and species identification functions. By deploying field recording equipment combined with spectrogram analysis (such as MFCC feature extraction), preliminary monitoring of bird populations in specific areas can be carried out. Existing technologies propose deep convolutional neural networks, which can be seen as consisting of two parts: one is the data processing stage. For example, in the process of sound signal processing, there is often a lack of extraction of physically nonlinear acoustic features, resulting in a lack of effective use of time domain representation information when processing voiceprint recognition tasks; the second is the stage of building a classification system. For example, traditional monitoring is prone to signal confusion in complex sound environments (rainfall / strong winds), and the same sound source is repeatedly recorded and stored in different devices. Regional natural enemy warning signals cannot be globally synchronized, which ultimately leads to the inability to establish a correlation between voiceprint signals and survival status (courtship / brooding / alertness), and the inability to reduce bird mortality. Summary of the Invention

[0003] (1) Technical problems solved In response to the shortcomings of the existing technology, the present invention provides a bird activity monitoring and early warning system based on voiceprint analysis, including a signal collection module, a signal training model, a judgment feedback module and an early warning analysis module, to achieve unified and targeted management of voiceprint signals within the target range, reduce the probability of bird death accidents, reflect the effectiveness of early warning management, and solve the problems raised in the background technology.

[0004] (2) Technical solution To achieve the above objectives, the present invention is implemented through the following technical solutions: In a first aspect, the present application provides a bird activity monitoring and early warning system based on voiceprint analysis, the system comprising: Signal collection module: collects voiceprint signals within the target range based on a preset monitoring mechanism; Signal training module: Preprocess the voiceprint signal of the target range and extract the logarithmic Mel spectrum features. The preprocessed voiceprint signal and logarithmic Mel spectrum features are imported into a pre-built energy calculation model to train and obtain the energy index of the target range and the category label of the bird activity scene; among which, the category label of the bird activity scene at least includes alert activity; Judgment and feedback module: judges the bird activity status based on the energy index of the target range and updates the monitoring mechanism according to the bird activity status; Early warning analysis module: Builds a rule engine based on alert activities and generates early warning events.

[0005] Furthermore, the target range includes a minimum range, a maximum range, and an intermediate range. The preprocessing of the voiceprint signal in the target range includes: AGC, noise reduction, and perturbation decomposition. The perturbation decomposition includes: The voiceprint signal within the target range is processed in segments, and each segment of the signal does not overlap with each other. The EMD algorithm is used to process each segment of the voiceprint signal in layers, and the segments of the same layer are spliced together to form a total layered signal. At the same time, the logarithmic Mel spectrum of several layers of signals is calculated.

[0006] Furthermore, the pre-built energy calculation model includes: The energy calculation model is built based on three convolutional layers. The processed results and logarithmic Mel-spectrogram features are imported into the energy calculation model, mapped to a high-dimensional space through a linear layer encoder, and connected to a four-layer S4 block. Through the S4 block, pooling layer, and decoder, it is mapped to several bird activity scene category labels. Each S4 block layer includes an S4 unit, a dropout layer, a residual connection layer, and a normalization layer. The last S4 block layer is connected to a feature weighted calculation unit. The feature weighted calculation unit receives the summary output of all S4 blocks, including the output attention weight, and inversely reconstructs the logarithmic Mel spectrum features extracted from the minimum range, middle range, and maximum range respectively. The inverse reconstruction process includes wavelet transform and obtains Mel cepstral wavelet coefficients; based on the attention feature weight, the Mel cepstral wavelet coefficients are squared in modulus, and weighted summed to obtain the energy index; wherein, the Mel cepstral wavelet coefficients are obtained by normalizing the Mel cepstral coefficients and the power regularization coefficients of the wavelet transform by mean and variance.

[0007] Furthermore, the energy index is obtained, including: Reversely reconstruct the corresponding logarithmic Mel spectrum features within the minimum range, the middle range, and the maximum range to obtain the corresponding first MFCC spectrogram, the second MFCC spectrogram, and the third MFCC spectrogram; at the same time, extract the Mel cepstral wavelet coefficients of each spectrogram; Add a time dimension to the first MFCC spectrogram, the second MFCC spectrogram, and the third MFCC spectrogram to obtain corresponding three-dimensional feature maps, and construct corresponding first feature matrices, second feature matrices, and third feature matrices based on the three-dimensional feature maps; wherein the first feature matrix includes M rows and N columns, and the element value of the i-th row and j-th column represents the j-th Mel-cepstral wavelet coefficient of the minimum range voiceprint signal at the i-th time point, where 1≤i≤M, 1≤j≤N, N is a custom parameter, and the second feature matrix and the third feature matrix have the same representation as the first feature matrix; Inputting the first characteristic matrix and the third characteristic matrix into a characteristic weighting calculation unit, the output result represents the energy index of several positions between the minimum range and the maximum range; The first characteristic matrix and the second characteristic matrix are input into a characteristic weighting calculation unit, and the output result represents the energy index of several positions between the middle range and the minimum range.

