Dolphin signal identification method and device based on diffusion model

Abstract

The invention discloses a dolphin signal identification method and device based on a diffusion model. The method comprises the following steps: S1, obtaining a dolphin signal data set; s2, training a diffusion model for dolphin signal denoising according to the historical dolphin signal data set; and S3, carrying out classification and identification on the denoised dolphin signal output by the trained diffusion model. By adopting the technical scheme of the invention, high-quality dolphin sonar signals can be generated.

Description

Technical Field

[0001] The present invention belongs to the technical field of audio recognition, and in particular relates to a dolphin signal recognition method and device based on a diffusion model. Background Art

[0002] The background of dolphin signal recognition technology mainly involves multiple fields such as ecological protection, fishery management and underwater acoustic communication. As a national protected animal, the protection and research of dolphins are of vital importance. Traditional monitoring methods are inefficient and prone to errors, so a method that can accurately identify and count the number of dolphins in real time is needed to support ecological protection work. In fishery management, the accidental capture of dolphins is a serious problem. A method that can monitor dolphin signals in real time and issue alarms in time is needed to reduce accidental capture and optimize fishery resource management. In the field of underwater acoustic communication, dolphin-like whistle underwater acoustic communication, as a new type of covert underwater acoustic communication technology, has attracted attention due to its good security and communication capabilities. However, the existing feature parameter recognition method cannot effectively identify the different modulation modes of dolphin-like whistle underwater acoustic communication signals, resulting in low recognition accuracy. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a dolphin signal recognition method and device based on a diffusion model.

[0004] To achieve the above object, the present invention adopts the following technical solution:

[0005] A dolphin signal recognition method based on a diffusion model, comprising:

[0006] Step S1, obtaining a dolphin signal dataset;

[0007] Step S2, training a diffusion model for dolphin signal denoising based on a historical dolphin signal dataset;

[0008] Step S3: classify and identify the denoised dolphin signal outputted by the trained diffusion model.

[0009] Preferably, in step S1, the positive samples and negative samples are mixed at different signal-to-noise ratios to obtain a dolphin signal data set; wherein the positive samples are pure dolphin signals, and the negative samples are background noise and mechanical noise.

[0010] Preferably, in step S2, the diffusion model is trained by forward expansion and reverse diffusion; wherein the training loss function is defined as:

[0011]

[0012] in, To minimize the loss function, the expected value is the signal x 0, noise∈ and the random value at time step t are averaged.

[0013] Preferably, in step S3, a convolutional neural network or a long short-term memory network is used to classify the denoised dolphin signal to identify whether it is a dolphin signal.

[0014] The present invention also provides a dolphin signal recognition device based on a diffusion model, comprising:

[0015] An acquisition module, used to acquire a dolphin signal dataset;

[0016] A training module for training a diffusion model for dolphin signal denoising based on a historical dolphin signal dataset;

[0017] The recognition module is used to classify and recognize the denoised dolphin signal output by the trained diffusion model.

[0018] Preferably, the acquisition module mixes positive samples and negative samples at different signal-to-noise ratios to obtain a dolphin signal data set; wherein the positive samples are pure dolphin signals, and the negative samples are background noise and mechanical noise.

[0019] Preferably, the training module trains the diffusion model by forward expansion and backward diffusion; wherein the training loss function is defined as:

[0020]

[0021] in, To minimize the loss function, the expected value is the signal x 0 , noise∈ and the random value at time step t are averaged.

[0022] Preferably, the recognition module uses a convolutional neural network or a long short-term memory network to classify the denoised dolphin signal to identify whether it is a dolphin signal.

[0023] The present invention aims to solve the key problems in dolphin signal recognition, especially how to accurately and efficiently recognize dolphin signals in complex marine environments to support applications such as ecological protection, fishery management and underwater acoustic communication. Traditional signal recognition methods mainly rely on manually extracted feature parameters, such as frequency, amplitude, duration, etc. These features are easily affected by noise and interference in complex marine environments, resulting in a decrease in recognition accuracy, poor model generalization ability, and poor adaptability to unknown or changing signal patterns. Although existing deep learning methods have advantages in theory, it is costly to obtain a large amount of high-quality annotated data, the annotation process is time-consuming and laborious, and the model training process requires a lot of computing resources and time, resulting in model update and deployment delays, affecting real-time monitoring and recognition efficiency. In specific application scenarios, such as monitoring dolphin populations and studying dolphin behavior in ecological protection, avoiding accidental capture and optimizing fishery resource management in fishery management, and imitating dolphin whistle underwater acoustic communication and signal camouflage and recognition in underwater acoustic communication, traditional methods have problems such as low accuracy and low efficiency. Therefore, the present invention aims to overcome the defects of traditional methods in feature extraction, model generalization, data labeling and real-time performance through dolphin signal recognition technology based on diffusion model, improve the accuracy and efficiency of dolphin signal recognition, and support practical applications such as ecological protection, fishery management and underwater acoustic communication.

