CN110456332A - A kind of underwater sound signal Enhancement Method based on autocoder - Google Patents
A kind of underwater sound signal Enhancement Method based on autocoder Download PDFInfo
- Publication number
- CN110456332A CN110456332A CN201910738375.0A CN201910738375A CN110456332A CN 110456332 A CN110456332 A CN 110456332A CN 201910738375 A CN201910738375 A CN 201910738375A CN 110456332 A CN110456332 A CN 110456332A
- Authority
- CN
- China
- Prior art keywords
- signal
- autocoder
- layer
- noise reduction
- noise
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/52—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S15/00
- G01S7/539—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S15/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The underwater sound signal Enhancement Method based on autocoder that the invention discloses a kind of, belongs to field of underwater acoustic signal processing.Difficult problem is extracted for ultrasonic echo feature in active sonar, the present invention devises the autocoder that noise reduction autocoder is combined with convolution noise reduction autocoder.Noise reduction advantage first with noise reduction autocoder in signal on the whole, pre-processes signals and associated noises；Optimization then in conjunction with convolution noise reduction autocoder to signal local feature carries out local noise reduction to signal, to realize signal enhancing.The method of the present invention can be inputted directly using the time domain waveform for receiving signal as feature, remain the amplitude and phase property of signal.The experimental results showed that the present invention not only effectively reduces the noise component(s) in signal, but also preferable recovery effects are reached in time domain and frequency domain.
Description
Technical field
The present invention relates to a kind of method of underwater sound signal enhancing, especially a kind of underwater sound signals based on depth learning technology
The method of enhancing, belongs to field of underwater acoustic signal processing.
Background technique
In signal processing, due to the influence of noise, so that long-range detection and Weak Signal Processing become very
Difficulty, noise reduction become the project of long-standing problem researcher, and reducing noise is that one effectively analyzed signal must not
The process that can lack.In traditional signal processing method, noise reduction is completed by filtering, and is using linear filter method
When, according to the characteristic distributions of signal in a frequency domain, as long as time series long enough, for the noise in period and quasi-periodic signal
It can thoroughly eliminate.But for the noise that nonlinear system generates, since signal and noise show on frequency spectrum and are
Broadband continuous spectrum, so that the filter effect of conventional method substantially reduces, this just needs to explore the new nonlinear properties that are suitable for
Noise-reduction method.
With the development of computer technology, neural network algorithm is widely applied, and deep learning is by Hinton G
Et al. proposed the neural network of multilayer learning structure having in 2006, although each hidden layer is general only in network structure
Simple nonlinear transformation is used, but the nonlinear combination of multitiered network can generate extremely complex non-linear change
It changes, therefore deep learning has powerful feature learning ability, can excavate out changing rule inherent in data.Deep learning
Since proposition, the extensive concern of domestic and international many scholars is caused, has not only been continued to introduce new on theoretical algorithm, in image
The application of the practical matters such as identification, picture noise reduction, Speech processing, simulation human brain is also increasing.
Summary of the invention
In order to overcome many disadvantages in conventional method, when the technical problem to be solved in the present invention, provides a kind of noise reduction effect
The method of the relatively good underwater sound signal enhancing based on depth learning technology.
In order to solve above-mentioned technical problem, the technical scheme is that
A kind of underwater sound signal Enhancement Method based on autocoder, comprising the following steps:
(1) neural network model and a mind based on convolution autocoder of a noise reduction autocoder are constructed
Through network model；The noise reduction autocoder inputs signals with noise, and output signal is defeated as the convolution autocoder
Enter signal, symmetrical structure, including coded portion and decoded portion is presented in the convolution autocoder, and coded portion believes input
It number is encoded, characteristic information is compressed to lower dimensional space, decoded coded portion for low-dimensional characteristic information solution and be pressed into clean signal；
(2) a series of echoes being likely to be received are emulated according to the linear FM signal parameters that active sonar emits to believe
Number, generate signals with noise and the corresponding signal pair of clean signal；
It (3) is training sample set and test sample collection by data sample random division；
(4) parameter of training sample set pre-training noise reduction autocoder neural network model is used, until training set sample damages
Function is lost to touch the mark；
(5) use the output of noise reduction autocoder as the input of convolution autocoder, clean signal is exported as it, with
The parameter of this pre-training convolution autocoder neural network model, until training set sample losses function touches the mark；
(6) with the input of signals with noise neural network as a whole, with the output of clean signal neural network as a whole,
Tuning is carried out to encoder with this data set, until training set sample losses function touches the mark；
(7) setting for completing coder parameters, obtains network model parameter, and joint noise reduction is used as after underwater sound signal is sampled
The input of the whole neural network of autocoder and convolution autocoder, the enhancing signal after obtaining noise reduction.