[0008] Furthermore, the bird activity status includes unknown bird activity, bird activity, and no bird activity; the detection mechanism includes a query state and a confirmation state.

[0009] Furthermore, the activity status of birds is judged, including: When the bird activity status is determined to be unknown bird activity: If the energy index of the target range is greater than or equal to the preset first threshold, a confirmation signal is sent and the detection mechanism enters the confirmation state; if the energy index of the target range is less than the preset first threshold, a query signal is sent and the detection mechanism enters the query state; When the bird activity status is determined to be bird activity: If the energy index of the target range is less than the preset second threshold, a confirmation signal is sent and the detection mechanism enters the confirmation state; if the energy index of the target range is greater than or equal to the preset second threshold, a query signal is sent and the detection mechanism enters the query state; When the bird activity status is judged as no bird activity: If the energy index of the target range is greater than or equal to the preset third threshold, a confirmation signal is sent and the detection mechanism enters the confirmation state; if the energy index of the target range is less than the preset third threshold, a query signal is sent and the detection mechanism enters the query state; The first threshold is a pre-set energy index for no bird activity, the second threshold is an energy index for the last time a bird flew into the target range, and the third threshold is an energy index for the last time a bird flew out of the target range.

[0010] Furthermore, a rule engine is built based on the alert activity and warning events are generated, including: Count the occurrence times of all alert activities and the corresponding target ranges, and calculate the standard deviation of the time intervals corresponding to the occurrence times of two consecutive alert activities within the same target range, and integrate and mark them as a standard deviation set; assign corresponding weight factors based on the target range, and integrate and mark the weight factors of each target range as a weight set; The corresponding risk ratio is obtained by combining the standard deviation set and the corresponding weight set, and each target range corresponds to a risk ratio, and the risk ratio includes low risk ratio, medium risk ratio and high risk ratio; If a high risk ratio is detected, it is marked as a first warning event; otherwise, it is marked as a second warning event.

[0011] Secondly, this application provides a bird activity monitoring and early warning method based on voiceprint analysis, including: Collect voiceprint signals within the target range based on the preset monitoring mechanism; Preprocess the voiceprint signal of the target range and extract the logarithmic Mel spectrum features. Import the preprocessed voiceprint signal and logarithmic Mel spectrum features into a pre-built energy calculation model to train and obtain the energy index of the target range and the bird activity scene category label; the bird activity scene category label at least includes alert activity; Determine the bird activity status based on the energy index of the target range and update the monitoring mechanism based on the bird activity status; Build a rule engine based on alert activities and generate warning events.

[0012] In a third aspect, the present application provides a computer device, comprising: at least one processor; and A memory storing a computer program executable on the processor, wherein the processor executes the steps of the method according to the second aspect when executing the program.

[0013] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described in the second aspect are performed.

[0014] (3) Beneficial effects The present invention provides a bird activity monitoring and early warning system based on voiceprint analysis, which has the following beneficial effects: The present invention constructs a preset energy calculation model, introduces the energy calculation model into the S4 layer, stacks it to form a deep neural network, and adds a feature weighting layer to accelerate training and improve feature quality. The logarithmic Mel spectrum is used as the input feature of the energy calculation model to train and obtain the energy index corresponding to the target range and the bird activity scene category label. The present invention generates an energy index at any position through the feature matrices of the three key ranges: minimum range, intermediate range, and maximum range. In this process, by fusing the spectral envelope information of MFCC and the time-frequency local characteristics of wavelet transform, the Mel-frequency cepstrum wavelet coefficients are obtained. This not only fully reflects the low-frequency and high-frequency components of the signal, but also effectively utilizes the time domain information, laying the foundation for the subsequent differentiation of bird activity status. The present invention uses different bird activity states, including no bird activity, bird activity, and unknown bird activity, to drive different sampling strategies for the detector, significantly optimizing resource allocation and reducing energy consumption while ensuring sampling accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a module diagram of a bird activity monitoring and early warning system according to an exemplary embodiment; Figure 2 2 is a diagram of a neural network architecture of an energy calculation model according to an exemplary embodiment. DETAILED DESCRIPTION