[0024] The present invention recognizes dolphin signals based on a diffusion model. First, the diffusion model can generate high-quality dolphin sonar signal samples that are closer to real signals in terms of details and fidelity, thereby improving the training effect and recognition accuracy of the recognition algorithm. Secondly, its training process is stable and is not prone to problems such as mode collapse, and can better learn the complex characteristics of dolphin signals. In addition, the diffusion model is highly flexible and can generate a variety of samples by adjusting parameters to adapt to the signal characteristics of different dolphin species and environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0026] Figure 1 This is a flow chart of a dolphin signal recognition method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] Embodiment 1:

[0030] like Figure 1 As shown, an embodiment of the present invention provides a dolphin signal recognition method based on a diffusion model, comprising:

[0031] Step S1, obtaining a dolphin signal dataset;

[0032] Step S2, training a diffusion model for dolphin signal denoising based on a historical dolphin signal dataset;

[0033] Step S3: classify and identify the denoised dolphin signal outputted by the trained diffusion model.

[0034] As an implementation of an embodiment of the present invention, in step S1, the dolphin signal is a high-frequency, complex bioacoustic signal, which is transmitted in a noisy underwater environment and is often interfered by water flow, background noise and other mechanical noise. The goal is to extract the dolphin signal s(t) from the mixed signal x(t) and identify it.

[0035] The mixed signal model is:

[0036] x(t))=s(t)+n(t)

[0037] Among them, s(t) is the target dolphin signal, which usually propagates in the ultrasonic frequency band and carries the communication and positioning information of the dolphin. The frequency range is usually 10-150kHz; n(t) is the noise signal, which may include underwater background noise and other interference.

[0038] The dolphin signal dataset contains:

[0039] Positive samples: collect pure dolphin signals as target signals for training.

[0040] Negative samples: Collect other underwater noises (such as background noise, mechanical noise, etc.) to construct negative samples.

[0041] Mixed signal: Positive samples and negative samples are mixed at different signal-to-noise ratios (SNR) to generate training samples. The SNR value range should be set according to the actual situation, such as from -10dB to 20dB, etc., to cover different noise interference levels.

[0042] As an implementation of the present invention, in step S2, the diffusion model is a generative model that gradually adds noise, and the goal is to generate a target signal through reverse process denoising. Training the diffusion model includes: a forward diffusion training process and a reverse diffusion training process.

[0043] In the forward diffusion process, the dolphin signal x is gradually transmitted 0 =s(t) adds noise to generate a series of x 1 ,x 2 ,....,x T , the final signal x T will converge to a Gaussian distribution. Specifically, the conditional distribution of the forward diffusion process is expressed as:

[0044]

[0045] Among them, β t is the diffusion step length, which controls the intensity of the noise at each step. The update rule for each diffusion step is as follows:

[0046]

[0047] in, is the cumulative diffusion coefficient.

[0048] By gradually adding noise, the signal distribution at any time step t can be obtained:

[0049]

[0050] This gradual way of adding noise allows the signal x(t)x(t)x(t) to be fully mixed in high-dimensional space so as to better cope with complex background noise and interference.

[0051] After clarifying the forward diffusion process, the next step is to train the diffusion model. The training goal of the diffusion model is to let the model predict the noise ∈ θ (x t ,t) is close to the real noise∈. The training loss function is defined as:

[0052]

[0053] in, To minimize the loss function, the expected value is the signal x 0, noise∈ and the random value of time step t are averaged. By minimizing this loss function, the model can learn to denoise and restore the original dolphin signal.

[0054] The purpose of the reverse diffusion process is to reverse the forward diffusion process by learning the denoising model p θ (x t-1 |x t ) to restore the original dolphin signal s(t). The conditional distribution of the reverse diffusion process is expressed as:

[0055] p θ (x t-1 |x t )=N( t-1 ;μ θ (x t ,t),∑θ(x t , t))

[0056] Among them, the denoised mean μ θ (x t ,t) is in the form of:

[0057]

[0058] The denoised signal is restored through the reverse diffusion process. The denoising process at each step is as follows:

[0059]

[0060] Finally, the reverse process gradually removes noise through multiple steps to generate the restored dolphin signal x 0 =s(t).

[0061] As an implementation of the present invention, in step S3, a convolutional neural network or a long short-term memory network is used to classify the denoised dolphin signal to identify whether it is a dolphin signal. The input of the classifier is the denoised signal, and the output is:

[0062]

[0063] Among them, W and b are classifier parameters, To obtain the classification results, we can extract dolphin signals from complex underwater noise signals. The specific network structure and parameter settings, such as the number of convolutional layers, the number of filters, the number of hidden units of LSTM, etc., need to be adjusted and optimized according to the actual data and task requirements.

[0064] The present invention has the following advantages:

[0065] 1. High-precision signal recovery and feature extraction

[0066] The introduction of the diffusion model makes the denoising and feature extraction of dolphin signals more accurate after preprocessing. Through the gradual denoising of the diffusion process and the layer-by-layer denoising of the anti-diffusion process, the random noise interference in the signal can be effectively removed while retaining the key feature information of the dolphin signal. For example, in the anti-diffusion formula: μ θ (x t ,t),Σθ(x t , t)), the model accurately predicts the noise ∈ at each step, thereby gradually restoring features that are closer to the original dolphin signal s(t), providing high-quality input for subsequent signal recognition, significantly improving the distinguishability and representativeness of the features, and making the differences between different types of dolphin signals more clearly displayed, laying the foundation for high-precision recognition.