Further, the signal that the signals with noise in step (2) and clean signal are constituted is to being all time domain waveform, step
(4) noise reduction autocoder outputs and inputs as signals with noise and clean signal in, in the complete noise reduction autocoder of pre-training
Afterwards, network corresponding to signals with noise is exported into the input as convolution autocoder, and certainly using clean signal as convolution
The output of dynamic encoder.
Further, in step (6), when the loss function of noise reduction autocoder is less than preset threshold T, convolution is automatic
When the loss function of encoder is less than preset threshold P, start using signals with noise as the input of combined coding device, clean signal is made
For the output of combined coding device, joint tuning is carried out to entire encoder network.
Further, the neuron number of each layer of the coded portion of the noise reduction autocoder with the number of plies be incremented by and
It reduces, the neuron number that each layer of decoded portion is incremented by with the number of plies and is increased, until being equal to the sampling number of clean signal.
Further, the network model of the convolution autocoder includes input layer, output layer and several convolutional layers,
The convolutional layer number of plies is more than or equal to 3 layers, and the port number of each convolutional layer is more than or equal to 30；In coded portion, each layer of convolution
Pond layer is added in layer below, and pond layer is maximum pond layer, and the port number of pond layer is equal with the port number of convolutional layer before；
Up-sampling layer is added behind decoded portion, each layer of convolutional layer, up-samples the channel of the port number and convolutional layer before of layer
Number is equal.
Further, the echo-signal received is emulated according to the following formula in step (2):
Wherein, r0It (t) is received echo-signal, x (t) is transmitting signal, and n (t) is receiving end Gauusian noise jammer, etc.
First item is direct sound wave on the right side of number, and Section 2 is multipath signal, and parameter L is the number for passing through the intrinsic sound ray of receiving point；Ai,
τiRespectively the i-th approach reaches receiving point signal amplitude and the time delay value relative to through acoustical signal, A0For the acoustical signal width that goes directly
Degree.
Beneficial effects of the present invention:, can be well since the present invention uses the data-driven learning ability of deep learning
The change due to channel in ocean is solved, the influence with ambient noise leads to the problem for extracting echo-signal difficulty, can pass through
The powerful non-linear mapping capability of neural network reduces the nonlinear noise in signal.The present invention is directly with the time domain waveform of signal
As the feature that outputs and inputs of neural network, other transformation and feature extraction is not done to signal, save a large amount of manpowers；This hair
Training dataset in bright is generated in the case where knowing the parameters of active sonar transmitting signal, and reality is not only conformed with
Active sonar detect condition, moreover it is possible to be network have stronger robustness, due to the characteristics such as the time-varying space-variant of ocean channel, people
For feature extraction often do not have robustness, and be a kind of data-driven learning algorithm by deep learning, can be according to existing
There are emulation data to automatically extract out the feature of more robust, so that noise abatement can achieve higher level.
Detailed description of the invention
Fig. 1 is algorithm block diagram of the invention；
Fig. 2 is the basic computational ele- ment neuron calculating process schematic diagram in deep learning convolutional neural networks；
Fig. 3 is the pond layer calculating process schematic diagram in deep learning convolutional neural networks；
Fig. 4 (a) is the schematic diagram without the full articulamentum of dropout；
Fig. 4 (b) is the schematic diagram by the full articulamentum of dropout；
Fig. 5 is the neural network model schematic diagram of noise reduction autocoder；
Fig. 6 is the neural network model schematic diagram of convolution autocoder；
Fig. 7 (a) is the variation diagram of noise reduction effect and the noise reduction autocoder network number of plies；
Fig. 7 (b) is the variation diagram that noise reduction effect and noise reduction autocoder compress smallest dimension；
Fig. 7 (c) is the variation diagram of noise reduction effect and the number of iterations；
Fig. 7 (d) is noise reduction effect and noise reduction autocoder is used alone and the change of convolution autocoder is used in combination
Change figure；
Fig. 8 (a) is the time domain waveform of clean signal；
Fig. 8 (b) is the noisy signal time domain waveform that signal-to-noise ratio is -10dB；
Fig. 8 (c) is the time domain waveform that network exports after the training of noise reduction autocoder is used alone；
Fig. 8 (d) is the time domain waveform of network output after the training of joint convolution autocoder；
Fig. 9 (a) is the time frequency distribution map for the noisy signal that signal-to-noise ratio is -10dB；
Fig. 9 (b) is the time frequency distribution map of network output after the training of joint convolution autocoder.
