Bird buzzing intelligent identification method and system based on ensemble learning

Abstract

The invention provides an integrated learning-based bird buzzing intelligent recognition method and system, and solves the problem that the recognition precision is insufficient when an existing single deep learning model processes high-noise and low-signal-to-noise-ratio audio data. In order to solve the problems that bird species with limited long-tail data distribution or limited data volume are particularly unsatisfactory, the method mainly comprises the following steps: S1, collecting and preprocessing bird buzzing data to construct a standardized bird buzzing database; s2, using a Bootstrap sampling scheme to perform random sampling with playback on audio samples in the standardized twitter database, and generating a plurality of different independent sub-data sets; s3, respectively constructing and training an independent deep learning model for each sub-data set, including model configuration, data preparation, model training and model tuning; and S4, based on a Bagging ensemble learning strategy, fusing output results of the plurality of models to generate a final prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of audio signal recognition, and in particular to a method and system for intelligently recognizing bird calls based on ensemble learning. Background Art

[0002] Birds are an important part of the ecosystem, and their population changes and behavioral characteristics are important criteria for assessing ecological health. Accurate identification of bird species is of great significance for ecological protection, biodiversity research, and environmental monitoring. Compared with visual recognition methods, bird song-based sound recognition technology has the advantages of non-invasiveness, wide applicability, and all-weather monitoring, and is particularly suitable for environments with dense vegetation or limited vision. Therefore, the development of efficient and intelligent bird song recognition methods has far-reaching significance for promoting ecological research and protection.

[0003] Traditional bird identification methods rely on manual observation and recording, which is time-consuming, labor-intensive and prone to errors. The use of advanced audio processing and classification and recognition algorithms can significantly improve the accuracy and efficiency of bird identification. At present, deep learning models such as convolutional neural networks and recurrent neural networks have been widely used in bird song recognition due to their powerful feature extraction capabilities and adaptability to complex data. These methods convert bird song signals into feature images (such as time domain waveforms, Mel-speech spectrograms, etc.), and then use deep learning models to extract the time-frequency features of bird songs from them, capturing the vocal differences of different species, thereby achieving automatic classification and recognition.

[0004] Although existing technologies have made significant progress, bird song recognition still faces many challenges. On the one hand, bird song data itself is complex and diverse: bird calls are more complex and changeable, and the voiceprints of different species or individuals vary significantly. At the same time, audio signals collected in natural environments are often accompanied by a variety of noises (such as wind, rain, other organisms and human activities), making feature extraction and classification tasks more difficult. On the other hand, a single deep learning model often has insufficient recognition accuracy when processing high-noise, low-signal-to-noise ratio audio data; it is particularly unsatisfactory for bird species with long-tail data distribution or limited data volume. The corresponding limitations are as follows: 1. Existing methods usually directly use a single dataset for model training, which easily leads to overfitting of the model to a specific pattern and poor generalization ability; 2. Existing methods usually rely on a single model structure and cannot fully utilize the characteristics and advantages of different deep learning models; 3. Existing methods usually adopt a simple averaging integration strategy, which limits the diversity and flexibility of model combinations and fails to fully utilize the complementary advantages of various models.

[0005] To this end, we propose an intelligent bird song recognition method and system based on ensemble learning to solve the above problems. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the defects of the prior art. The present invention proposes a method and system for intelligent recognition of bird calls based on ensemble learning. The method adopts the Bagging ensemble learning strategy, integrates the advantages of multiple models, and enhances the robustness and generalization ability of the recognition system. In particular, it shows high recognition accuracy and reliability when facing complex background noise and rare bird data.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: an intelligent bird song recognition method based on ensemble learning, comprising the following steps: S1. Collect and preprocess bird song data to build a standardized bird song database; S2, using the Bootstrap sampling scheme, randomly sample the audio samples in the standardized bird song database with replacement to generate multiple different independent sub-datasets; S3. For each sub-dataset, build and train an independent deep learning model, including model configuration, data preparation, model training, and model tuning; S4. Based on the Bagging ensemble learning strategy, the output results of multiple models are integrated to generate the final prediction results.