[0016] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0017] Example 1 The embodiment of the present invention provides a bird activity monitoring and early warning system based on voiceprint analysis; Figure 1 is a module diagram of a bird activity monitoring and early warning system according to an exemplary embodiment; Figure 2 is a neural network architecture diagram of an energy calculation model according to an exemplary embodiment; see Figures 1 to 2 This system integrates the Internet of Things, neural networks, and algorithm optimization to achieve unified and targeted management of voiceprint signals within the target range. It can not only identify voiceprint signals within the target range and detect bird activities, but also issue early warnings based on voiceprint analysis (for example, natural enemy warnings), thereby reducing the probability of bird deaths and demonstrating the effectiveness of early warning management. The system includes: a signal collection module, a signal training module, a judgment feedback module and an early warning analysis module, and the signal collection module, the signal training module, the judgment feedback module and the early warning analysis module are communicatively connected; The following is an explanation of each module: Signal collection module: Used to collect voiceprint signals within a target range and transmit the voiceprint signals to a database through IoT technology; the target range includes a minimum range, a maximum range, and an intermediate range; The bird's nest is detected by setting up a set of detectors: the detectors can be a sound sensor array or a microphone array, including omnidirectional microphones and directional microphones; this article does not limit the selection of detectors; The target range is: If it is an acoustic sensor array, it is deployed at 5 meters, 22.5 meters, and 50 meters; For microphone arrays, omnidirectional microphones should only be deployed within a range of ≤10 meters (e.g., 5 meters), while directional microphones should be deployed within a range of >10 meters (e.g., 22.5 meters, 50 meters). For example, omnidirectional microphones have a wide coverage range (360°) but are susceptible to interference from ambient noise, making them suitable for dense deployment over short distances (<10 meters). Directional microphones focus in a specific direction (e.g., parabolic microphones), and their effective range can be extended to 50 meters through noise suppression algorithms and directional microphones. The detector captures the soundprint signals at equal intervals within a range of 5 to 50 meters from the bird's nest, and obtains several soundprint signals. The range within 5 meters is marked as the minimum range, the range within 50 meters is marked as the maximum range, and the range within 22.5 meters is marked as the middle range; In this application, 5 meters, 22.5 meters and 50 meters are given as examples. The actual distance is given according to the actual situation and will not be described in detail here.

