CN112215054B - Depth generation countermeasure method for denoising underwater sound signal - Google Patents

Abstract

The invention discloses a deep generation countermeasure method for denoising underwater sound signals, and belongs to the technical field of underwater sound signal denoising. Firstly, sampling and feature extraction are carried out on an original underwater sound signal, then the extracted signal is sent into a Gauss-limited Boltzmann machine, and semi-supervised pre-training of a probability generation model is carried out; and finally, constructing a deep generation countermeasure model, sending the data generated in the probability generation model and the real label data stream into a Bernoulli limited Boltzmann machine countermeasure model, and performing supervised training. Aiming at the characteristic extraction characteristic of the underwater sound signal, the generation countermeasure model is introduced into the probability model of the limited Boltzmann machine, so that the problems of strong dependence and overfitting of the limited Boltzmann machine in the training process caused by the complex signal carried by the underwater sound are effectively eliminated, and the training model has stronger self-applicability.

Description

A Deep Generative Adversarial Method for Underwater Acoustic Signal Denoising

technical field

The invention belongs to the technical field of underwater acoustic signal noise reduction, and can effectively reproduce the original useful signal from the underwater acoustic signal with a large signal-to-noise ratio.

Background technique

In the existing underwater acoustic signal denoising, there are traditional denoising methods, modal decomposition method based on time domain and overall modal decomposition method based on frequency domain, which need to set some empirical parameters in advance, so that the denoising process depends on experience. The classical mode decomposition method can make the denoising process no longer need to set the function in advance, but it is easy to produce mode mixing and boundary effects in the decomposition process. In order to overcome boundary mixing, the underwater acoustic signal denoising method based on CEEMDAN, refined composite multi-scale dispersion entropy and wavelet threshold denoising has achieved good results in analyzing the complexity of chaotic signals, but has no good robustness.

In recent years, deep learning algorithms have been introduced in the field of underwater acoustic denoising to increase the robustness of the system. The self-encoding network is a typical deep learning algorithm. In the denoising algorithm, no assumption is made about the independence between signal and noise. The robustness of its denoising system is better than that of traditional denoising methods. The Boltzmann machine principle is added to the auto-encoding network to further enhance the robustness of the de-noising auto-encoding network and make the network model more robust. However, the depth of Due to its fully connected feature, the auto-encoding network will increase the number of network parameters in the case of a small number of inputs with increased eigenvalues. Too many parameters will bring about gradient instability and increase the calculation amount of the entire system, so It is necessary to reduce the number of parameters on the premise of ensuring the denoising results. A new breakthrough in the field of deep learning generative modeling is the generative adversarial network, which has achieved great success in the field of computer vision, can generate realistic images, and improve model parameters and optimization algorithms on the basis of generative confrontation. Good results. The generative adversarial model utilizes the mutual optimization of the generative model and the adversarial model, which can achieve good results under a limited number of training sets, effectively solving the limitations of underwater acoustic denoising in deep learning applications.

In the process of noise reduction of underwater acoustic signals, when the corresponding parameters need to be set in advance, that is, in the case of high experience requirements, and the robustness of the system is poor, the traditional noise reduction method is no longer applicable, so deep learning uses The advantages of noise reduction are shown. At present, the generative adversarial network in deep learning, the convolutional neural network used in both the generative model and the adversarial model, because the underwater acoustic signal is characterized by fewer training samples, in the process of small sample training Among them, the adaptive ability of the convolutional neural network is weak, and the robustness of the learned network is poor.

SUMMARY OF THE INVENTION

In order to solve the problem of small sample training unique to underwater acoustic signals, this paper proposes a deep generative adversarial method for underwater acoustic signal denoising technology. The restricted Boltzmann network in this method adds sample statistical characteristics to the training of the samples, so the training effect for small samples is very good. According to the characteristics of underwater acoustics and the characteristics of underwater acoustic background noise, the generative adversarial network is further optimized, so that the training model is suitable for denoising of underwater acoustic channels.

The technical solution adopted by the present invention to solve the technical problem: a deep generation confrontation method for denoising underwater acoustic signals, and its characteristics include the following steps:

Step 1: Sampling and feature extraction of the original underwater acoustic signal. This method adopts the MFCC feature extraction method. The Mel cepstral coefficient (referred to as MFCC) is the cepstral parameter extracted in the Mel scale frequency domain, and the Mel scale Describes the nonlinear characteristics of the human ear frequency, and its relationship with frequency can be approximated by the following formula:

where f is the sampling frequency.