[0008] Furthermore, the data collection preprocessing process in step S1 is: According to the distribution of birds, bird species in the target area are selected as research objects, a list of target species is formed, and a corresponding species name comparison table is constructed; Download the target bird's call data from the public bird sound resource library to form an original bird song database; The bird song data in the database are cleaned and format converted, and then different audio files are adjusted to audio with unified parameters through data resampling, length trimming, intensity normalization and other means.

[0009] Furthermore, the specific process of step S2 is as follows: All the bird song audios in the standardized bird song database are used as the original data set; Set sampling parameters to determine the number of sub-datasets and the sample size of each sub-dataset. The sample size of each sub-dataset is the same as that of the original data set. A sample is randomly selected from the original data set and placed in the current sub-data set, and then the sample is placed back into the original data set so that it may be selected again in subsequent sampling. This process is repeated until the current sub-data set reaches a predetermined size, and the predetermined number of sub-data sets is generated.

[0010] Furthermore, the specific steps of model configuration in step S3 are as follows: Select three deep learning model architectures: ConvNext, ConvNextV2, and ECA-NFNet as the basic models; Initialize the model using pre-trained weights obtained from a public platform; According to the actual classification task requirements, adjust the output layer size of the model to make it consistent with the number of target classification categories; add a Softmax layer to generate category probability distribution to adapt to multi-classification tasks; PaddedCMAP is used as the loss function to adapt to the unbalanced characteristics of bird song data. Use the Adam optimizer in combination with the cosine annealing scheduler to achieve dynamic learning rate adjustment.

[0011] Furthermore, the specific steps of data preparation in step S3 are as follows: According to the sample size, choose random partitioning or K-fold cross-validation partitioning to obtain the training set and test set; Add background noise of different intensities and types to the audio samples in the training set and pre-process the background noise. Set the noise intensity range simultaneously. For each bird song audio sample, randomly select a noise signal and a relative noise intensity to multiply to obtain an adjusted noise signal. Align and superimpose the adjusted noise signal with the bird song signal. Normalize the mixed signal to obtain a new audio sample. The time series audio data in the training set and validation set are converted into Mel-spectrograms as model input features.

[0012] Furthermore, the specific steps of time-series audio data conversion are as follows: The audio signal is pre-emphasized and the processing process is expressed by a difference equation: ,in is the current sample of the input signal, is the previous sample, is the current sample of the output signal, is the pre-emphasis coefficient, ranging from 0.9 to 1.0; The audio signal is divided into short frames and a window function is applied to each frame. Then, the data of each frame is fast Fourier transformed to convert it from the time domain to the frequency domain to obtain the spectrum data. The amplitude square of each frequency component in the spectrum is synchronously calculated to obtain the power spectrum of the signal. The power spectrum is filtered using a set of Mel filters covering the entire audible frequency range, the output of the filter bank is logarithmically compressed to form a Mel spectrogram, and the generated spectrogram is normalized.

[0013] Furthermore, the specific steps of model training in step S3 are as follows: According to the complexity of the task and the size of the dataset, set the maximum number of training rounds and batch size. Start the current round, load the normalized Mel-spectrogram, randomly divide the training set into multiple batches according to the set batch size, input the training data of the current batch into the model, the model calculates each layer step by step, outputs the predicted category probability distribution, and simultaneously calculates the error between the model prediction result and the true label according to the set loss function, calculates the gradient of the loss value to the model parameters through the back-propagation mechanism, and uses the optimizer to adjust the model weight according to the learning rate and gradient information; Traverse each batch of training data, repeat until all batches of data in the current round are trained, save the current model weight ratio, and use the validation set to evaluate the model performance; After the model is verified, it is saved and iteratively trained until the early stopping condition is met or the set maximum number of training rounds is reached. The early stopping condition is to take the verification loss value as the basis for judging the model performance, set the minimum improvement threshold and patience parameter, and if the improvement in model performance is lower than the minimum threshold in n consecutive rounds, the early stopping mechanism is triggered to terminate the training process early.

[0014] Furthermore, the specific steps of model tuning in step S3 are as follows: Define the tuning space and determine the hyperparameters that need to be adjusted and their ranges, including learning rate, batch size, regularization coefficient, and Dropout ratio; Bayesian optimization is used as a hyperparameter tuning method. The performance of different hyperparameter combinations is evaluated through multiple trainings, the effect of each set of hyperparameter configurations is verified, and the search space is gradually adjusted to determine the optimal hyperparameter combination. The optimal hyperparameter combination is selected based on the tuning results and final training is performed. After the training is completed, the optimal weights of each model are saved separately.