[0018] Signal training module: It is used to pre-process the voiceprint signal of the target range, extract and merge the logarithmic Mel spectrum features, import the pre-processed results and the logarithmic Mel spectrum features into the pre-built energy calculation model, and train to obtain the energy index of the target range and the category label of the bird activity scene; Among them, preprocessing includes: AGC, noise reduction and perturbation decomposition; Utilize AGC and noise reduction processing to adapt to sound changes in different time periods and bird scenes, ensuring the stability and clarity of the voiceprint signal; AGC (Automatic Gain Control): Used to adjust the amplitude of the voiceprint signal. For example, in different bird scenarios, if a bird suddenly approaches (signal enhancement), the AGC will quickly reduce the gain for a certain period of time to avoid partial overload caused by the chirping; if the bird suddenly flies away (signal weakening), the AGC will slowly restore the gain for a certain period of time to prevent frequent gain fluctuations from introducing noise; Noise reduction processing: used to analyze the spectrum of the voiceprint signal, identify and reduce the frequency components that belong to noise interference, and adjust the filter parameters in real time according to the changes in environmental noise interference (such as wind, rain, traffic noise, etc.) to maximize the suppression of noise interference in the voiceprint signal, remove noise interference from the original voiceprint signal, and thus improve the quality and intelligibility of the useful voiceprint signal; Perturbation decomposition: used to analyze the nonlinear characteristics of voiceprint signals. The voiceprint signals within the target range are segmented, and each signal segment does not overlap. Based on the EMD algorithm (empirical mode decomposition), each signal segment is decomposed into IMF components. The signals are layered from high to low frequency, and the signal segments of the same frequency layer are spliced to form a total layered signal. At the same time, the logarithmic Mel spectrum of the signals in several layers is calculated and combined as the input feature of the energy calculation model. Using perturbation decomposition processing, based on the nonlinearity of sound propagation, the soundprint signal is decomposed into the superposition of several linear systems, which facilitates the accuracy of feature extraction; different vibration components in the soundprint signal (such as: snake vibration pattern (alert activity); high-frequency bird singing pattern (courtship activity); high-frequency bird flapping signal pattern (bird alert)) are separated for feature extraction and pattern recognition; Through the above content, the technical effect of collecting and analyzing voiceprint signals in the target range in real time, safely and accurately is achieved, and the control problem of unstable data collection during the data collection process is solved; The energy calculation model includes: The processed results and log-mel spectrum features are imported into an energy calculation model, which is built based on three convolutional layers. The model is mapped to a high-dimensional space (e.g., a 128-dimensional space) via a linear layer encoder and connected to a four-layer S4 block. The model is mapped to several bird activity scene category labels via the S4 block, pooling layer, and decoder. The bird activity scene category labels include at least: alert activity. Each S4 block layer includes an S4 unit, a dropout layer, a residual connection layer, and a normalization layer. The S4 unit introduces an attention mechanism to dynamically assign attention feature weights. Specifically, set the neural network parameters of the energy calculation model: Convolutional layer: First layer: two-dimensional (2D) convolution: 16 filters, convolution kernel size is 7*7; Second layer: two-dimensional (2D) convolution: 16 filters, convolution kernel size is 7*7; two-dimensional maximum pooling: pool size is 5*5; The third layer: two-dimensional (2D) convolution: 32 filters, convolution kernel size is 7*7; 2D maximum pooling: pool size is 4*100; S4 block: S4 unit: introduces the attention mechanism and dynamically allocates attention feature weights; Dropout layer: dropout rate 30%; Residual connection layer: merges input and processed features; Normalization layer: batch normalization to stabilize feature distribution; Fully connected layer: 100 neurons, the activation function is ReLu function; Output layer: Softmax function: The bird activity scene category label with the maximum probability is taken as the classification result; Pooling layer: In addition, bird activity scene category labels also include: territory declaration, migration flight, foraging activities, and nesting communication. The following is an explanation of the designed activity scene category labels: Territorial declaration: high-frequency (3k-6kHz), repetitive, regular calls, often accompanied by long syllable intervals (such as the knocking sounds of woodpeckers); migratory flight: low-frequency continuous vibrations (100-500Hz) interwoven with short chirps (such as the flight vibration signals of nocturnal migration); foraging activities: such as begging for food by young birds: high-frequency (4k-8kHz), short and repetitive chirps (such as the "chirp" sounds of chicks), with low signal-to-noise ratio but concentrated energy; nesting communication: low-frequency friction sounds (such as the sound of biting branches) combined with soft chirps (200Hz-1kHz), showing intermittent pulses in the time domain; alert activities: sudden high-frequency screams (such as the alarm calls of approaching hawks), with a sharp increase in energy and short duration (less than 0.5 seconds); A feature weighting calculation unit is added after the last S4 block. It receives the summary output of all S4 blocks, including the output attention weights, and inversely reconstructs the logarithmic Mel spectrum features extracted from the minimum range, middle range, and maximum range, including wavelet transform, to obtain the Mel cepstral wavelet coefficients. Based on the attention feature weights, the Mel cepstral wavelet coefficients are squared and weighted summed to obtain the energy index. Among them, the Mel cepstral wavelet coefficient is obtained by fusing the Mel cepstral coefficient and the power regularization coefficient of the wavelet transform through mean and variance processing; Specifically, the Mel-frequency cepstral coefficients are obtained during the inverse reconstruction of the logarithmic Mel-frequency spectrum features. Meanwhile, the FFT part in the Mel-frequency cepstral coefficient extraction process is replaced by wavelet transform to obtain the regularized power cepstral coefficients of each layer. The two feature parameters are fused by mean and variance normalization to obtain