A mosquito species classification method based on voiceprint recognition, a training method for a mosquito species voiceprint recognition classification model, and a computer program.
By combining a parameterized filter bank based on the Sinc function and the MSCNet deep learning model with frequency domain and time-dependent features, the problem of low accuracy in existing mosquito voiceprint recognition was solved, and efficient mosquito classification in the field was achieved.
Patent Information
- Application Number
- CN202610384031.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-26
- Publication Date
- 2026-07-03
AI Technical Summary
Existing mosquito voiceprint recognition methods mainly rely on frequency domain features, resulting in low recognition accuracy and failing to meet the needs of large-scale screening.
Frequency domain features were extracted using a parameterized filter bank based on the Sinc function, and combined with time-dependent features. Mosquito species were classified using the MSCNet deep learning model, which included data preprocessing, a combination of SincConv layers, LSTM layers and classifiers, and model training was performed using multiple sets of training samples simulating field scenarios.
It improves the accuracy of mosquito species identification and enables efficient and accurate mosquito species classification in complex field environments.
Smart Images
Figure CN122337211A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mosquito species classification technology, and in particular to a mosquito species classification method based on voiceprint recognition, a training method for a mosquito species voiceprint recognition classification model, and a computer program. Background Technology
[0002] Diseases such as malaria and dengue fever are transmitted among humans via mosquitoes. To prevent the spread of these diseases, mosquito control is necessary. Some mosquitoes do not bite humans and therefore do not require significant manpower or resources to eradicate. Furthermore, different mosquito species may require different control methods. Therefore, to prevent diseases such as malaria and dengue fever, it is essential to accurately identify the mosquito species in the target area and then implement targeted mosquito control measures accordingly. There are three main existing methods for identifying mosquito species: One method is morphological classification, where professionals observe morphological features such as wing veins and antennae under a microscope. This identification method has a high accuracy rate, but it relies entirely on the professional skills of the professionals. A single test takes more than 30 minutes, which cannot meet the needs of large-scale screening. Secondly, there are molecular biology methods, where professionals use DNA barcoding, protein labeling and other technologies to detect mosquitoes. These identification methods have high detection accuracy but require specialized laboratory equipment, have high sample processing costs, and cannot be applied in the field. Thirdly, there is the voiceprint recognition method. This method uses the species-specificity of mosquito wingbeat voiceprints for classification. It has the advantages of non-contact and rapid detection and can be used for field detection. However, the accuracy of mosquito species identification is not high because it only relies on the frequency domain features extracted from the mosquito wingbeat audio. Summary of the Invention
[0003] The technical problem to be solved by this invention is how to improve the accuracy of mosquito species identification based on voiceprint recognition.
[0004] To address the aforementioned technical problems, this invention provides a mosquito species classification method based on voiceprint recognition, comprising the following steps: S1. Extract the frequency domain features of mosquito wingbeat audio; S3. Determine the corresponding mosquito species category based on the frequency domain characteristics and output it to the user; Between steps S1 and S3, S2 is executed to obtain the time-dependent features of the above frequency domain features, and to obtain the comprehensive features that fuse the frequency domain features and the time domain features. S3 specifically derives the corresponding mosquito species category based on the comprehensive features.
[0005] Furthermore, S1 extracts the frequency domain features of mosquito wingbeat audio using a parameterized filter bank based on the Sinc function.
[0006] Furthermore, in S1, the frequency domain features are subjected to dimensionality reduction processing.
[0007] Further, before S1, S0 is executed to receive the collected mosquito wingbeat audio.
[0008] Furthermore, in S0, the audio of mosquito wing flapping is normalized and noise-reduced.
[0009] Furthermore, S1 is specifically executed using a sinusoidal filter convolutional layer, S2 is specifically executed using an LSTM layer, and S3 is specifically executed using a classifier.
[0010] This invention also provides a training method for a mosquito species voiceprint recognition classification model, comprising the following steps: A set of training samples is constructed by taking the audio of mosquito wing flapping as input and the mosquito species as output. The mosquito species voiceprint recognition and classification model was trained using multiple sets of the above-mentioned training samples, enabling it to perform the mosquito species classification method based on voiceprint recognition as described above.
[0011] The present invention also provides a computer program product, including a computer program that, when executed, enables the mosquito species classification method based on voiceprint recognition as described above.
[0012] The mosquito species classification method based on voiceprint recognition proposed in this invention identifies mosquito species not only based on the frequency domain features of mosquito wingbeat audio, but also based on the time domain features. Specifically, S2 obtains the time dependence features of the frequency domain features of mosquito wingbeat audio, and obtains a comprehensive feature that integrates the frequency domain features and the time domain features. S3 derives the corresponding mosquito species category based on the comprehensive feature, and the identification accuracy is high. Attached Figure Description
[0013] Figure 1 This is a flowchart illustrating the mosquito species classification method based on voiceprint recognition provided in this invention.
[0014] Figure 2 This is a flowchart illustrating the model architecture and data processing of the mosquito species voiceprint recognition and classification model provided in this invention. Detailed Implementation
[0015] The present invention will be further described in detail below with reference to specific embodiments.
[0016] In this embodiment, a computer program product was obtained by technicians using computer code programming. This computer program product includes a computer program, denoted as the mosquito species voiceprint recognition and classification model. The mosquito species voiceprint recognition and classification model adopts the MSCNet (Multi-Scale Network) deep learning model architecture, such as... Figure 2As shown, the model architecture consists of a data preprocessing layer, a SincConv layer (sine filter convolutional layer), a three-stage convolutional-pooling layer, an LSTM layer, and a classifier, connected sequentially. The structure of each layer of the mosquito species voiceprint recognition and classification model is detailed below. (1) Data preprocessing layer The data preprocessing layer includes a normalization unit and a noise reduction unit, which can perform normalization and noise reduction on the input audio signal.
[0017] (2) SincConv layer The SincConv layer contains a parameterized bandpass filter bank, which includes 256 parameterized bandpass filters based on the Sinc function. The parameterized bandpass filters are located as follows:
[0018] in, This represents the low-frequency boundary of the i-th filter. This represents the high-frequency boundary of the i-th filter.
[0019] The low-frequency and high-frequency cutoff frequencies of each filter are learnable parameters, initially set to 50Hz for the low frequency and 50Hz for the initial bandwidth. During model training, the filter parameters are adaptively optimized based on the classification loss function through backpropagation, automatically matching the bandwidth of each filter to the discriminative spectral characteristics of different mosquito species' wingbeat vocalizations, thus achieving adaptive capture of mosquito species' vocalization frequency domain features. The SincConv layer of the mosquito species vocalization recognition and classification model uses a parameterized bandpass filter bank to perform parallel frequency domain mapping on the input audio signal. The parameterized bandpass filter can automatically adjust the bandwidth, thus capturing fine-grained frequency domain features.
[0020] In other embodiments, Gabor filters or wavelet transforms can be used instead of the SincConv layer to achieve frequency domain feature extraction.
[0021] (3) Three-level convolution-pooling layer A three-stage convolution-pooling structure is adopted: The first convolutional layer consists of 256 64×1 convolutional kernels with Leaky ReLU activation function. Second convolutional layer: 128 32×1 convolutional kernels; The third convolutional layer: 64 16×1 convolutional kernels; Max pooling layer: The pooling size is 2, which reduces the feature dimension and retains key information.
[0022] (4) LSTM layer An LSTM layer with 32 hidden units is introduced, and the time-dependent features of the wing flapping signal are captured through a gating mechanism. The update equation is as follows:
[0023]
[0024]
[0025]
[0026]
[0027] in, , , , These are the input gate, forget gate, output gate, and cell state, respectively. For the Sigmoid activation function, This represents the element-wise product.
[0028] In other embodiments, a CNN+Transformer architecture can be used instead of the LSTM layer to capture long temporal dependencies.
[0029] (5) Classifier The classifier consists of two fully connected layers and an output layer connected in sequence.
[0030] Fully connected layer 1: Output dimension 256, activated by Leaky ReLU function; Fully connected layer 2: output dimension 128, Dropout (preservation probability 0.5).
[0031] Output layer: Activated by the Softmax function, outputting the probability distribution of the target mosquito species.
[0032] The output layer selects the category with the highest probability as the classification result for output.
[0033] This embodiment uses mosquito wing-beat audio as input and mosquito species as output to form a set of training samples, and constructs multiple such sets of training samples. Specifically, the training samples are usually obtained in a single environment, such as a laboratory environment. Therefore, this embodiment performs noise injection, frequency domain transformation, and signal perturbation on the training samples to simulate mosquito wing-beat audio samples collected in various field scenarios. The specific process is as follows: Noise injection: -20dB to -10dB of environmental noise is superimposed on the mosquito wing flapping audio of the training samples to simulate the complex acoustic environment under field collection conditions. The model is trained using these training samples to improve the robustness and generalization ability of the model in low signal-to-noise ratio scenarios. Frequency domain transformation: Without destroying the original spectrum structure, the frequency of the mosquito wingbeat audio of the training samples is stretched by ±5% to simulate the natural fluctuation of the mosquito wingbeat frequency. Using this type of training sample to train the model can improve the model's ability to adapt to frequency shifts. Signal perturbation: The mosquito wing flapping audio signal of the training sample is randomly shifted by 0.005~0.01 seconds to simulate the uncertainty of the signal starting position during actual collection. Using this type of training sample to train the model can improve the robustness of the model to time shift.
[0034] The mosquito voiceprint recognition and classification model was trained using multiple sets of the above-processed training samples, enabling it to perform tasks such as... Figure 1 The mosquito species classification method shown.
[0035] Users install the pre-trained mosquito species voiceprint recognition and classification model on their computers. The computer processor then runs this model to achieve the following: Figure 1 The mosquito species classification method based on voiceprint recognition is shown below. The steps of this mosquito species classification method are described in detail below.
[0036] S0. Receive the collected mosquito wingbeat audio and perform normalization and noise reduction processing on it.
[0037] In this embodiment, the audio of mosquito wing flapping is collected via a microphone, and then input into a mosquito species voiceprint recognition and classification model. After receiving the mosquito wing flapping audio, the mosquito species voiceprint recognition and classification model uses a data preprocessing layer to acquire the audio signal from the mosquito wing flapping audio at a sampling rate of 8000Hz. The acquired audio signal is then segmented into a two-dimensional tensor with a frame length of 0.1 seconds and a frame step size of 0.01 seconds.
[0038] Where N is the number of frames and F is the frequency domain feature length. Each audio segment retains a maximum of 100 frames.
[0039] The data preprocessing layer then standardizes and normalizes the segmented audio signal to reduce the impact of volume fluctuations on subsequent feature extraction. A 50Hz high-pass filter is then used to remove low-frequency noise from the segmented audio signal. The mosquito wing-flashing audio signal processed by the data preprocessing layer is output to the SincConv layer.
[0040] S1. Extract the frequency domain features of mosquito wingbeat audio.
[0041] The mosquito species voiceprint recognition and classification model then activates the SincConv layer, which extracts frequency domain features through a parameterized filter bank. A three-stage convolutional-pooling layer is then used to perform dimensionality reduction and nonlinear mapping on the frequency domain features. The SincConv layer outputs the dimensionality-reduced frequency domain features to the LSTM layer.
[0042] S2. Obtain the time-dependent features of the above frequency domain features to obtain the comprehensive features that fuse the frequency domain features and time domain features.
[0043] The mosquito species voiceprint recognition classification model uses an LSTM layer to capture time-dependent features in the frequency domain through a gating mechanism, and outputs a comprehensive feature that fuses the frequency domain features and time domain features to the classifier.
[0044] S3. Based on the comprehensive characteristics, the corresponding mosquito species category is determined and output to the user.
[0045] The classifier uses two fully connected layers to map the combined features to the category space, uses the Softmax function of the output layer to output the probability distribution of each mosquito species, and then outputs the category corresponding to the highest probability as the classification result to the user.
[0046] The above description is merely an embodiment of the present invention and does not limit the scope of patent protection. Any non-substantial changes or substitutions made by those skilled in the art based on the present invention will still fall within the scope of patent protection.
Claims
1. A mosquito species classification method based on voiceprint recognition, comprising the following steps: S1. Extract the frequency domain features of mosquito wingbeat audio; S3. Determine the corresponding mosquito species category based on the frequency domain characteristics and output it to the user; Its characteristics are: Between steps S1 and S3, S2 is executed to obtain the time-dependent features of the above frequency domain features, and to obtain the comprehensive features that fuse the frequency domain features and the time domain features. S3 specifically derives the corresponding mosquito species category based on the comprehensive features.
2. The mosquito species classification method based on voiceprint recognition as described in claim 1, characterized in that, S1 extracts the frequency domain features of mosquito wingbeat audio using a parameterized filter bank based on the Sinc function.
3. The mosquito species classification method based on voiceprint recognition as described in claim 1, characterized in that, In S1, the frequency domain features are subjected to dimensionality reduction processing.
4. The mosquito species classification method based on voiceprint recognition as described in claim 1, characterized in that, Before S1, execute S0 to receive the collected mosquito wing flapping audio.
5. The mosquito species classification method based on voiceprint recognition as described in claim 4, characterized in that, In S0, the audio of mosquito wing flapping is normalized and noise-reduced.
6. The mosquito species classification method based on voiceprint recognition as described in any one of claims 1 to 5, characterized in that, S1 is executed using a sinusoidal filter convolutional layer, S2 is executed using an LSTM layer, and S3 is executed using a classifier.
7. A training method for a mosquito species voiceprint recognition and classification model, characterized by: Includes the following steps: A set of training samples is constructed by taking the audio of mosquito wing flapping as input and the mosquito species as output. The mosquito species voiceprint recognition classification model is trained using multiple sets of the above-mentioned training samples, so that it can perform the mosquito species classification method based on voiceprint recognition as described in any one of claims 1 to 6.
8. A computer program product comprising a computer program, characterized in that, When executed, the computer program is able to implement the mosquito species classification method based on voiceprint recognition as described in any one of claims 1 to 6.