[0015] Furthermore, the specific steps of step S4 are as follows: By calling multiple models to predict the input data separately, we can get the predicted output of each model. , where N is the number of models, indicating the prediction result of the i-th model; Select the appropriate integration strategy according to specific needs. The Gaussian distribution weighted method is used by default. The Gaussian distribution weighted method is used to calculate the Gaussian distribution weight of each model according to the prediction error, and then the prediction results of each model are weighted averaged to obtain the final prediction result. The specific calculation formula is: , in, is the prediction result of the ith model, is the Gaussian distribution weight of the model, N is the number of models, is the prediction error of the ith model, is the standard deviation of all model prediction errors; The fused results are used as the final prediction results, and the top 5 categories and their probabilities are output.

[0016] The system constructed according to the above-mentioned bird song intelligent recognition method based on ensemble learning includes: The data collection and preprocessing module is used to clean, convert and standardize the collected bird song data to establish a standardized bird song database; The sub-dataset generation module is used to randomly sample the standardized bird song database using the Bootstrap sampling scheme based on the Bagging ensemble learning strategy to generate multiple different independent sub-datasets; The model training module is used to build and train an independent deep learning model for each sub-dataset for subsequent integration of prediction results. It specifically includes model configuration, data preparation, model training, and model tuning steps.

[0017] The result integration module is used to fuse the output results of multiple models based on the Bagging ensemble learning strategy to generate the final prediction results.

[0018] Compared with the prior art, the beneficial effects of the present invention include: 1. The present invention provides a bird voiceprint recognition system based on deep learning, which significantly improves the accuracy and timeliness of bird voiceprint recognition. The model recognition accuracy rate reaches 96%, and the prediction speed exceeds 10 audio / second; 2. The present invention provides a random sampling strategy based on Bootstrap sampling to generate multiple different sub-data sets, introduce data diversity, and enhance the generalization ability of the model and the stability of the recognition system; 3. The present invention provides a multi-model integration strategy based on Gaussian distribution weighting, which integrates multiple independent deep learning models, fully utilizes the advantages and complementarity of each model, and improves the accuracy of the recognition system. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the accompanying drawings, the same reference numerals are used to refer to the same components. Among them: Figure 1 A schematic diagram of a recognition method flow chart according to an embodiment of the present invention is shown; Figure 2The specific implementation effect diagram of the identification method proposed according to one embodiment of the present invention is schematically shown. DETAILED DESCRIPTION

[0020] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the whole of the present invention or as a limitation or restriction to the technical solution of the present invention.

[0021] According to one embodiment of the present invention, Figure 1-Figure 2 Shown.

[0022] like Figure 1 As shown in the figure, for the overall process, the bird song intelligent recognition method based on ensemble learning includes the following steps: S1. Collect and preprocess bird song data to build a standardized bird song database; S2, using the Bootstrap sampling scheme, randomly sample the audio samples in the standardized bird song database with replacement to generate multiple different independent sub-datasets; S3. For each sub-dataset, build and train an independent deep learning model, including model configuration, data preparation, model training, and model tuning; S4. Based on the Bagging ensemble learning strategy, the output results of multiple models are integrated to generate the final prediction results.

[0023] The following is a detailed description of the above steps: 1. Data collection and preprocessing module 1.1 Data Collection According to the distribution of birds, bird species in the target area are selected as research objects, a list of target species is formed, and a species name comparison table of "Latin name-Chinese name-English name" is constructed. The call data of target birds are downloaded from public bird sound resource libraries (such as the xeno-conto website and the bird-knowing app) to form an original bird call database.

[0024] 1.2 Data Preprocessing Due to different data sources, bird song data has great differences in audio length, sampling rate, encoding format, etc., and needs to be properly preprocessed. Preprocessing includes the following parts: 1) Data cleaning: Set the screening thresholds for audio length, sampling rate, file size and other parameters to remove low-quality audio samples.

[0025] 2) Format conversion: Convert audio files in different encoding formats to a unified encoding format (such as MP3 format).