Specific embodiment
Combined with specific embodiments below, the present invention is furture elucidated, it should be understood that these embodiments are merely to illustrate the present invention
Rather than limit the scope of the invention, after the present invention has been read, those skilled in the art are to various equivalences of the invention
The modification of form falls within the application range as defined in the appended claims.
A kind of underwater sound signal Enhancement Method based on autocoder disclosed by the embodiments of the present invention, main includes following step
It is rapid:
(1) the autocoder neural network model of input recurrence identical with output neuron number, net are constructed
The frame diagram of network is as shown in Figure 1, training network uses noise reduction autocoder and convolution autocoder (Convolutional
Denoising Automatic-Encoder, CDAE) joint autocoder (DAE+CDAE) mode for combining.It is instructing in advance
Practice the stage, training set (train_clean) is known as noisy signal (train_noise), train_noise conduct plus noise
The input signal of DAE network, echo signal of the train_clean as DAE network train DAE network by reversed tuning,
Train_noise, test_noise are inputted DAE, obtain new training set (train1), new test set (test1), it is pre- to instruct
Practicing the stage terminates.By train1As the input signal of CDAE, echo signal of the train_clean as CDAE, by reversed
Tuning trains CDAE network, after training is completed, to test1Network test is carried out, final denoised signal is obtained, is completed whole
A training process.Specific noise reduction autocoder is as shown in figure 5, specifically convolution autocoder is as shown in fig. 6, carrying out
When cataloged procedure, to pond layer is added behind each layer of convolutional layer, pond layer is maximum pond layer.
(2) in the case where the parameters of the linear FM signal of known active sonar transmitting, a series of possibility are emulated
The echo-signal received generates signals with noise and the corresponding signal pair of clean signal, and noise reduction is converted to after pre-processing certainly
The input form of dynamic encoder；Specifically, the echo-signal made an uproar to echo-signal and band emulates, by artificial echo signal and
The signals with noise of emulation is normalized, data divide, and is fabricated to sample of every M sampled point as a primary data sample
It is right；
(3) by all data samples to being randomly divided into two class of training sample set and test sample collection；
(4) noise reduction autocoder neural network is trained with training sample set, method therefor is by reversely passing
Broadcast algorithm algorithm.Judge whether to reach preset index by the loss function of noise reduction autocoder neural network that (loss function is small
In equal to preset threshold T), if touched the mark, retains the parameter of network, the input of noise reduction autocoder neural network is cut
It is changed to test data sample set, if do not touched the mark, on the basis of original network model parameter, continues to use training sample set
Noise reduction autocoder neural network is trained, to constantly update the parameter of network.Loss function is judged again, is recycled past
It is multiple, until loss function touches the mark；
(5) band of training set sample of making an uproar is inputted in the network of above-mentioned trained noise reduction autocoder, is output it
As the input of convolution autocoder, and the input form for being transformed into convolutional neural networks will be output and input.
(6) convolution autocoder neural network is trained with above-mentioned data set, method therefor is by reversely passing
Broadcast algorithm algorithm.Judge whether the loss function of convolution autocoder neural network reaches preset index, if reaching finger
Mark (loss function is less than or equal to preset threshold P), retains the parameter of network, the input of convolution autocoder neural network is cut
It is changed to test data sample set, if do not touched the mark, on the basis of original network model parameter, continues to use training sample set
Convolution autocoder neural network is trained, to constantly update the parameter of network.Loss function is judged again, is recycled past
It is multiple, until loss function touches the mark；Finally with the input of signals with noise neural network as a whole, using clean signal as whole
The output of somatic nerves network, to encoder carry out tuning, until training set sample losses function touch the mark (loss function be less than etc.
In preset threshold Q)；
(7) setting for completing coder parameters, obtains network model parameter, and joint noise reduction is used as after underwater sound signal is sampled
The input of the whole neural network of autocoder and convolution autocoder, i.e., the enhancing signal after exportable noise reduction, will roll up
The output transform of product autocoder is normal time domain waveform form, and shows output result.
In above-mentioned steps, signal that signals with noise and clean signal are constituted is to being all time domain waveform, noise reduction in step (4)
Outputting and inputting for autocoder makes an uproar band after the complete noise reduction autocoder of pre-training for signals with noise and clean signal
Network corresponding to signal exports the input as convolution autocoder, and using clean signal as convolution autocoder
Output.
In step (6), when the loss function of noise reduction autocoder is less than T, the loss function of convolution autocoder
When less than P, start output of the clean signal as combined coding device, to whole using signals with noise as the input of combined coding device
A encoder network carries out joint tuning.
The neuron number that each layer of the coded portion of noise reduction autocoder is incremented by with the number of plies and is reduced, and decoded portion is every
One layer of neuron number is incremented by with the number of plies and is increased, until being equal to the sampling number of clean signal.