a new fused feature parameter: Mel-frequency cepstral wavelet coefficients. Statistics fusion based on mean and variance: MFCC statistics: calculate the mean and variance of the Mel-frequency cepstral coefficients; wavelet statistics: calculate the mean and variance of the regularized power of each layer; concatenate the MFCC statistics and wavelet statistics into a joint feature vector to obtain the Mel-frequency cepstral wavelet coefficients; As a long time series, the voiceprint signal needs to consider its long-range dependencies when using convolutional neural networks for learning and representation, and the S4 layer is precisely capable of completing this task. It can appear as a layer in a convolutional neural network and can be stacked to form a deep neural network. The Mel-frequency cepstral wavelet coefficients, by fusing the spectral envelope information of MFCC and the time-frequency local characteristics of the wavelet transform, can fully reflect the low-frequency and high-frequency components of the signal, laying the foundation for subsequent differentiation of bird activity status. Among them, the energy index obtained through training includes: Reversely reconstruct the corresponding logarithmic Mel spectrum features within the minimum range, the middle range, and the maximum range to obtain the corresponding first MFCC spectrogram, the second MFCC spectrogram, and the third MFCC spectrogram; at the same time, extract the Mel cepstral wavelet coefficients of each spectrogram; Add a time dimension to the first MFCC spectrogram, the second MFCC spectrogram, and the third MFCC spectrogram to obtain corresponding three-dimensional feature maps, and construct corresponding first feature matrices, second feature matrices, and third feature matrices based on the three-dimensional feature maps; wherein the first feature matrix includes M rows and N columns, and the element value of the i-th row and j-th column represents the j-th Mel-cepstral wavelet coefficient of the minimum range voiceprint signal at the i-th time point, where 1≤i≤M, 1≤j≤N, N is a custom parameter, and the second feature matrix and the third feature matrix have the same representation as the first feature matrix; Inputting the first characteristic matrix and the third characteristic matrix into a characteristic weighting calculation unit, the output result (value) represents the energy index of several positions between the minimum range and the maximum range; Inputting the first characteristic matrix and the second characteristic matrix into a characteristic weighting calculation unit, the output result (value) represents the energy index of several positions between the middle range and the minimum range; Feature weighted calculation unit: Adds learnable embedding code to each position (for example, assigns position information to each target range and associates it with the feature matrix to prevent the same position from being supervised by any two ranges at the same time) to distinguish the position space hierarchy (for example, the embedding dimension is 32, and the representation of position information is dynamically optimized through training). An attention mechanism is introduced for feature fusion to capture dependencies across position ranges. Masked attention is used to constrain position range interactions (minimum position-maximum position, middle position-minimum position) to accurately separate voiceprint signals at different positions. By combining the feature matrices corresponding to the minimum range (equivalent to the starting point) and the maximum range (equivalent to the end point), position energy covering the entire range is generated. By combining the feature matrices corresponding to the middle range (equivalent to the relay point) and the minimum range (equivalent to the starting point), the position energy for the first half is generated. By fusing the first and third feature matrices, and the second and third feature matrices, the attention weights are dynamically modified to map the weighted features to the target range corresponding to the target position. That is, the feature expression of key positions is enhanced through position-aware weighted fusion. In addition, if all combinations (minimum range and maximum range, minimum range and middle range, middle range and maximum range) are directly calculated, the training will be complicated; Energy index: Based on the attention feature weight, the modulus value of the Mel-frequency cepstral wavelet coefficient is squared and weighted summed to obtain the energy index E of the target range: Where ck represents the kth Mel-cepstral wavelet coefficient, θ k represents the kth attention weight output by the S4 unit, where k is in the range [1, K], and K is the total number of Mel-frequency cepstral wavelet coefficients, which is determined by the length of the signal segment; According to the above steps, energy indices at any position are generated using feature matrices in three key ranges: the minimum range, the middle range, and the maximum range. The feature weighted calculation unit can be considered an independent pre-training layer (or autoencoder), with the first feature matrix, the second feature matrix, and the third feature matrix as input data for the layer. The energy indices at several positions between the minimum range and the maximum range, and the energy indices at several positions between the middle range and the minimum range, are used as pseudo labels for certain input data, and the pseudo labels are used for supervision. The design of this layer accelerates training or improves the quality of features at a specific layer. The energy indices at the middle positions are generated as pseudo labels using an interpolation algorithm. Judgment feedback module: The bird activity status is determined based on the energy index of several locations, and a target range that meets the bird activity status is selected as the object for performing bird activity analysis; The activity status of birds is judged based on the energy index of several locations, including: When the bird activity status is determined to be unknown bird activity: If the energy index of the target range is greater than or equal to the preset first threshold, a confirmation signal is sent and the detection mechanism enters the confirmation state; if the energy index of the target range is less than the preset first threshold, a query signal is sent and the detection mechanism enters the query state; Among them, unknown bird activity means that the system cannot determine whether there is bird activity or no bird activity in the target range; If the system detects that the entropy value of the Mel-frequency cepstral wavelet coefficient of the voiceprint signal within the target range is greater than the preset entropy threshold (e.g., 0.85), high entropy indicates disorder. In the training of the energy calculation model, it will be impossible to match the