[0026] 3) Data standardization: Through data resampling, length trimming, intensity normalization and other means, different audio files are adjusted to audio with unified parameters such as sampling rate, length, intensity, and number of channels.

[0027] After the above-mentioned preprocessing steps such as data cleaning, conversion and standardization, a standardized bird song database with accurate data and complete information is established.

[0028] 2. Sub-dataset generation module Based on the Bagging ensemble learning strategy, the Bootstrap sampling scheme is used to randomly sample the standardized bird song database to generate multiple different independent sub-datasets. The specific steps are as follows: 1) Initialize the sample set. All the bird song audio in the standardized bird song database is used as the original data set.

[0029] 2) Set sampling parameters. Determine the number of sub-datasets and the sample size of each sub-dataset. The sample size of each sub-dataset is generally the same as that of the original data set.

[0030] 3) Perform random sampling with replacement. Randomly select a sample from the original data set and put it into the current sub-data set, and then put the sample back into the original data set so that it may be selected again in subsequent sampling. Repeat this process until the current sub-data set reaches the predetermined size.

[0031] 4) Generate multiple sub-datasets. Repeat step 3 until a predetermined number of sub-datasets are generated.

[0032] 3. Model training module For each sub-dataset, an independent deep learning model is built and trained for subsequent integration of prediction results, including model configuration, data preparation, model training, model tuning and other steps.

[0033] 3.1 Model Configuration 1) Model architecture selection. Three deep learning model architectures, ConvNext, ConvNextV 2, and ECA-NFNet, were selected as the basic models. The choice of model architecture is flexible, and other architectures (such as ViT, ResNet, EfficientNet, etc.) can be expanded according to needs to adapt to different research scenarios and data characteristics.

[0034] 2) Pre-trained weight loading. Use pre-trained weights obtained from public platforms (such as Hugging Face or TensorFlow Hub) to initialize the model. These weights trained based on large data sets can significantly improve the convergence speed and initial performance of the model. Supports the use of random initialization strategies to directly generate random weights.

[0035] 3) Customized model adjustment. According to the actual classification task requirements, adjust the output layer size of the model to make it consistent with the number of target classification categories. Add a Softmax layer to generate category probability distribution to adapt to multi-classification tasks. Support adding a Dropout layer or regularization strategy to enhance the robustness of the model.

[0036] 4) Loss function setting. PaddedCMAP is used as the loss function by default to adapt to the unbalanced characteristics of bird song data. Other loss functions are supported, such as cross entropy loss or focal loss.

[0037] 5) Optimizer settings. The default Adam optimizer is used, combined with the cosine annealing scheduler to achieve dynamic learning rate adjustment. Other optimizers are supported, such as AdamW, SGD, or RMSprop.

[0038] 3.2 Data Preparation 3.2.1 Data Division Depending on the sample size, choose random partitioning or K-fold cross-validation partitioning mode. Random partitioning mode: randomly shuffle the data set and then divide it into training set and test set according to a predetermined ratio. Cross-validation mode: randomly divide the data set into K subsets of equal size. In each iteration, select one subset as the test set, and combine the remaining K-1 subsets as the training set.

[0039] 3.2.2 Training Data Augmentation Add background noise of different intensities and types to the audio samples in the training set to improve the model's adaptability to noise and changes in real scenes. The specific steps are as follows: 1) Noise data acquisition: Collect background noises in various real scenes from public noise databases, such as environmental sounds, wind sounds, rain sounds, and insect sounds.

[0040] 2) Noise data preprocessing: The collected background noise is resampled, cropped, normalized, and other operations are performed to match the duration and sampling rate of the bird song audio samples.

[0041] 3) Set the noise intensity range. According to the degree of noise impact on bird song signals in actual scenarios, set a reasonable relative noise intensity range to simulate the degree of noise interference in different environments and ensure that the main features of bird song signals are not obscured.

[0042] 4) Superimpose noise signals. For each bird song audio sample, randomly select a noise signal and a relative noise intensity, and multiply the noise signal by the relative intensity to obtain an adjusted noise signal. Align and superimpose the adjusted noise signal with the bird song signal, and normalize the mixed signal to obtain a new audio sample.