Depth convolution autocoder network model includes input layer, output layer and several convolutional layers, the convolutional layer
The number of plies is more than or equal to 3 layers, and the port number of each convolutional layer is all larger than equal to 30；After coded portion, each layer of convolutional layer
Pond layer is added in face, and pond layer is maximum pond layer, and the port number of pond layer is equal with the port number of convolutional layer before；It is solving
Up-sampling layer is added behind each layer of convolutional layer, up-samples the port number of layer and the port number phase of convolutional layer before for code part
Deng.
Below for the present embodiments relate to related encoder related notion and calculating process detailed explain
It states.
A, neuron and its calculating
As shown in Fig. 2, neuron is the basic operation unit in neural network, it is the certain of mimic biology nerve cell
Characteristic and the mathematicization model built, the input of neuron can for multiple variables but between be divided into weight, in neuron
Linear combination is carried out, and itself and needs are compared with a certain threshold value as neuron self attributes, then can just be motivated
Neuron response output, and generally non-linear relation between the response output and the linear combination of input of neuron, so
The calculating of the mathematical model of neuron is as follows:
If
Wherein xiFor each input of neuron, wiFor corresponding weight,For nonlinear activation function vector, i.e.,N is upper one layer of neuron number, and b is biasing.
Nonlinear activation functionIt is generally following several:
Nonlinear activation function in the network model of the present embodiment uses second.
B, neural network
The weight and threshold value of neural network are generally obtained by back-propagation algorithm, the calculating process in relation to back-propagation algorithm
It is introduced later.
C, pond layer is handled
As shown in figure 3, the parameter for needing to adjust also can be many due to when there are many number of the neuron of neural network,
Pondization operation can be carried out to one layer of neural network, i.e., it is down-sampled to input data progress, both reduce calculation amount in this way, also increases
Strong generalization ability.As shown, by one number of multiple number synthesis in the range of pond, it is most common to have maximum pond, be averaged
Chi Hua, summation pond etc..The pond layer of the network model of the present embodiment takes Chi Huafan using maximum pond layer, the i.e. result in pond
The maximum number enclosed.If pond range impartial cannot be drawn due to the shape of input data and the shape of pond range
Point, pondization operation can be being carried out suitably with 0 cover.
D, dropout is handled
As shown in Fig. 4 (a) and Fig. 4 (b), for the generalization ability for further increasing model, Hinton proposition is once being trained
Certain neurons can be hidden according to certain probability in the process, i.e., parameter related with the neuron can't be trained specifically
It is updated in journey, keeps the numerical value of last training, and in the last test stage, all neurons can all participate in calculating.
E, back-propagation algorithm
As shown in figure 5, DAE is improved on the basis of autocoder (Automatic-Encoder, AE), it is one
Kind executes the network structure of data compression, is learnt by neural network to sample data, can be by being automatically learned
Compression function and decompression function.The main thought of DAE is an autocoder trained first, can be in the encoder input layer
Addition random noise manually rebuilds input data in output layer；It then, can be to input number by the encoder model after training
According to being compressed and being decompressed, noise reduction is realized in this course, to can generate better feature for subsequent detection mission
It indicates.With the DAE basic network topology for abandoning structure as shown, wherein X is original signal,For signals with noise, y is hidden
Containing layer,For output layer.If input feature vector is X, output feature is Y, then desired objective result is by mapping valueAs far as possible
Ground is close to x, it is possible to and a square of reconstructed error function, the i.e. loss function of DAE are constructed, it is minimized, from
And the parameter after being optimized:
Wherein N is the number of neuron, Wopt,Wo′pt,bopt,bo′ptNetwork weight before respectively updating, it is updated
Network weight, the network biasing before update, updated network biasing.
Then it is updated by weight parameter of the stochastic gradient descent method to whole network, solves the optimal of objective function
Solution.Noise reduction autocoder algorithm initializes weight and offset parameter first, and then iteration update finds out optimal solution, specific algorithm
Steps are as follows:
(1) random initializtion: to all l, if Δ W(l)=0, Δ b(l)=0；
(2) the number of iterations: from i=1 to m:
A) it is calculated using BP algorithmWith
B) it calculates
C) it calculates
(3) weight parameter is updated:
Wherein α is learning rate, and λ is the preset parameter of network.In conjunction with the above full articulamentum and each layer of parameters
Partial derivative calculation method, and so on, with reference to the accompanying drawings 5 and attached drawing 6 schematic diagram, each parameter of preceding layer is deduced by later layer
Partial derivative, thus acquire entire depth study convolutional neural networks model partial derivative, finally, press parameter space negative gradient
Direction undated parameter, this is back-propagation algorithm.