Step 2: Send the extracted signal into a Gauss-Restricted Boltzmann Machine for semi-supervised pre-training of the probability generation model.

The probabilistic generation model is composed of a stack of Gaussian Restricted Boltzmann Machines. First, the depth model uses unlabeled data and uses the Gauss Restricted Boltzmann Machine algorithm to perform unsupervised pre-training from the bottom layer to the top to obtain the depth. Initial values of network hyperparameters. After unsupervised pre-training, the network adjusts the weights by supervised training using labeled data. The Gauss-restricted Boltzmann machine is a generative random network consisting of a visible layer and a hidden layer. The weight θ of the network consists of the connection weight matrix ω of the visible layer and the hidden layer and the bias vector c of the visible layer, The bias vector b of the hidden layer is composed of, when a set of visible layer states v and hidden layer states h are given, the energy function and likelihood function distribution of the restricted Boltzmann machine are expressed as:

是配分函数，当可见层和隐层其中之一固定状态时，受限玻尔兹曼机的条件概率分布可以表述为Among them, v _i ∈{0,1}; h _j ∈{0,1};

is the partition function. When one of the visible layer and the hidden layer is in a fixed state, the conditional probability distribution of the restricted Boltzmann machine can be expressed as

in,

The generation part is to use a sample z under the mapping of a set of prior distribution Z to generate the distribution of x samples. The generator is an effective mapping that can simulate the distribution of real data to generate new samples related to the training set. Not a traditional input to output mapping but a mapping of input data streams to output data streams.

Step 3: Build a deep generative adversarial model, and send the data generated in the probabilistic generative model and the real label data stream into the Bernoulli restricted Boltzmann machine adversarial model for supervised training;

The generative adversarial model is composed of a generative model and an adversarial model. In the generative model, due to the complexity of the underwater acoustic signal features, when the deep neural network is back-propagated to optimize the network weights, there are too many layers, and the training system It is unstable, so the probability generation model is adopted. The structural feature is that the training process is layer-by-layer training. Each Boltzmann machine is trained separately and then superimposed to effectively avoid the inverse caused by the complex structure between layers. The propagation parameters are updated slowly and the error is large. The probability generation model is a network with no connection in the layer and full connection in the layer. The eigenvalues of the input layer of the network are equal to the eigenvalues of the output layer, and four layers of Gaussian restricted Boltzmann networks are superimposed in the middle.

判别器的功能就是可以判断出真实的样本分布X为真实数据，而生成器生成的样本分布

为假数据，生成器根据对抗器的输出结果不断的进行优化，直到对抗器无法判断输入数据是真实数据还是生成数据，说明输入数据的分布已经非常接近真实数据的分布了，这种训练成为生成对抗模型的最大最小博弈训练，其目的是如式6所示：The Bernoulli restricted Boltzmann machine adversarial model is the optimization part of the whole model. The way of generating the partial learning optimal mapping is done through adversarial training. The self-encoding model in the adversarial model uses Bernoulli restricted glass. The Ertzmann machine model is a good way to add probability judgment to data training and fine-tuning, and finally adds a discriminant layer, which is a binary classifier. There are two input data for the adversarial model, one is from the real sample distribution X, One is the real sample distribution simulated by the generator

The function of the discriminator is to determine that the real sample distribution X is the real data, and the sample distribution generated by the generator

For fake data, the generator continuously optimizes according to the output of the adversary, until the adversary cannot judge whether the input data is real data or generated data, indicating that the distribution of the input data is very close to the distribution of the real data, and this training becomes the generated data. The max-min game training of the adversarial model, whose purpose is as shown in Equation 6:

On this basis, we can add some additional conditional values to perform mapping and classification. In the denoising model, the additional conditional value is the noisy sample x _c . After adding, Equation 6 can be changed to Equation 7:

For the stability of the training process, the generator and the adversary are trained separately, and the minimum values of the two functions are optimized respectively, as shown in equations 8 and 9:

The discriminant network model is mainly responsible for the confrontation. When the discriminant system can easily discriminate that the clean sample is 1 and the generated sample is 0 at the beginning, when the generated sample of the discriminant generator is 0, the generator starts to optimize its own parameters, so that it starts from the input sample. The mapped value is close to the value of the real sample. When the generated sample is so close to the real sample that the discriminator cannot discriminate, the discriminator starts to optimize its own parameters, so that the generated sample and the real sample can be well discriminated. The input in the generation model is a A two-dimensional array. When the generated samples and noisy samples are input, the discriminator judges as 0; when the input of clean samples and noisy samples, the discriminant result is 1.