[0043] 3.2.3 Mel spectrum conversion The time series audio data in the training set and validation set are converted into Mel Spectrograms as model input features. The specific conversion process is as follows: 1) Audio signal preprocessing. Perform pre-emphasis on the audio signal to enhance the high-frequency components. The processing process can be expressed by a differential equation: ,in is the current sample of the input signal, is the previous sample, is the current sample of the output signal, It is the pre-emphasis coefficient, which generally ranges from 0.9 to 1.0.

[0044] 2) Framing and windowing. Divide the audio signal into short frames with some overlap between frames. Apply a windowing function (such as a Hamming window) to each frame to reduce discontinuities at the frame boundaries.

[0045] 3) Fast Fourier Transform (FFT): Perform a fast Fourier transform on each frame of data to convert it from the time domain to the frequency domain to obtain spectrum data.

[0046] 4) Power spectrum calculation: Calculate the square of the amplitude of each frequency component in the spectrum to obtain the power spectrum of the signal.

[0047] 5) Apply Mel filters. Filter the power spectrum using a set of Mel filters (such as triangular filters) that cover the entire audible frequency range.

[0048] 6) Logarithmic compression: Logarithmic compression is performed on the output of the filter bank to form a Mel-spectrogram.

[0049] 7) Spectrum normalization: Normalize the generated spectrum graph to improve the generalization ability of the model.

[0050] 3.3 Model Training 1) Training parameter setting: According to the complexity of the task and the size of the dataset, set the maximum training round (epoch) and batch size and other parameters.

[0051] 2) Start the current round. Load the training data (normalized Mel-spectrogram) and randomly divide the training set into multiple batches according to the set batch size.

[0052] 3) Data input and forward propagation: The current batch of training data is input into the model, and the model calculates each layer step by step and outputs the predicted category probability distribution.

[0053] 4) Loss calculation: According to the set loss function (PaddedCMAP), the error between the model prediction result and the true label is calculated.

[0054] 5) Back propagation and gradient update: The gradient of the loss value to the model parameters is calculated through the back propagation mechanism, and the optimizer (Adam) is used to adjust the model weights according to the learning rate and gradient information.

[0055] 6) Complete the current round. Traverse each batch of training data and repeat steps 3 to 5 until the training of all batches of data is completed.

[0056] 7) Model verification. Input the verification data into the model after this round of training, and calculate the category probability distribution predicted by the model, the verification loss value (PaddedCMAP), and other evaluation indicators (accuracy, F1 score, Focal Loss, etc.).

[0057] 8) Model saving. Record key information such as the training loss value, validation loss value, and evaluation index value of the current round. If the validation index value exceeds the historical optimal level, save the weight of the current model.

[0058] 9) Early stopping condition judgment. Select the validation loss value (or other evaluation index value) as the basis for judging the model performance, set the minimum improvement threshold (Δ=0.01) and patience parameter (n=10). If the improvement of model performance is lower than the minimum threshold in n consecutive rounds, the early stopping mechanism is triggered and the training process is terminated early.

[0059] 10) Iterative training. Repeat steps 2 to 9 until the early stopping condition is met or the maximum number of training rounds is reached.

[0060] 3.4 Model Tuning 1) Define the tuning space. Determine the hyperparameters that need to be adjusted and their ranges, including learning rate, batch size, regularization coefficient, Dropout ratio, etc.

[0061] 2) Select a tuning method. By default, Bayesian optimization is used as the hyperparameter tuning method to improve search efficiency and convergence speed. Other tuning methods, such as random search or grid search, are supported to meet different tuning requirements and computing resource conditions.

[0062] 3) Tuning process: Evaluate the performance of different hyperparameter combinations through multiple trainings, verify the effect of each set of hyperparameter configurations, gradually adjust the search space, and determine the optimal hyperparameter combination.

[0063] 4) Model selection. Select the optimal hyperparameter combination based on the tuning results and perform final training. After training is completed, save the optimal weights of each model separately.

[0064] 4. Result Integration Module Based on the Bagging ensemble learning strategy, the output results of multiple models are integrated to generate the final prediction results. The specific steps are as follows: 1) Obtain the prediction results of each model. By calling multiple models to predict the input data separately, obtain the prediction output of each model , where N is the number of models, Represents the prediction result of the i-th model.

[0065] 2) Select an integration strategy. Select an appropriate integration strategy based on specific needs. The Gaussian distribution weighted method is used by default. Other integration strategies are supported, such as simple average method, voting method, stacking method, etc.