As shown in fig. 6, CDAE is mainly formed by encoding (encoder) and decoding (decoder) two parts, instructed by layering
Practice to optimize overall structure.Symmetrical structure is presented in CDAE network, and this structure first carries out input signal in preceding two convolutional layer
Coding, is compressed to lower dimensional space for characteristic information, is decoded hidden layer in rear two layers of convolutional layer, thus by low-dimensional feature
Information solution is pressed into clean signal.
The input of each convolutional layer is 3 dimensional feature data, and filter, Chi Hua, up-sampling are all using 2 dimension operators, every
In a convolutional layer, CDAE will input information MAP to the stronger feature of more abstract, robustness by the denoising transformation of study.
Encoder section in CDAE is made of multiple convolutional layers, active coating and pond layer, and convolutional layer is made of one group of filter, these
Filter extracts feature from their input layer, and the active coating in the present invention is to apply nonlinear amendment list to characteristic pattern
Member.In the layer of pond, pond function selects maximum pond function (Max-pooling), and Max-pooling is by mapping specific sky
Between maximum value in range invariant, down-sampling is carried out to active coating and generates the new mappings space with dimensionality reduction.It is solving
Code part is made of convolutional layer, active coating and up-sampling layer (Up-sampling), and up-sampling layer is by the activation to front
Layer is up-sampled, and the mew layer of higher-dimension is generated.
The loss function defined formula network update side of the loss function of CDAE and the loss function of overall network and DAE
Formula be also it is the same, details are not described herein again.
In view of underwater acoustic channel belongs to the double random inhomogeneous medium channels in interface of out-of-flatness in experimentation, thus emit letter
Number by time-varying in underwater acoustic channel communication process, space-variant and multipath extension are serious, receive signal waveform and are distorted.Underwater sound letter
Road has the characteristic of multipath extension, emits sound ray of the signal Jing Guo different approaches and successively reaches receiving hydrophone, last connects
The collection of letters number is that the interference superposition of arriving signal is propagated through each sound ray.If transmitting signal is x (t), receiving end Gauusian noise jammer
For n (t), then signal r is received0(t) are as follows:
Wherein, first item is direct sound wave on the right side of equal sign, and Section 2 is multipath signal, and parameter L is the sheet by receiving point
Levy the number of sound ray；Ai, τiRespectively the i-th approach reaches receiving point signal amplitude and the time delay value relative to through acoustical signal, A0
For the acoustical signal amplitude that goes directly.
In the emulation of training data, using unpolluted LFM as echo signal, it is first determined bandwidth be 2kHz~
8kHz generates the LFM signal of 10ms, then radom insertion carrys out simulated target echo in some time point under the sample rate of 48k
The position of appearance, generates one section of echo signal with 1200 sampled points, symbiosis at 200500 such echo signals,
In 200000 signals as training set, remaining 500 are used as test set；On the basis of echo signal, if intrinsic line number is
3, signal amplitude be followed successively by section [0.9,0.6), [and 0.6,0.3), some random value of [0.3,0.1], time delay value is followed successively by
(rand*100) a sampled point is at random -20dB~5dB white Gaussian noise plus signal-to-noise ratio, directly makees its time-domain signal
For input signal, it can preferably retain the phase information of input signal.In training network, changes the DAE network number of plies, can obtain
The variation diagram of reinforcing effect and the DAE network number of plies, such as Fig. 7 (a)；When changing compression smallest dimension, reinforcing effect and compression can be obtained
The variation diagram of smallest dimension, such as Fig. 7 (b)；When changing the number of iterations, the variation diagram of reinforcing effect and the number of iterations, such as Fig. 7 can be obtained
(c)；It only includes the signal enhancing effect of DAE and network comprising two kinds of situations of DAE, CDAE that network, which is shown, in Fig. 7 (d).Thus
It determines optimal network structure, after completing training to the network structure, enhanced result figure can be obtained, such as Fig. 8 (d), In
Shown in clean signal such as Fig. 8 (a) before training, signal-to-noise ratio is shown in noisy signal such as Fig. 8 (b) of -10dB, in only DAE
Network structure under enhancing signal such as Fig. 8 (c).
In order to further verify the feasibility and validity of proposition method of the present invention, the transmission of Song Hua River subglacial signal has been carried out
Experiment.Emit signal using LFM signal, LFM signal pulsewidth is 10ms, bandwidth is 2~8kHz.Signal input will be received originally
Be restored signal in invention network, carries out time frequency analysis to the signal for restoring front and back, obtains it and analyze result such as Fig. 9 (a), 9
(b) shown in.