[0067] 2. Strong robust recognition capability

[0068] The embodiments of the present invention show strong robustness when facing dolphin signals in complex marine environments. Signals in the marine environment are often interfered by various factors, such as changes in water media, background noise, etc. During the training process, the diffusion model learns a large amount of dolphin signal data with different signal-to-noise ratios and different frequency offsets, so that the model can adapt to various complex situations. In the actual recognition process, even if the signal is interfered with to a certain extent, the model can still accurately identify the type of dolphin signal based on the extracted robust features. For example, when the signal-to-noise ratio of the dolphin signal is low, the traditional method may be difficult to identify because the noise submerges the signal characteristics, but the embodiment of the present invention relies on the diffusion model to deeply denoise and enhance the features of the signal, and can still maintain a high recognition accuracy rate, effectively improving the reliability and practicality of dolphin signal recognition in practical applications.

[0069] 3. Strong adaptability of classification and recognition performance

[0070] Combining the features extracted by the diffusion model with advanced classifiers, the embodiments of the present invention have excellent adaptability. For different types of dolphin signals, in addition, when faced with new types of dolphin signals or unknown signal patterns, by adjusting the parameters and training strategies of the diffusion model, the model can quickly adapt and achieve effective recognition, greatly expanding the application scope and scenarios of dolphin signal recognition, and providing a more flexible and reliable signal recognition tool for the fields of marine biological research, underwater acoustic communication, etc.

[0071] Embodiment 2:

[0072] The embodiment of the present invention further provides a dolphin signal recognition device based on a diffusion model, comprising:

[0073] An acquisition module, used to acquire a dolphin signal dataset;

[0074] A training module for training a diffusion model for dolphin signal denoising based on a historical dolphin signal dataset;

[0075] The recognition module is used to classify and recognize the denoised dolphin signal output by the trained diffusion model.

[0076] As an implementation method of the embodiment of the present invention, the acquisition module mixes positive samples and negative samples at different signal-to-noise ratios to obtain a dolphin signal data set; wherein the positive samples are pure dolphin signals, and the negative samples are background noise and mechanical noise.

[0077] As an implementation method of an embodiment of the present invention, the training module trains the diffusion model through forward expansion and reverse diffusion; wherein the training loss function is defined as:

[0078]

[0079] in, To minimize the loss function, the expected value is the signal x 0 , noise∈ and the random value at time step t are averaged.

[0080] As an implementation method of the embodiment of the present invention, the recognition module uses a convolutional neural network or a long short-term memory network to classify the denoised dolphin signal to identify whether it is a dolphin signal.

[0081] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A dolphin signal recognition method based on a diffusion model, characterized in that: include: Step S1, obtaining a dolphin signal dataset; Step S2, training a diffusion model for dolphin signal denoising based on a historical dolphin signal dataset; Step S3: classify and identify the denoised dolphin signal outputted by the trained diffusion model.

2. The dolphin signal recognition method based on the diffusion model according to claim 1, characterized in that: In step S1, positive samples and negative samples are mixed at different signal-to-noise ratios to obtain a dolphin signal data set, wherein the positive samples are pure dolphin signals, and the negative samples are background noise and mechanical noise.

3. The dolphin signal recognition method based on the diffusion model according to claim 2, characterized in that: In step S2, the diffusion model is trained by forward expansion and backward diffusion, wherein the training loss function is defined as: in, To minimize the loss function, the expected value is the average of the signal x0, the noise ∈, and the random values ​​of the time step t.

4. The dolphin signal recognition method based on the diffusion model as claimed in claim 3, characterized in that: In step S3, a convolutional neural network or a long short-term memory network is used to classify the denoised dolphin signal to identify whether it is a dolphin signal.

5. A dolphin signal recognition device based on a diffusion model, characterized in that: include: An acquisition module, used to acquire a dolphin signal dataset; A training module for training a diffusion model for dolphin signal denoising based on a historical dolphin signal dataset; The recognition module is used to classify and recognize the denoised dolphin signal output by the trained diffusion model.

6. The dolphin signal recognition device according to claim 5, characterized in that: The acquisition module mixes positive samples and negative samples at different signal-to-noise ratios to obtain a dolphin signal dataset; among them, the positive samples are pure dolphin signals, and the negative samples are background noise and mechanical noise.

7. The dolphin signal recognition device according to claim 6, characterized in that: The training module trains the diffusion model through forward expansion and backward diffusion; the training loss function is defined as: in, To minimize the loss function, the expected value is the average of the signal x0, the noise ∈, and the random values ​​of the time step t.

8. The dolphin signal recognition device according to claim 7, characterized in that: The recognition module uses a convolutional neural network or a long short-term memory network to classify the denoised dolphin signal and identify whether it is a dolphin signal.

Images