[0066] 3) Integrate the prediction results of each model. The Gaussian distribution weighted method calculates the Gaussian distribution weight of each model based on its prediction error, and then performs weighted average of the prediction results of each model to obtain the final prediction result. The specific calculation formula is: , in, is the prediction result of the ith model, is the Gaussian distribution weight of the model, N is the number of models, is the prediction error of the ith model, is the standard deviation of all model prediction errors.

[0067] 4) Output the final prediction result. The fused result is used as the final prediction result, and the top 5 categories and their probabilities are output.

[0068] Similarly, the system constructed according to the above method steps is also within the protection scope of the present invention. Specifically, the system may include: The data collection and preprocessing module is used to clean, convert and standardize the collected bird song data to establish a standardized bird song database; The sub-dataset generation module is used to randomly sample the standardized bird song database using the Bootstrap sampling scheme based on the Bagging ensemble learning strategy to generate multiple different independent sub-datasets; The model training module is used to build and train an independent deep learning model for each sub-dataset for subsequent integration of prediction results. It specifically includes model configuration, data preparation, model training, and model tuning steps.

[0069] The result integration module is used to fuse the output results of multiple models based on the Bagging ensemble learning strategy to generate the final prediction results.

[0070] In addition, the above-mentioned method steps and system can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0071] The technical scope of the present invention is not limited to the contents in the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. An intelligent bird song recognition method based on ensemble learning, characterized in that: The steps include: S1. Collect and preprocess bird song data to build a standardized bird song database; S2, using the Bootstrap sampling scheme, randomly sample the audio samples in the standardized bird song database with replacement to generate multiple different independent sub-datasets; S3. For each sub-dataset, build and train an independent deep learning model, including model configuration, data preparation, model training, and model tuning; S4. Based on the Bagging ensemble learning strategy, the output results of multiple models are integrated to generate the final prediction results.

2. The method for intelligently identifying bird calls based on ensemble learning according to claim 1, characterized in that: The data collection preprocessing process in step S1 is: According to the distribution of birds, bird species in the target area are selected as research objects, a list of target species is formed, and a corresponding species name comparison table is constructed; Download the target bird's call data from the public bird sound resource library to form an original bird song database; The bird song data in the database are cleaned and format converted, and then different audio files are adjusted to audio with unified parameters through data resampling, length trimming, intensity normalization and other means.

3. The method for intelligently identifying bird calls based on ensemble learning according to claim 1, characterized in that: The specific process of step S2 is as follows: All the bird song audios in the standardized bird song database are used as the original data set; Set sampling parameters to determine the number of sub-datasets and the sample size of each sub-dataset. The sample size of each sub-dataset is the same as that of the original data set. A sample is randomly selected from the original data set and placed in the current sub-data set, and then the sample is placed back into the original data set so that it may be selected again in subsequent sampling. This process is repeated until the current sub-data set reaches a predetermined size, and the predetermined number of sub-data sets is generated.

4. The method for intelligently identifying bird calls based on ensemble learning according to claim 1, characterized in that: The specific steps of model configuration in step S3 are as follows: Select three deep learning model architectures: ConvNext, ConvNextV2, and ECA-NFNet as the basic models; Initialize the model using pre-trained weights obtained from a public platform; According to the actual classification task requirements, adjust the output layer size of the model to make it consistent with the number of target classification categories; add a Softmax layer to generate category probability distribution to adapt to multi-classification tasks; PaddedCMAP is used as the loss function to adapt to the unbalanced characteristics of bird song data. Use the Adam optimizer in combination with the cosine annealing scheduler to achieve dynamic learning rate adjustment.

5. The method for intelligently identifying bird calls based on ensemble learning according to claim 1, characterized in that: The specific steps of data preparation in step S3 are as follows: According to the sample size, choose random partitioning or K-fold cross-validation partitioning to obtain the training set and test set; Add background noise of different intensities and types to the audio samples in the training set and pre-process the background noise. Set the noise intensity range simultaneously. For each bird song audio sample, randomly select a noise signal and a relative noise intensity to multiply to obtain an adjusted noise signal. Align and superimpose the adjusted noise signal with the bird song signal. Normalize the mixed signal to obtain a new audio sample. The time series audio data in the training set and validation set are converted into Mel-spectrograms as model input features.