The time-frequency figure as shown in Fig. 9 (a), 9 (b), it can be found that the ambient noise after restoring is obviously reduced, in echo signal
Period in, signal, which is more concentrated on, to be had in frequency range, this is also to be consistent with the result of theoretical simulation.Due to
Experimental signal is difficult to and emits signal alignment, can not calculate its accurate signal-to-noise ratio, the present invention is by being approximately to make an uproar by no signal section
Acoustical signal, having signal segment is approximately echo signal, calculates its average energy value, it can be deduced that signal restore before signal-to-noise ratio be
3.75dB, signal-to-noise ratio is 17.56dB after recovery.Therefore, result above is demonstrating the depth of the invention designed to a certain degree
It practises neural network structure and is solving the problems, such as the validity in sonar signals enhancing.
Claims (6)
1. a kind of underwater sound signal Enhancement Method based on autocoder, it is characterised in that: the following steps are included:
(1) neural network model and a nerve net based on convolution autocoder of a noise reduction autocoder are constructed
Network model；The noise reduction autocoder inputs signals with noise, and output signal is believed as the input of the convolution autocoder
Number, the convolution autocoder is presented symmetrical structure, including coded portion and decoded portion, coded portion by input signal into
Characteristic information is compressed to lower dimensional space by row coding, is decoded coded portion for low-dimensional characteristic information solution and is pressed into clean signal；
(2) a series of echo-signals being likely to be received are emulated according to the linear FM signal parameters that active sonar emits,
Generate signals with noise and the corresponding signal pair of clean signal；
It (3) is training sample set and test sample collection by data sample random division；
(4) parameter for using training sample set pre-training noise reduction autocoder neural network model, until training set sample losses letter
Number touches the mark；
(5) use the output of noise reduction autocoder as the input of convolution autocoder, clean signal is pre- with this as its output
The parameter of training convolutional autocoder neural network model, until training set sample losses function touches the mark；
(6) with the input of signals with noise neural network as a whole, with the output of clean signal neural network as a whole, with this
Data set carries out tuning to encoder, until training set sample losses function touches the mark；
(7) setting for completing coder parameters, obtains network model parameter, automatic as joint noise reduction after underwater sound signal is sampled
The input of the whole neural network of encoder and convolution autocoder, the enhancing signal after obtaining noise reduction.
2. a kind of underwater sound signal Enhancement Method based on autocoder according to claim 1, it is characterised in that: step
(2) signal that signals with noise and clean signal in are constituted is to being all time domain waveform, noise reduction autocoder in step (4)
Outputting and inputting will be corresponding to signals with noise after the complete noise reduction autocoder of pre-training for signals with noise and clean signal
Network exports the input as convolution autocoder, and using clean signal as the output of convolution autocoder.
3. a kind of underwater sound signal Enhancement Method based on autocoder according to claim 1, it is characterised in that: in step
Suddenly in (6), when the loss function of noise reduction autocoder is less than preset threshold T, the loss function of convolution autocoder is less than
When preset threshold P, start output of the clean signal as combined coding device using signals with noise as the input of combined coding device,
Joint tuning is carried out to entire encoder network.
4. a kind of underwater sound signal Enhancement Method based on autocoder according to claim 1, it is characterised in that: described
Noise reduction autocoder each layer of coded portion of neuron number with the number of plies be incremented by and reduce, each layer of decoded portion
Neuron number is incremented by with the number of plies and is increased, until being equal to the sampling number of clean signal.
5. a kind of underwater sound signal Enhancement Method based on autocoder according to claim 1, it is characterised in that: described
The network model of convolution autocoder include input layer, output layer and several convolutional layers, the convolutional layer number of plies is greater than
Equal to 3 layers, the port number of each convolutional layer is more than or equal to 30；Pond layer, Chi Hua are added behind coded portion, each layer of convolutional layer
Layer is maximum pond layer, and the port number of pond layer is equal with the port number of convolutional layer before；In decoded portion, each layer of convolution
Up-sampling layer is added in layer below, and the port number for up-sampling layer is equal with the port number of convolutional layer before.