known bird activity scene category, and the energy index will be between the energy index threshold corresponding to no bird activity and the energy index threshold corresponding to bird activity, avoiding high entropy caused by environmental interference. The calculation of the entropy value is a conventional technique and will not be explained in detail here. When the above conditions (Mel-frequency cepstral wavelet coefficient entropy value > preset entropy value threshold, energy index threshold corresponding to no bird activity and energy index threshold corresponding to bird activity) are met, it indicates unknown bird activity; When the bird activity status is determined to be bird activity: If the energy index of the target range is less than the preset second threshold, a confirmation signal is sent and the detection mechanism enters the confirmation state; if the energy index of the target range is greater than or equal to the preset second threshold, a query signal is sent and the detection mechanism enters the query state; When the bird activity status is judged as no bird activity: If the energy index of the target range is greater than or equal to the preset third threshold, a confirmation signal is sent and the detection mechanism enters the confirmation state; if the energy index of the target range is less than the preset third threshold, a query signal is sent and the detection mechanism enters the query state; The first threshold is the pre-set energy index for no bird activity, the second threshold is the energy index for the last time a bird entered the target range (just when bird activity was detected), and the third threshold is the energy index for the last time a bird flew out of the target range (just when no bird activity was detected). In this embodiment, the detection mechanism reflects the working state of the detector, including the query state and the confirmation state, and is suitable for checking bird activity monitoring and whether the bird activity state has changed. When the bird activity state is the confirmation state, the detector is controlled to start, so that its detection cycle is 32kHz-48kHz sampling, lasting 0.5-2 seconds and the number of cycles is 3 times. At the same time, when entering the confirmation state, it is necessary to judge whether the bird activity state corresponding to the target range has changed 3 times. When all three situations indicate that the bird activity state corresponding to the target range has changed, it can be finally determined that the bird activity state corresponding to the target range has changed. When the bird activity state is the query state, the detector is controlled to start, so that its detection cycle is 24kHz-32kHz sampling, lasting 5-10 seconds. Different sampling strategies are driven by the detector according to the different bird activity states, which significantly optimizes resource allocation and reduces energy consumption while ensuring sampling accuracy. In addition, when monitoring changes in bird activity, sound source positioning can be performed. Aiming at the shortcomings of the traditional MFCC feature extraction process, this paper adds a feature weighting calculation unit on the basis of the full-sequence convolutional neural network, introduces wavelet transform in the feature extraction process, improves the amplitude-frequency characteristics of the filter, and proposes Mel-cepstral wavelet coefficients to fully retain the detailed information in the audio. Finally, the feature and classification purpose are combined to form a weighted full-sequence convolutional neural network. Early warning analysis module: Used to build a rule engine based on alert activities and generate warning events; the warning events include the first warning event and the second warning event; Build a rule engine based on alert activities and generate warning events, including: Count the occurrence time of all alert activities and the corresponding target range, and calculate the standard deviation of the time interval corresponding to the occurrence time of two consecutive alert activities within the same target range, and integrate and mark them as a standard deviation set; Assign corresponding weight factors based on the target range, and integrate the weight factors of each target range into a weight set; It should be noted that the weight factor is determined using the coefficient of variation method. The coefficient of variation method is a method of assigning weights to each indicator based on the degree of variation between the current value of each evaluation indicator and the target value. If the numerical difference of a certain indicator is large, it can clearly distinguish the evaluated objects, indicating that the indicator has rich discrimination information, and thus the indicator should be given a larger weight. Conversely, if the numerical difference of each evaluated object on a certain indicator is small, then the ability of this indicator to distinguish the evaluated objects is weak, and thus the indicator should be given a smaller weight. This method directly uses the information contained in each indicator to obtain the indicator weight through calculation, and therefore is objective. The corresponding risk ratio is obtained by combining the standard deviation set and the corresponding weight set, and each target range corresponds to a risk ratio, and the risk ratio includes low risk ratio, medium risk ratio and high risk ratio; Among them, the risk ratio of the i-th target range is for: Where, Indicates the weight assigned to the target range i, is the standard deviation of the time interval between two alert activities in target range i; Each target range corresponds to a risk ratio, including: comparing the risk ratio with a preset risk interval: if the risk ratio is less than the preset risk interval, marking it as a low risk ratio; if the risk ratio is within the preset risk interval, marking it as a medium risk ratio; if the risk ratio exceeds the preset risk interval, marking it as a high risk ratio; Specifically, the preset risk interval is set based on the historical risk ratio: the risk ratios corresponding to the low risk ratio and the medium risk ratio are obtained, the average value and the standard deviation are calculated, and the sum of the average value and 2 times the standard deviation is taken as the minimum value of the risk interval, and the sum of the average value and 3 times the standard deviation is taken as the maximum value of the risk interval; If a high risk ratio is detected, it will be marked as a first warning event; Otherwise, it is marked as the second warning event; According to the detection of the first warning event and the second warning event, the warning supervision strategies corresponding to the two types of events are executed. The warning supervision strategies are set according to the actual situation and will not be described in detail here.