6. The method for intelligently identifying bird calls based on ensemble learning according to claim 5, characterized in that: The specific steps for timing audio data conversion are as follows: The audio signal is pre-emphasized and the processing process is expressed by a difference equation: ,in is the current sample of the input signal, is the previous sample, is the current sample of the output signal, is the pre-emphasis coefficient, ranging from 0.9 to 1.0; The audio signal is divided into short frames and a window function is applied to each frame. Then, the data of each frame is fast Fourier transformed to convert it from the time domain to the frequency domain to obtain the spectrum data. The amplitude square of each frequency component in the spectrum is synchronously calculated to obtain the power spectrum of the signal. The power spectrum is filtered using a set of Mel filters covering the entire audible frequency range, the output of the filter bank is logarithmically compressed to form a Mel spectrogram, and the generated spectrogram is normalized.

7. The method for intelligently identifying bird calls based on ensemble learning according to claim 6, characterized in that: The specific steps of model training in step S3 are as follows: According to the complexity of the task and the size of the dataset, set the maximum number of training rounds and batch size. Start the current round, load the normalized Mel-spectrogram, randomly divide the training set into multiple batches according to the set batch size, input the training data of the current batch into the model, the model calculates each layer step by step, outputs the predicted category probability distribution, and simultaneously calculates the error between the model prediction result and the true label according to the set loss function, calculates the gradient of the loss value to the model parameters through the back-propagation mechanism, and uses the optimizer to adjust the model weight according to the learning rate and gradient information; Traverse each batch of training data, repeat until all batches of data in the current round are trained, save the current model weight ratio, and use the validation set to evaluate the model performance; Iterate training until the early stopping condition is met or the set maximum number of training rounds is reached. The early stopping condition is to take the validation loss value as the basis for judging the model performance, set the minimum improvement threshold and patience parameter, and if the improvement in model performance is lower than the minimum threshold in n consecutive rounds, the early stopping mechanism is triggered to terminate the training process early.

8. The method for intelligently identifying bird calls based on ensemble learning according to claim 1, characterized in that: The specific steps of model tuning in step S3 are as follows: Define the tuning space and determine the hyperparameters that need to be adjusted and their ranges, including learning rate, batch size, regularization coefficient, and Dropout ratio; Bayesian optimization is used as a hyperparameter tuning method. The performance of different hyperparameter combinations is evaluated through multiple trainings, the effect of each set of hyperparameter configurations is verified, and the search space is gradually adjusted to determine the optimal hyperparameter combination. The optimal hyperparameter combination is selected based on the tuning results and final training is performed. After the training is completed, the optimal weights of each model are saved separately.

9. The method for intelligently identifying bird calls based on ensemble learning according to claim 1, characterized in that: The specific steps of step S4 are as follows: By calling multiple models to predict the input data separately, we can get the predicted output of each model. , where N is the number of models, indicating the prediction result of the i-th model; Select the appropriate integration strategy according to specific needs. The Gaussian distribution weighted method is used by default. The Gaussian distribution weighted method is used to calculate the Gaussian distribution weight of each model according to the prediction error, and then the prediction results of each model are weighted averaged to obtain the final prediction result. The specific calculation formula is: ， in, is the prediction result of the ith model, is the Gaussian distribution weight of the model, N is the number of models, is the prediction error of the ith model, is the standard deviation of all model prediction errors; The fused results are used as the final prediction results, and the top 5 categories and their probabilities are output.

10. A system constructed according to any one of claims 1 to 9 for intelligent bird song recognition based on ensemble learning, characterized in that: include: The data collection and preprocessing module is used to clean, convert and standardize the collected bird song data to establish a standardized bird song database; The sub-dataset generation module is used to randomly sample the standardized bird song database using the Bootstrap sampling scheme based on the Bagging ensemble learning strategy to generate multiple different independent sub-datasets; The model training module is used to build and train an independent deep learning model for each sub-dataset for the integration of subsequent prediction results. It specifically includes model configuration, data preparation, model training, and model tuning steps. The result integration module is used to fuse the output results of multiple models based on the Bagging ensemble learning strategy to generate the final prediction result.

Images