The beneficial effects of the present invention are: in view of the feature extraction characteristics of underwater acoustic signals, a generative confrontation model is introduced into the restricted Boltzmann probability model, which effectively eliminates the restricted Boltzmann caused by complex signals carried by underwater acoustics. The strong dependence and over-fitting problems of the machine in the training process make the training model more self-applicable.

The invention proposes a deep generation confrontation algorithm to improve the signal-to-noise ratio in the feature extraction technology of underwater acoustic signals. The assumptions of the sample, however, some assumptions cannot be fully satisfied in the actual underwater environment. The coding model of the generative adversarial network is a typical deep learning network model, which can be well matched without assuming independence. The noisy signal is denoised, but in the common generative adversarial denoising model, both the generative network and the adversarial network use convolutional neural networks, which have good robustness to data with a large number of training samples, but are not suitable for small samples of underwater sound. Therefore, a Gaussian restricted Boltzmann machine with probabilistic and statistical characteristics and a Bernoulli first Boltzmann machine are added to the generation unit and the confrontation unit of the generative adversarial model. First, semi-supervised learning is used to optimize the network parameters to form A deep generative adversarial network model.

The deep generative adversarial network proposed by the present invention hopes to remove the noise of the original noisy signal to obtain a clean underwater acoustic signal. The main sources of the underwater acoustic noise are ocean background noise and propeller noise, including linear noise and nonlinear noise, and The number of training samples for underwater acoustic signals is limited, and the deep generative adversarial method learning data is not a simple mapping relationship between learning data input and output, but learning the mapping relationship between the statistical characteristics and statistical characteristics of a given sample, and then use the confrontation method. In order to carry out supervised learning and continuously update the statistical characteristics of the input underwater acoustic signal, in the case of a mixed linear and nonlinear system and a small number of training sets, the underwater acoustic signal can also be well reconstructed, making it close to The deep generative adversarial model with statistical characteristics obtained has strong robustness and adaptive ability.

The present invention will be described in detail below with reference to specific embodiments.

Description of drawings

Figure 1 is the discriminant model of the system

Figure 2 is a generative adversarial network model.

Figure 3: Red is the time-domain map of underwater acoustics after denoising using the deep generative adversarial method, blue is the time-domain map of the noisy signal, and green is the time-domain map of the clean signal.

The left picture of Fig. 4 is the spectrogram of the clean underwater acoustic signal, and the right picture is the spectrogram of the signal after denoising by applying the deep generative adversarial method.

Detailed ways

Step 1 First, the samples are divided into frames and processed in batches.

Step 2 Then send the processed data into the generative model for model training. The generative model is a semi-supervised model, so there is a difference between the trained data and the clean data

Step 3: Add the data generated by the generator to the noisy data, and send it to the discriminant model together with the original clean and noisy data for discrimination. The discriminator can discriminate well at the beginning. The data of the generated model is a fake sample, and the output is 0. The original clean and noisy data is the real sample and the output is 1. According to the result discriminated by the discriminator, the generator starts to simulate its own generated data, so that the data is as close to the real data as possible, so that until the discriminator has no way to distinguish, the generator will generate The data is sent to the discriminator again, and the discriminator starts to fail to discriminate. When the input is two samples, the output is 0.5. After a period of training, it can be well discriminated that the generated sample is 0. Repeat steps 3 and 4 , until the output value of the discriminator stabilizes at 0.5, at which point the model generated by the generator is a clean sample.

It can be seen from Figures 3 and 4 that, under the same signal-to-noise ratio in this embodiment, the denoising effect of the underwater acoustic signal of small samples is due to the traditional denoising method, and the time domain diagram of the noisy signal of the trained denoising model, It has a high degree of coincidence with the signal time domain map of the clean sample. It is found in the frequency domain that the energy of the clean sample is well concentrated at one point, and the energy after the denoising model is offset, but the energy will still be concentrated in At a specific point, according to many experiments, it is found that the frequency shift is a fixed value, so with the empirical value of the frequency shift, it is possible to denoise the noisy speech very well.