[0019] Example 2 Embodiment 2 of the present invention provides a bird activity monitoring and early warning method based on voiceprint analysis, comprising the following steps: Collect voiceprint signals within the target range based on the preset monitoring mechanism; Preprocess the voiceprint signal of the target range and extract the logarithmic Mel spectrum features. Import the preprocessed voiceprint signal and logarithmic Mel spectrum features into a pre-built energy calculation model to train and obtain the energy index of the target range and the bird activity scene category label; the bird activity scene category label at least includes alert activity; Determine the bird activity status based on the energy index of the target range and update the monitoring mechanism based on the bird activity status; Build a rule engine based on alert activities and generate warning events.

[0020] Example 3 Embodiment 3 of the present invention provides a computer device, including: at least one processor; and A memory storing a computer program executable on the processor, wherein the processor executes the steps of the method described in embodiment 2 when executing the program.

[0021] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the steps of the method described in embodiment 2 are performed.

[0022] The readable storage medium may include, but is not limited to, a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0023] In a possible implementation manner, the present invention can also be implemented in the form of a program product, which includes a program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the bird activity monitoring and early warning method based on voiceprint analysis in Example 2.

[0024] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0025] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0026] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. Bird activity monitoring and early warning system based on voiceprint analysis, characterized by: The system comprises: Signal collection module: collects voiceprint signals within the target range based on a preset monitoring mechanism; Signal training module: pre-processes the voiceprint signal of the target range and extracts the logarithmic Mel spectrum features. The pre-processed voiceprint signal and logarithmic Mel spectrum features are imported into a pre-built energy calculation model to train and obtain the energy index of the target range and the bird activity scene category label; wherein the bird activity scene category label at least includes alert activity; Judgment and feedback module: judging the bird activity status based on the energy index of the target range, and updating the monitoring mechanism according to the bird activity status; Early warning analysis module: builds a rule engine based on the alert activities and generates early warning events.

2. The bird activity monitoring and early warning system based on voiceprint analysis according to claim 1 is characterized in that: The target range includes a minimum range, a maximum range, and an intermediate range. The preprocessing of the voiceprint signal in the target range includes: AGC, noise reduction, and perturbation decomposition. The perturbation decomposition includes: The voiceprint signal within the target range is processed in segments, and each segment of the signal does not overlap with each other. The EMD algorithm is used to process each segment of the voiceprint signal in layers, and the segments of the same layer are spliced together to form a total layered signal. At the same time, the logarithmic Mel spectrum of several layers of signals is calculated.

3. The bird activity monitoring and early warning system based on voiceprint analysis according to claim 2 is characterized in that: The pre-built energy calculation model includes: The energy calculation model is built based on three convolutional layers. The processed results and logarithmic Mel-spectrogram features are imported into the energy calculation model, mapped to a high-dimensional space through a linear layer encoder, and connected to a four-layer S4 block. Through the S4 block, pooling layer, and decoder, it is mapped to several bird activity scene category labels. Each S4 block layer includes an S4 unit, a dropout layer, a residual connection layer, and a normalization layer. The last S4 block layer is connected to a feature weighted calculation unit. The feature weighted calculation unit receives the summary output of all S4 blocks, including the output attention weights, and inversely reconstructs the logarithmic Mel spectrum features extracted from the minimum range, middle range, and maximum range respectively. The inverse reconstruction process includes wavelet transform and obtains Mel cepstral wavelet coefficients; Based on the attention feature weight, the modulus value of the Mel-cepstral wavelet coefficient is squared and weighted summed to obtain the energy index; among them, the Mel-cepstral wavelet coefficient is obtained by normalizing the Mel-cepstral coefficient and the power regularization coefficient of the wavelet transform by mean and variance.