6. a kind of underwater sound signal Enhancement Method based on autocoder according to claim 1, it is characterised in that: step
(2) echo-signal received is emulated according to the following formula:
Wherein, r0It (t) is received echo-signal, x (t) is transmitting signal, and n (t) is receiving end Gauusian noise jammer, and equal sign is right
Side first item is direct sound wave, and Section 2 is multipath signal, and parameter L is the number by the intrinsic sound ray of receiving point；Ai, τiPoint
Not Wei the i-th approach reach receiving point signal amplitude and the time delay value relative to through acoustical signal, A0For the acoustical signal amplitude that goes directly.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910738375.0A CN110456332B (en) | 2019-08-12 | 2019-08-12 | Underwater acoustic signal enhancement method based on automatic encoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910738375.0A CN110456332B (en) | 2019-08-12 | 2019-08-12 | Underwater acoustic signal enhancement method based on automatic encoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110456332A true CN110456332A (en) | 2019-11-15 |
| CN110456332B CN110456332B (en) | 2022-08-05 |
Family
ID=68485838
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910738375.0A Active CN110456332B (en) | 2019-08-12 | 2019-08-12 | Underwater acoustic signal enhancement method based on automatic encoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110456332B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110956957A (en) * | 2019-12-23 | 2020-04-03 | 苏州思必驰信息科技有限公司 | Training method and system of speech enhancement model |
| CN111968614A (en) * | 2020-08-24 | 2020-11-20 | 湖南工业大学 | Active noise control device of vehicle global space based on convolution-fuzzy network |
| CN112053421A (en) * | 2020-10-14 | 2020-12-08 | 腾讯科技（深圳）有限公司 | Signal noise reduction processing method, device, equipment and storage medium |
| CN112446415A (en) * | 2020-10-09 | 2021-03-05 | 山东中医药大学 | Fusion subtraction automatic encoder algorithm for image feature extraction |
| CN112464837A (en) * | 2020-12-03 | 2021-03-09 | 中国人民解放军战略支援部队信息工程大学 | Shallow sea underwater acoustic communication signal modulation identification method and system based on small data samples |
| CN112986950A (en) * | 2020-12-25 | 2021-06-18 | 南京理工大学 | Single-pulse laser radar echo feature extraction method based on deep learning |
| WO2021164267A1 (en) * | 2020-02-21 | 2021-08-26 | 平安科技（深圳）有限公司 | Anomaly detection method and apparatus, and terminal device and storage medium |
| CN115118557A (en) * | 2022-06-28 | 2022-09-27 | 南华大学 | Underwater sound OFDM communication channel feedback method and system based on deep learning |
| CN112446415B (en) * | 2020-10-09 | 2024-04-30 | 山东中医药大学 | Method for fusion-subtracting automatic encoder for image feature extraction |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107590778A (en) * | 2017-07-31 | 2018-01-16 | 南京理工大学 | A kind of own coding method based on lossless constraint noise reduction |
| CN108537271A (en) * | 2018-04-04 | 2018-09-14 | 重庆大学 | A method of resisting sample is attacked based on convolution denoising self-editing ink recorder defence |
| CN109389171A (en) * | 2018-10-11 | 2019-02-26 | 云南大学 | Medical image classification method based on more granularity convolution noise reduction autocoder technologies |
-
2019
- 2019-08-12 CN CN201910738375.0A patent/CN110456332B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107590778A (en) * | 2017-07-31 | 2018-01-16 | 南京理工大学 | A kind of own coding method based on lossless constraint noise reduction |
| CN108537271A (en) * | 2018-04-04 | 2018-09-14 | 重庆大学 | A method of resisting sample is attacked based on convolution denoising self-editing ink recorder defence |
| CN109389171A (en) * | 2018-10-11 | 2019-02-26 | 云南大学 | Medical image classification method based on more granularity convolution noise reduction autocoder technologies |
Non-Patent Citations (6)
| Title |
|---|
| BO DU等: "Stacked Convolutional Denoising Auto-Encoders for Feature Representation", 《IEEE TRANSACTIONS ON CYBERNETICS》 * |
| MOEIN OWHADI-KARESHK等: "Representation learning by Denoising Autoencoders for Clustering-based Classification", 《2015 5TH INTERNATIONAL CONFERENCE ON COMPUTER AND KNOWLEDGE ENGINEERING (ICCKE)》 * |
| SHAIKH AKIB SHAHRIYAR等: "Speech Enhancement Using Convolutional Denoising Autoencoder", 《2019 INTERNATIONAL CONFERENCE ON ELECTRICAL, COMPUTER AND COMMUNICATION ENGINEERING (ECCE)》 * |
| SUDIPTA SINGHA ROY等: "NOISY IMAGE CLASSIFICATION USING HYBRID DEEP LEARNING METHODS", 《JOURNAL OF INFORMATION AND COMMUNICATION TECHNOLOGY》 * |
| 郭赛: "基于深度学习的海面目标检测研究", 《中国优秀硕士学位论文全文数据库》 * |
| 陈越超 等: "基于降噪自编码器的水中目标识别方法", 《声学与电子工程》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110956957A (en) * | 2019-12-23 | 2020-04-03 | 苏州思必驰信息科技有限公司 | Training method and system of speech enhancement model |