Claims (1)

1. A depth-generated countermeasure method for denoising underwater acoustic signals is characterized by comprising the following steps:

step 1: sampling and feature extracting are carried out on the original underwater sound signal; the used feature extraction method is a Mel cepstrum coefficient feature extraction method, wherein Mel cepstrum coefficients are abbreviated as MFCC, MFCC is cepstrum parameters extracted in Mel scale frequency domain, Mel scale describes nonlinear characteristics of human ear frequency, and the relationship between the Mel cepstrum coefficients and the frequency is approximately expressed by the following formula:

wherein f is the sampling frequency;

step 2: sending the extracted signals into a Gauss limited Boltzmann machine for semi-supervised pre-training of a probability generation model;

Firstly, performing unsupervised pre-training on a depth model layer by layer from the bottom layer upwards by using data without labels by using a limited Gauss boltzmann machine algorithm to obtain an initial value of a hyper-parameter of a depth network, and after unsupervised pre-training, performing supervised training on the network by using data with labels to adjust a weight; the Gauss limited Boltzmann machine is a generating random network, which consists of a visible layer and a hidden layer, wherein a weight theta of the network consists of a connection weight matrix omega of the visible layer and the hidden layer, a bias vector c of the visible layer and a bias vector b of the hidden layer; given a set of visible layer states v, and hidden layer states h, the energy function and likelihood function distribution of a constrained boltzmann machine is represented as:

wherein v is_i∈{0,1}；h_j∈{0,1}；

Is a partition function, when one of the visible layer and the hidden layer is in a fixed state, the conditional probability distribution of the restricted Boltzmann machine can be expressed as

Wherein,

the generation part generates the distribution of x samples by using a sample S under the mapping of a group of prior distribution S, the generator is an effective mapping which can simulate real data distribution to generate new samples related to a training set, and the generation model learns the mapping from input data flow to output data flow instead of the traditional mapping from input to output;

And 3, step 3: constructing a deep generation countermeasure model, sending data generated in the probability generation model and a real label data stream into a Bernoulli limited Boltzmann machine countermeasure model, and carrying out supervised training;

the generation countermeasure model consists of a generation model and a countermeasure model, wherein the generation model adopts a probability generation model; the countermeasure model is a Bernoulli limited Boltzmann machine countermeasure model which is an optimized part of the whole model, the mode of generating part learning optimal mapping is completed by countermeasure training, a self-coding model in the countermeasure model adopts a Bernoulli limited Boltzmann machine model, the probability judgment is well added to data training and fine tuning,finally, a discrimination layer is added, namely a binary classifier is adopted, and input data of the countermeasure model are divided into two parts, namely a real sample distribution X from the generator and a real sample distribution simulated by the generator

The function of the discriminator is to determine the true sample distribution X as true data, and the generator generates the sample distribution

For false data, the generator continuously optimizes according to the output result of the countermeasure until the countermeasure cannot judge whether the input data is real data or generated data, which shows that the distribution of the input data is very close to that of the real data, and the training becomes the maximum and minimum game training for generating the countermeasure model, and the purpose is as shown in formula 6:

In the de-noising model, the additional condition value is a noisy sample x₁Formula 6 may become represented by formula 7 after addition:

training the generator and the countermeasure device separately, and respectively optimizing the minimum value of two functions, as shown in formulas 8 and 9:

the judgment network model is mainly responsible for antagonism, the judgment system easily judges that a clean sample is 1 and a generated sample is 0 at the beginning, when the generated sample of the judgment generator is 0, the generator starts to optimize self parameters to enable a value mapped by the generator from an input sample to be close to a value of a real sample, when the generated sample is close to the real sample and the judgment device cannot accurately judge whether the generated sample is the clean sample or the generated sample, the judgment device starts to optimize the self parameters to enable the generated sample and the real sample to be well judged, a two-dimensional array is input into the generation model, and when the generated sample and a noisy sample are input, the judgment device judges that the generated sample is 0; when the clean sample and the noisy sample are input, the discrimination result is 1.

Images