4. The bird activity monitoring and early warning system based on voiceprint analysis according to claim 3 is characterized in that: The energy index obtained includes: Reversely reconstruct the corresponding logarithmic Mel spectrum features within the minimum range, the middle range, and the maximum range to obtain the corresponding first MFCC spectrogram, the second MFCC spectrogram, and the third MFCC spectrogram; at the same time, extract the Mel cepstral wavelet coefficients of each spectrogram; Add a time dimension to the first MFCC spectrogram, the second MFCC spectrogram, and the third MFCC spectrogram to obtain corresponding three-dimensional feature maps, and construct corresponding first feature matrices, second feature matrices, and third feature matrices based on the three-dimensional feature maps; wherein the first feature matrix includes M rows and N columns, and the element value of the i-th row and j-th column represents the j-th Mel-cepstral wavelet coefficient of the minimum range voiceprint signal at the i-th time point, where 1≤i≤M, 1≤j≤N, N is a custom parameter, and the second feature matrix and the third feature matrix have the same representation as the first feature matrix; Inputting the first characteristic matrix and the third characteristic matrix into a characteristic weighting calculation unit, the output result represents the energy index of several positions between the minimum range and the maximum range; The first characteristic matrix and the second characteristic matrix are input into a characteristic weighting calculation unit, and the output result represents the energy index of several positions between the middle range and the minimum range.

5. The bird activity monitoring and early warning system based on voiceprint analysis according to claim 1 is characterized in that: The bird activity status includes unknown bird activity, bird activity, and no bird activity; the detection mechanism includes a query state and a confirmation state.

6. The bird activity monitoring and early warning system based on voiceprint analysis according to claim 5 is characterized in that: Determining the activity status of the bird includes: When the bird activity status is determined to be unknown bird activity: If the energy index of the target range is greater than or equal to the preset first threshold, a confirmation signal is sent and the detection mechanism enters the confirmation state; if the energy index of the target range is less than the preset first threshold, a query signal is sent and the detection mechanism enters the query state; When the bird activity status is determined to be bird activity: If the energy index of the target range is less than the preset second threshold, a confirmation signal is sent and the detection mechanism enters the confirmation state; if the energy index of the target range is greater than or equal to the preset second threshold, a query signal is sent and the detection mechanism enters the query state; When the bird activity status is judged as no bird activity: If the energy index of the target range is greater than or equal to the preset third threshold, a confirmation signal is sent and the detection mechanism enters the confirmation state; if the energy index of the target range is less than the preset third threshold, a query signal is sent and the detection mechanism enters the query state; The first threshold is a pre-set energy index for no bird activity, the second threshold is an energy index for the last time a bird flew into the target range, and the third threshold is an energy index for the last time a bird flew out of the target range.

7. The bird activity monitoring and early warning system based on voiceprint analysis according to claim 1 is characterized in that: The step of building a rule engine based on the alert activity and generating an early warning event includes: Count the occurrence times of all alert activities and the corresponding target ranges, and calculate the standard deviation of the time intervals corresponding to the occurrence times of two consecutive alert activities within the same target range, and integrate and mark them as a standard deviation set; assign corresponding weight factors based on the target range, and integrate and mark the weight factors of each target range as a weight set; The corresponding risk ratio is obtained by combining the standard deviation set and the corresponding weight set, and each target range corresponds to a risk ratio, and the risk ratio includes low risk ratio, medium risk ratio and high risk ratio; If a high risk ratio is detected, it is marked as a first warning event; otherwise, it is marked as a second warning event.

8. Bird activity monitoring and early warning method based on voiceprint analysis, characterized in that: include: Collect voiceprint signals within the target range based on the preset monitoring mechanism; Preprocessing the voiceprint signal of the target range and extracting the logarithmic Mel spectrum features, importing the preprocessed voiceprint signal and the logarithmic Mel spectrum features into a pre-built energy calculation model, and training to obtain the energy index of the target range and the bird activity scene category label; wherein the bird activity scene category label at least includes alert activity; determining the bird activity status based on the energy index of the target range, and updating the monitoring mechanism according to the bird activity status; A rule engine is built based on the alert activities and warning events are generated.

9. A computer device comprising: at least one processor; as well as A memory storing a computer program that can be run on the processor, wherein the processor performs the steps of the method according to claim 8 when executing the program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein the computer program performs the steps of the method according to claim 8 when executed by a processor.

Images