| CN110956957B (en) * | 2019-12-23 | 2022-05-17 | 思必驰科技股份有限公司 | Training method and system of speech enhancement model |
| WO2021164267A1 (en) * | 2020-02-21 | 2021-08-26 | 平安科技（深圳）有限公司 | Anomaly detection method and apparatus, and terminal device and storage medium |
| CN111968614A (en) * | 2020-08-24 | 2020-11-20 | 湖南工业大学 | Active noise control device of vehicle global space based on convolution-fuzzy network |
| CN111968614B (en) * | 2020-08-24 | 2023-09-19 | 湖南工业大学 | Active noise control device of vehicle global space based on convolution-fuzzy network |
| CN112446415A (en) * | 2020-10-09 | 2021-03-05 | 山东中医药大学 | Fusion subtraction automatic encoder algorithm for image feature extraction |
| CN112446415B (en) * | 2020-10-09 | 2024-04-30 | 山东中医药大学 | Method for fusion-subtracting automatic encoder for image feature extraction |
| CN112053421B (en) * | 2020-10-14 | 2023-06-23 | 腾讯科技（深圳）有限公司 | Signal noise reduction processing method, device, equipment and storage medium |
| CN112053421A (en) * | 2020-10-14 | 2020-12-08 | 腾讯科技（深圳）有限公司 | Signal noise reduction processing method, device, equipment and storage medium |
| CN112464837A (en) * | 2020-12-03 | 2021-03-09 | 中国人民解放军战略支援部队信息工程大学 | Shallow sea underwater acoustic communication signal modulation identification method and system based on small data samples |
| CN112986950A (en) * | 2020-12-25 | 2021-06-18 | 南京理工大学 | Single-pulse laser radar echo feature extraction method based on deep learning |
| CN115118557A (en) * | 2022-06-28 | 2022-09-27 | 南华大学 | Underwater sound OFDM communication channel feedback method and system based on deep learning |
| CN115118557B (en) * | 2022-06-28 | 2023-07-25 | 南华大学 | Underwater acoustic OFDM communication channel feedback method and system based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110456332B (en) | 2022-08-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110456332A (en) | A kind of underwater sound signal Enhancement Method based on autocoder | |
| CN110136731A (en) | Empty cause and effect convolution generates the confrontation blind Enhancement Method of network end-to-end bone conduction voice | |
| CN107680611B (en) | Single-channel sound separation method based on convolutional neural network | |
| CN109841226A (en) | A kind of single channel real-time noise-reducing method based on convolution recurrent neural network | |
| CN106529428A (en) | Underwater target recognition method based on deep learning | |
| CN108038471A (en) | A kind of underwater sound communication signal type Identification method based on depth learning technology | |
| CN110246510B (en) | End-to-end voice enhancement method based on RefineNet | |
| CN105488466B (en) | A kind of deep-neural-network and Acoustic Object vocal print feature extracting method | |
| CN106331433A (en) | Video denoising method based on deep recursive neural network | |
| CN113094993B (en) | Modulation signal denoising method based on self-coding neural network | |
| CN105677035A (en) | EEMD (Ensemble Empirical Mode Decomposition) and wavelet threshold based motor imagery electroencephalogram signal denoising method | |
| CN111723701B (en) | Underwater target identification method | |
| CN107967920A (en) | A kind of improved own coding neutral net voice enhancement algorithm | |
| CN109658949A (en) | A kind of sound enhancement method based on deep neural network | |
| CN115602152B (en) | Voice enhancement method based on multi-stage attention network | |
| CN112466320B (en) | Underwater sound signal noise reduction method based on generation countermeasure network | |
| CN115442191B (en) | Communication signal noise reduction method and system based on relative average generation countermeasure network | |
| CN112214929B (en) | Radar interference suppression method for intermittent sampling repeated forwarding type interference | |
| Yang et al. | An adaptive contourlet HMM–PCNN model of sparse representation for image denoising | |
| CN112309411A (en) | Phase-sensitive gated multi-scale void convolutional network speech enhancement method and system | |
| CN107368840A (en) | A kind of underwater manoeuvre Small object recognition methods based on MFCC and SVMs | |
| Galajit et al. | Semi-fragile speech watermarking based on singular-spectrum analysis with CNN-based parameter estimation for tampering detection | |
| CN113935240A (en) | Artificial seismic wave simulation method based on generative confrontation network algorithm | |
| CN112307926A (en) | Acoustic passive ship target classification method based on generation countermeasure network | |
| CN115293214A (en) | Underwater sound target recognition model optimization method based on sample expansion network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |