CN114818789A - Ship radiation noise identification method based on data enhancement - Google Patents
Ship radiation noise identification method based on data enhancement Download PDFInfo
- Publication number
- CN114818789A CN114818789A CN202210362029.9A CN202210362029A CN114818789A CN 114818789 A CN114818789 A CN 114818789A CN 202210362029 A CN202210362029 A CN 202210362029A CN 114818789 A CN114818789 A CN 114818789A
- Authority
- CN
- China
- Prior art keywords
- data
- radiation noise
- convolution
- encoder
- time delay
- 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.)
- Pending
Links
- 230000005855 radiation Effects 0.000 title claims abstract description 95
- 238000000034 method Methods 0.000 title claims abstract description 47
- 238000012549 training Methods 0.000 claims abstract description 103
- 238000009826 distribution Methods 0.000 claims abstract description 73
- 239000013598 vector Substances 0.000 claims abstract description 45
- 238000012360 testing method Methods 0.000 claims abstract description 29
- 238000005070 sampling Methods 0.000 claims abstract description 26
- 230000017105 transposition Effects 0.000 claims abstract description 20
- 238000004422 calculation algorithm Methods 0.000 claims description 21
- 238000009792 diffusion process Methods 0.000 claims description 17
- 238000010606 normalization Methods 0.000 claims description 17
- 239000012634 fragment Substances 0.000 claims description 9
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 230000003190 augmentative effect Effects 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 6
- 230000008034 disappearance Effects 0.000 claims description 4
- 238000004880 explosion Methods 0.000 claims description 4
- 230000006870 function Effects 0.000 description 10
- 238000012545 processing Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 6
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 239000000284 extract Substances 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 238000013136 deep learning model Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 241000165780 Preussia/Sporomiella species complex Species 0.000 description 1
- 238000012952 Resampling Methods 0.000 description 1
- 238000010521 absorption reaction Methods 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000001902 propagating effect Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 239000013535 sea water Substances 0.000 description 1
- 230000005236 sound signal Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a ship radiation noise identification method based on data enhancement, which comprises the following steps: collecting a radiation noise sample containing ship information; windowing the sample to obtain a plurality of small sections of radiation noise, regularizing the windowed radiation noise sample, and attaching corresponding labels to different samples; extracting MFCC characteristics of all samples, and inputting the MFCC characteristics into a time delay convolution encoder of a variational self-encoder data enhancement model to obtain distribution parameters of a corresponding characteristic space; sampling is carried out from the characteristic space distribution to obtain a sampling characteristic vector, and the sampling characteristic vector is input into a transposition time delay convolution decoder of a variational self-encoder data enhancement model to obtain corresponding reconstruction data; training a variational self-encoder data enhancement model, and generating a large amount of generated data after training is finished to expand an original training sample; training a classifier by using the expanded training samples; and carrying out prediction on the test data to obtain a corresponding predicted ship type.
Description
Technical Field
The invention relates to a ship radiation noise identification method, in particular to a ship radiation noise identification method based on data enhancement.
Background
China has a large sea area and abundant ocean resources, and with the rapid development of industries such as modern marine industry and fishery, the research of underwater acoustic signals is more and more emphasized by people. Radiated underwater noise is a common underwater acoustic signal that is mostly generated by a ship traveling through water. Since the radiation noise of a ship can reflect hull information such as the operating speed of the ship, the cargo capacity of the ship, the operating conditions of parts, etc., the radiation noise signal is often used as an important information source for analyzing the ship. Most of the current ship radiation noise analysis work is performed by trained professionals, and the application range of the technology is very limited due to the training cost of the professionals and the efficiency of manual identification. In order to reduce the cost of the radiation noise identification technology and expand the application range of the technology, more and more researchers are put into the research of the radiation noise automatic identification.
However, underwater acoustic signals are more susceptible to underwater channels and to a seemingly mushrooming aquatic environment than signals propagating in the air. Due to random motion of the sea surface, unevenness and variation of the sea bottom with time, uneven distribution of the water body and the like, the underwater acoustic channel is not uniform in space but is randomly variable in time domain. In addition, the underwater acoustic signal has the characteristics of low propagation speed, long code element period and time variation and instability caused by a complex underwater channel. Meanwhile, the marine environment is more diverse than the land, various organisms, seawater movement and noise generated by ship movement exist underwater, the absorption frequencies of different water areas to sound wave signals are different, and the signal propagation speeds in the water areas with different depths are different, so that the signal actually propagated underwater can be interfered by multipath effect, Doppler effect and various noise signals. Due to the variety of problems and the complexity of processing, underwater acoustic signal identification has been a challenging topic.
The existing underwater radiation noise identification research is mostly developed based on a classical signal processing algorithm, the algorithm finishes the work of signal noise reduction and signal feature extraction by constructing a complex signal processing model, and then finishes classification by comparing the similarity of features among different signals. The method has better interpretability and theoretical basis, but is easily influenced by environmental factors, and when an experimental scene changes, the model parameters need to be adjusted by means of knowledge in related fields. In recent years, with the development of deep learning theory and the updating of data computing equipment, the deep neural network technology is rapidly developed, and huge achievements are obtained in the fields of image recognition, voice recognition, natural language processing and the like. More and more scholars now propose the use of deep learning models to build a radiated noise signal recognition model. Compared with the traditional time-frequency analysis method, the deep learning model can extract the nonlinear features with more expressive power, and meanwhile, the recognition performance is improved to a greater degree. However, the deep learning algorithm mostly needs a large-scale data set to support the training of the model, and the radiation noise signal has a very limited experimental data scale due to high acquisition cost, complex acquisition method and other reasons, so that the development of data enhancement research and the expansion of the noise signal data set are very important for improving the recognition performance of the model.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to solve the technical problem of providing a ship radiation noise identification method based on data enhancement aiming at the defects of the prior art.
In order to solve the technical problem, the invention discloses a ship radiation noise identification method based on data enhancement, which comprises the following steps:
step 1, constructing a variational self-encoder data enhancement model, wherein the model consists of a time delay convolution encoder and a transposition time delay convolution decoder; collecting a radiation noise sample containing ship driving information, taking the type of a ship emitting the radiation noise as a class label of the ship, obtaining an original radiation noise data set B1, and dividing a data set B1 into a training set B2 and a test set B3 according to the proportion;
step 2, performing windowing operation on the training set B2 to segment a plurality of fragment signals from the original radiation noise signals, wherein the ship class labels of the fragment signals are the same as the labels of the corresponding original radiation noise signals, and the Mel frequency cepstrum coefficient MFCC characteristics of the fragment signals are used for training a variational self-encoder data enhancement model to obtain a training set B4;
step 3, inputting the data I in the training set B4 into a time delay convolution encoder to obtain the probability distribution of the input data I in the depth characteristic space, wherein the probability distribution is normal distribution, and the time delay convolution encoder outputs the mean value M and the standard deviation S of the probability distribution;
step 4, randomly sampling by using the mean value M and the standard deviation S output by the time delay convolution encoder to obtain a new eigenvector V, inputting the eigenvector V into a transposed time delay convolution decoder, and decoding the eigenvector V layer by layer to obtain reconstructed Mel frequency cepstrum coefficient MFCC data O;
step 6, inputting the training set data into a trained variational self-encoder data enhancement model to generate more reconstruction data, wherein the class labels of the reconstruction data are consistent with the input data, and then expanding a training set B4 by using the reconstruction data to obtain an expanded training set B5;
step 7, training a ResNet-18 classifier by using the expanded training set B5, and saving the result file of the trained classifier as a result file F;
and 8, identifying each radiation noise signal in the test set B3 by using the result file F to obtain the ship category to which the test radiation noise belongs, and finishing the identification of the ship radiation noise based on data enhancement.
In step 1 of the invention, a passive sonar device is used for collecting radiation noise emitted by ships from and to a port, radiation noise raw data, namely radiation noise samples, are obtained, the type of the ship emitting the signal is taken as the category of the radiation noise signal, all the collected radiation noise samples and the category data of the radiation noise signal form a raw radiation noise data set B1, and the data set B1 is proportionally divided into a training set B2 and a test set B3.
The windowing operation in step 2 of the invention comprises: the method comprises the steps of using a windowing method, defining the size of a fixed window as W, dividing original data in a training set B2 into a plurality of windowed signals with the length of the segment as W, wherein the class label of each windowed signal is the class label of the original data to which the windowed signal belongs, extracting corresponding Mel frequency cepstrum coefficient MFCC characteristics from segment signals obtained by windowing, wherein the characteristics are two-dimensional time-frequency characteristics, and forming a training set B4 by utilizing the Mel frequency cepstrum coefficient MFCC characteristics of the segment signals of the window for training a variational self-encoder data enhancement model.
In step 3 of the invention, a Mel frequency cepstrum coefficient I is selected from a training set B4 as the input of a variational self-encoder data enhancement model, the variational self-encoder data enhancement model consists of a time delay convolutional encoder and a transposed time delay convolutional decoder, the input data is firstly learned through a time delay convolutional encoder structure to obtain the characteristic space probability distribution corresponding to the input data, and then the reconstructed data is obtained through the characteristic space sampling vector decoding to complete the task of data enhancement;
the time delay convolution encoder is formed by cascading three time delay convolution units, wherein each time delay convolution unit comprises four parts of convolution, transposition operation, batch normalization and residual connection; the depth features of input data I are extracted through a time delay convolution encoder, the depth features are expanded into one-dimensional features, and the one-dimensional features are input into two independent fully-connected neural networks to obtain a mean value M and a standard deviation S of random distribution of the depth features. .
In step 4 of the invention, random sampling is carried out in multidimensional normal distribution with the mean value M and the standard deviation S obtained from a time delay convolution encoder to obtain a new characteristic vector V, and the following heavy parameter sampling strategy is adopted:
V’～N(0,1)
V＝V’*S+M
wherein N (0,1) represents a standard normal distribution with a mean value of 0 and a standard deviation of 1; firstly, sampling from standard normal distribution N (0,1) to obtain a random vector V' by a multiparameter sampling strategy, and then generating a characteristic vector V which accords with multidimensional normal distribution with a mean value M and a standard deviation S through the characteristic of normal distribution;
the transposition time delay convolution decoder is another component of the variational self-encoder data enhancement model, and is formed by cascading three transposition time delay convolution units, wherein each transposition time delay convolution unit comprises four parts of transposition operation, diffusion convolution, batch normalization and residual connection; and decoding the feature vector V into a feature map by a transposed time delay convolution decoder, and finally obtaining a reconstructed output O with the same size as the input Mel frequency cepstrum coefficient MFCC feature.
In step 5 of the invention, a reconstruction error between input data I and reconstruction output data O and a deviation between depth feature vector probability distribution and standard normal distribution are calculated to obtain a loss function L of the evaluation data enhancement model performance, and the method comprises the following steps:
L＝mse_loss(O,I)-KL(N(M,S^2)|N(0,1))
the mean square error between output O and input I is calculated by a reconstruction error term mse _ loss (O, I) in the loss function, and the closer the output O and the input I are, the smaller the reconstruction error is and the larger the denormalization is; the regularization term KL (N (M, S ^2) | N (0,1)) in the loss function calculates the KL divergence between the depth feature vector probability distribution N (M, S ^2) and the standard normal distribution N (0, 1);
these error values are used to calculate parameter update values for the model using an error back propagation algorithm (ref: RUMELMATHART D E, HINTON G E, WILLIAMS R J. learning representation by back-compressing errors [ J ]. nature,1986,323(6088):533 @.) and then parameter update is performed on the variation self-encoder data enhancement model by a random small-batch gradient descent algorithm (ref: Li M, Zhang T, Chen Y, et al. effective mini-batch tracking for storage optimization [ C ]/Proceedings of the 20th ACM SIGKDD internal control Knowledge and data mining 2014: 661. 670.).
In step 6 of the present invention, each data in the training set B4 is input into a variational self-encoder data enhancement model, a delay convolution encoder of the model outputs probability distribution of the data in a feature space, a feature vector is obtained based on the probability distribution sampling in the feature space, the feature vector is input into a transposed delay convolution decoder, reconstructed data of the same type as the original data is generated, the training set B4 is extended by using the reconstructed data, and an extended training set B5 is obtained.
In step 7 of the invention, the expanded training set B5 is utilized, the data in the training set B5 is input into a ResNet-18 classifier in combination with the corresponding sample class label thereof for training, the deviation between the output prediction class label and the real sample class label in the training process is calculated by using a cross entropy loss function, the updated value of the ResNet-18 model parameter is calculated by a back propagation algorithm, the parameter of the ResNet-18 model is updated by using a random gradient descent algorithm, and multiple rounds of training are repeated until the change trend of the cross entropy loss function value is converged; and after the trained ResNet-18 classifier is obtained, saving the training result as a result file F.
In step 8 of the invention, the ResNet-18 classifier obtained in step 7 is used for identifying the test signals in the test set B3, firstly, windowing operation is carried out on the original test signals to obtain segmented radiation noise signals identical to training data, then the Mel frequency cepstrum coefficient MFCC characteristics of the segmented signals are extracted, and the characteristics are input into the ResNet-18 classifier for prediction to obtain the ship category to which the segmented radiation noise signals belong.
The transposed time delay convolution process in step 4 of the present invention comprises:
transposition operation: this operation exchanges the channel dimensions and high dimensions of the input feature vectors;
diffusion convolution: diffusing each element of the transposed feature vector into a matrix with the same size of a transposed convolution kernel by the diffusion convolution operation; then calculating convolution results between the diffusion matrix and the transposed convolution kernel, repeating the operation on each element in the input characteristics to obtain a series of intermediate results of convolution operation, and then performing matrix addition on the intermediate results to obtain a result of diffusion convolution;
batch normalization: the vector distribution drift caused by multilayer nonlinear operation is relieved, a batch normalization layer is accessed after each diffusion convolution, and the characteristic distribution of batch training data is dynamically adjusted;
residual connection: and a residual connecting branch is added between the input of the transposition time delay convolution and the output of the batch normalization layer, so that the gradient disappearance and gradient explosion of a deep model are relieved.
Has the advantages that:
the invention overcomes the problem that the training data of the current radiation noise identification model based on deep learning is insufficient, constructs a variational self-encoder data enhancement model based on a time delay convolution structure, extracts the frame dimension depth characteristic of radiation noise through a time delay convolution module, generates enhancement data by using a data reconstruction mechanism of the variational self-encoder data enhancement model, and completes the extension of a training data set. The method effectively improves the scale of the original training data set and improves the identification precision of the radiation noise identification model.
The method has the obvious advantages of expanding the scale of the radiation noise data set, effectively relieving the problem of insufficient data of the deep learning model and improving the robustness and the recognition performance of the recognition model.
Drawings
The foregoing and/or other advantages of the invention will become further apparent from the following detailed description of the invention when taken in conjunction with the accompanying drawings.
FIG. 1 is a flow chart of the operation of the system of the present invention.
Fig. 2 is a graph of the effect of the algorithm on the radiated noise data set a in the present invention.
Fig. 3 is a graph of the effect of the algorithm of the present invention on the radiated noise data set B.
FIG. 4 is a comparison of a heat map of generated MFCC data obtained from the time-delayed convolutional variational self-encoder data enhancement model in accordance with the present invention with the corresponding original MFCC data.
Detailed Description
The invention is further explained below with reference to the drawings and the embodiments.
Fig. 1 is a flow chart showing the operation of the system of the present invention, which includes 8 steps.
Step 1, collecting a radiation noise sample containing ship driving information, taking the type of a ship emitting the radiation noise as a class label of the ship, obtaining an original radiation noise data set B1, and performing data analysis according to the following steps of 2: 1, dividing the data set B1 into a training set B2 and a testing set B3;
step 2, performing windowing operation on the training set B2 to segment a plurality of fragment signals from the original radiation noise signals, wherein the ship class labels of the fragment signals are the same as the corresponding original signal labels, and the Mel cepstrum coefficients (MFCC) features of the fragment signals form a data set B4;
step 3, inputting the data I of the data set B4 into a time delay convolution encoder to obtain the probability distribution of the input data in the depth characteristic space, wherein the probability distribution is normal distribution, and the encoder outputs the mean value M and the standard deviation S of the distribution;
step 4, randomly sampling by using the mean value M and the standard deviation S output by the time delay convolution encoder to obtain a new feature vector V, inputting the V into a transposed time delay convolution decoder, and decoding the feature vector layer by layer to obtain reconstructed Mel frequency cepstrum coefficient MFCC data O;
step 6, inputting the training set data into a trained variational self-encoder data enhancement model to generate more reconstruction data, wherein the class labels of the reconstruction data are consistent with the input data, and then expanding a training set B4 by using the reconstruction data to obtain an expanded training set B5;
step 7, training a ResNet-18 classifier by using the expanded training set B5, and saving the result file of the trained classifier as a result file F;
and 8, identifying each radiation noise signal in the test set B3 by using the file F to obtain the ship category to which the test radiation noise belongs.
In step 1, passive sonar equipment is used for collecting radiation noise emitted by ships from and to ports, and original radiation noise data are obtained, the ship radiation noise is a sound signal generated after the combined action of various noise sources on the ships and water media where the ship radiation noise is located, and the main radiation noise sources comprise: propellers, rotating and reciprocating machines, various pumps, etc. Identifying the type of vessel to which the radiated noise belongs is a common type of underwater acoustic signal processing task. After the data acquisition is completed, the type of ship emitting the signal is taken as the category of the radiated noise signal, and the acquired data is composed into a raw radiated noise data set B1, according to 2: the raw radiation noise data in data set B1 was scaled by a scale of 1, resulting in training set B2 and test set B3.
In step 2, a windowing method is used, because the length of the acquired original radiation noise signal is generally long, the complexity of directly using the acquired original radiation noise signal as an input of a classification model is too high. Thus, here, a fixed window size W is used, (Mel frequency cepstrum coefficients, MFCCs) specifies that the fixed window size W is 10000, and the raw radiated noise data in the training set B2 is divided into several windowed segments of length W, each of which has a class label of the raw data to which it belongs. MFCC features of the windowed segmented noise signal are then extracted, the MFCC features are two-dimensional time-frequency features which describe the acoustic features of the signal and are often used for acoustic signal recognition tasks. The labeled MFCC features are used to form a training set B4. (reference: Davis S, Mermelstein P. Complex of parametric representation for monomeric word recognition in constitutive speech sensing [ J ]. IEEE transactions on optics, speed, and signal processing,1980,28(4): 357-
In step 3, data is selected from a training set B4 as an input I of a variational self-encoder data enhancement model, the time delay convolution encoder structure extracts the depth characteristics of the data through a time delay convolution unit, and the time delay convolution unit comprises four parts of convolution, transposition operation, batch normalization and residual error connection.
1. And (4) convolution. In the conventional convolutional neural network model, a convolution kernel of 3 × 3 size is mostly used, because the convolution kernel of this size can better capture the detail texture in the picture. However, for the ship radiation noise, the signal sampling frequency is concentrated between 8kHz and 16kHz, and at this time, the field of view corresponding to the convolution characteristics obtained by the convolution kernel with the size of 3 × 3 is too small to describe the signal characteristics well. The time delay convolution is an extended structure based on time domain convolution, and the height of a kernel of the time delay convolution is a fixed value and is the same as the height of an input characteristic. The time delay convolution adopts the design of a large convolution kernel, and the size of the receptive field corresponding to the convolution characteristic is improved.
2. And (5) transposition operation. In the output of the time delay convolution, the channel dimension and the high dimension of the output characteristic diagram are exchanged, so that the time-frequency distribution characteristic of the input data is still obtained after the time delay convolution.
3. And (4) batch normalization. Because the convolution kernel size used by the time delay convolution is large, in order to relieve the problem of difficult parameter optimization caused by a large convolution kernel, a batch normalization layer is accessed after the convolution kernel transposition operation to dynamically adjust the characteristic distribution of batch training data.
4. And residual connecting. A residual connecting branch is added between the input of the time delay convolution unit and the output of the batch normalization layer, so that the problems of gradient disappearance and gradient explosion of a deep model are solved.
And then expanding the depth time delay convolution characteristic into a one-dimensional characteristic, inputting the one-dimensional characteristic into two independent fully-connected neural networks, and obtaining the mean value M and the standard deviation S of the random distribution of the depth characteristic. (reference: Kingma D P, Welling M. auto-encoding variant weights [ J ]. arXiv preprinting arXiv:1312.6114,2013)
In step 4, random sampling is performed in multidimensional normal distribution with the mean value M and the standard deviation S to obtain a new feature vector V. Because the random sampling operation is not integrable, in order to update the network parameters of the encoder part in the subsequent parameter updating stage, the following heavy parameter sampling strategy is adopted:
V’～N(0,1)
V＝V’*S+M
where N (0,1) represents a standard normal distribution with a mean of 0 and a standard deviation of 1. Firstly, sampling from standard normal distribution N (0,1) to obtain a random vector V' by a resampling sampling strategy, and then generating a characteristic vector V which accords with multidimensional normal distribution with the mean value M and the standard deviation S through the characteristic of normal distribution;
the transposed time delay convolutional decoder consists of a series of transposed time delay convolutional layers with input and output characteristics opposite to those of the time delay convolutional layers in the time delay convolutional encoder, so that the feature vector generated by the encoder can be decoded step by step to obtain reconstructed output. By transposing the time delay convolution encoder structure, the feature vector V is gradually decoded into a feature map with a larger size, and finally a reconstructed output O with the same size as the input MFCC feature is obtained. The transposition time delay convolution is the reverse process of the time delay convolution and mainly comprises the following steps:
1. and (5) transposition operation. This operation exchanges the channel dimensions and high dimensions of the input feature vectors;
2. and (4) diffusion convolution. The diffusion convolution operation diffuses each element of the transposed feature vector into a matrix of the same size as the transposed convolution kernel. Then calculating convolution results between the diffusion matrix and the transposed convolution kernel, repeating the operation on each element in the input characteristics to obtain a series of intermediate results of convolution operation, and then performing matrix addition on the intermediate results to obtain a result of diffusion convolution;
3. and (4) batch normalization. In order to alleviate the problem of vector distribution drift caused by multilayer nonlinear operation, a batch normalization layer is accessed after each diffusion convolution to dynamically adjust the characteristic distribution of batch training data;
4. and residual connecting. A residual connecting branch is added between the input of the transposed delay convolution and the output of the batch normalization layer, so that the problems of gradient disappearance and gradient explosion of a deep model are solved.
In step 5, a reconstruction error between the input data I and the reconstruction output data O and a deviation between the depth feature vector probability distribution and the standard normal distribution are calculated to obtain a loss function of the evaluation data enhancement model performance:
L＝mse_loss(O,I)-KL(N(M,S^2)|N(0,1))
the mean square error between the output O and the input I is calculated by a reconstruction error term mse _ loss (O, I) in the loss function, and the closer the output O and the input I are, the smaller the reconstruction error is, and the larger the regularization is; the regular term KL (N (M, S ^2) | N (0,1)) in the loss function calculates the Kullback-Leibler divergence between the depth feature vector probability distribution N (M, S ^2) and the standard normal distribution N (0,1), which is called KL divergence for short, and the index reflects the difference degree between the two distributions;
an error back-propagation algorithm is used to calculate the parameter update values of the model using these error values, and then the variational self-encoder data enhancement model is parameter updated by a random small batch gradient descent algorithm.
In step 6, the MFCC characteristics of each radiation noise signal in the training set B4 are input into a variational self-encoder data enhancement model, the time-delay convolution encoder structure of the data enhancement model firstly outputs the probability distribution of the data in a characteristic space, then a characteristic vector is obtained in the characteristic space based on the probability distribution sampling, the characteristic vector is input into a transposed time-delay convolution decoder to generate reconstructed data of the same type as the original data, and the reconstructed data is used for expanding the training set B3 to obtain an expanded training set B5.
In step 7, the data in the expanded training set B5 and the corresponding sample label are input into a ResNet-18 (18-layer Residual convolutional neural Network-18, ResNet-18) classifier for training. Calculating the deviation between a prediction label output by ResNet-18 and a real sample label in the training process by using a cross entropy loss function, calculating an updated value of a ResNet-18 model parameter by using a back propagation algorithm, updating the parameter of the ResNet-18 model by using a random gradient descent algorithm, and repeating multiple rounds of training until the change trend of the cross entropy loss function value is converged. And after the trained ResNet-18 classifier is obtained, saving the training result as a result file F.
In step 8, signals in the test set B3 are identified by using the ResNet-18 (18-layer Residual convolutional neural Network-18, ResNet-18) classifier obtained in step 7, firstly, windowing operation is carried out on an original test signal to obtain a segmented radiation noise signal which is the same as training data, then MFCC characteristics of the segmented signal are extracted, and the segmented radiation noise signal is input into the ResNet-18 classifier for prediction to obtain a ship class to which the segmented radiation noise signal belongs. (reference: He K, Zhang X, Ren S, et al. deep residual learning for image recognition [ C ]// Proceedings of the IEEE conference on computer vision and pattern recognition.2016: 770-778.)
Examples
In order to carry out preprocessing before system operation, the method needs to train a system algorithm model, and a training set is radiation noise signals of different ship types.
The invention uses passive sonar device to collect the radiation noise sent by the ship at different port, then carries on manual label of ship category, finally obtains two groups of radiation noise data sets with ship category label information, each group of data sets includes 50 original radiation noise signals.
The present invention extracts 33% of the original radiated noise signal from the two radiated noise data sets, respectively, as a test signal. For each test data set, processing is carried out by using the steps shown in the figure one, ship class detection is carried out by using a trained ResNet-18 classifier, and the class prediction accuracy on the test set is calculated by comparing the real noise signal ship class with the predicted noise ship class.
Training and evaluating a system model by using the radiation noise training set and the radiation noise test set according to the following steps:
1. model training based on radiated noise signals:
1.1, using a windowing method, defining a fixed window size W to be 10000, dividing original radiation noise data in a training set into a plurality of windowed signals with the length of W, wherein the class label of each windowed signal is the class label of the original data to which the windowed signal belongs, and then extracting the MFCC (Mel frequency cepstrum coefficient) characteristics of each windowed segmented signal to form a training data set of a model;
1.2, the scale of the sample is normalized, namely the size of the sample is the same, so that the influence of the training sample scale on model training can be eliminated, and different types of labels are converted into numerical values, such as 0,1, 2;
1.3, using the processed sample set as the input of a variational self-encoder data enhancement model, extracting the depth characteristics of input data through a time delay convolution encoder structure, then expanding the depth time delay convolution characteristics into one-dimensional characteristics, and inputting two independent fully-connected neural networks to obtain a mean value M and a standard deviation S of random distribution of the depth characteristics;
and 1.4, randomly sampling from multi-dimensional normal distribution with the mean value of M and the standard deviation of S to obtain a new feature vector, gradually decoding the feature vector into a feature map with a larger size through a transposed time delay convolution encoder structure, and finally obtaining the reconstructed output with the same size as the input data.
1.5, calculating a reconstruction error between input data and reconstruction output data and a deviation between depth feature vector probability distribution and standard normal distribution, and performing parameter updating on a variational self-encoder data enhancement model by utilizing parameter updating values of the error item calculation model;
1.6 inputting the training set data into a trained variational self-encoder data enhancement model to generate more reconstruction data, wherein the class labels of the reconstruction data are consistent with the input data, and then expanding the training set by using the reconstruction data;
1.7 training a ResNet-18 classifier by using the expanded training set, and storing model parameters obtained by training.
2. Testing
2.1 for each original signal in the test set, firstly performing windowing processing same as training data to obtain a series of windowed segmented signals;
2.2 extracting the MFCC characteristics of each windowed segmented signal, and then taking the MFCC characteristics as the input of a ResNet-18 classifier to predict to obtain a ship class label;
2.3 calculate the accuracy and AUC value (ROC area of Curve, Aer Under ROC currve, AUC) of the ResNet-18 classifier on the test set by comparing the predicted label and the true label of the windowed segmented signal.
2.3 for the tag class of the original signal, the number of tags of each windowed segment signal contained in the original signal can be counted by a voting method, and the ship class tag with the largest occurrence frequency is taken as the predicted tag of the original signal.
Based on the training and testing steps, a ship radiation noise data enhancement and identification system is finally obtained, and the accuracy of radiation noise identification based on the data enhancement method reaches over 75 percent. And the defect that the traditional deep learning method is poor in robustness under a small data set is overcome by using the data enhancement strategy. Therefore, the method is applied to the identification of the radiation noise of the ship, and has the advantages of good robustness and high prediction accuracy.
Fig. 2 lists the case where the identification algorithm part of the present invention identifies the type of segmented signal on the inventive acquired radiated noise data set 1. The results show that the invention has excellent performance in terms of statistical accuracy. Some indices in the table have the following meanings: "Algorithm" indicates different recognition models, FCN stands for full convolution neural network classifier, ResNet-18 stands for residual convolution neural network structure used in the present invention; "data enhancement" indicates different data enhancement methods, "none" indicates an original training set that is not augmented by the data enhancement method, "MFCC" indicates a training set that is augmented by the data enhancement method that directly imposes a disturbance on the MFCC feature, "wavelet" indicates a training set that is augmented by the data enhancement method that imposes a disturbance on the original signal in the wavelet domain using the discrete wavelet decomposition method, and "VAE-tdc" indicates a training set that is augmented by the variational self-encoder data enhancement model proposed by the present invention. Auccuary and AUC represent two measurement indexes, namely accuracy and ROC curve area, and the larger the two indexes, the better the identification effect of the code model.
Fig. 3 lists the case where the identification algorithm portion of the present invention identifies the type of segmented signal on the radiated noise data set 2 acquired by the present invention. The meaning of each index of the experimental result is consistent with that of fig. 2, and fig. 2 illustrates that the variational self-encoder data enhancement model provided by the invention has a remarkable improvement effect on different actually-measured radiation noise signals, thereby further proving the effectiveness thereof.
Fig. 4 shows a comparison diagram of a heat map corresponding to reconstructed MFCC data obtained through the variational self-encoder data enhancement model and original MFCC data, which can be clearly seen that enhanced data generated by the variational self-encoder data enhancement model has richer changes compared with the original data, and the enhanced data retains the distribution characteristics of the original data, which is very advantageous for improving the robustness and the identification accuracy of the identification model.
The invention provides a method and a concept for identifying ship radiation noise based on data enhancement, and a method and a way for implementing the technical scheme are many, the above description is only a preferred embodiment of the invention, and it should be noted that, for those skilled in the art, a plurality of improvements and embellishments can be made without departing from the principle of the invention, and these improvements and embellishments should also be regarded as the protection scope of the invention. All the components not specified in the present embodiment can be realized by the prior art.
Claims (10)
1. A ship radiation noise identification method based on data enhancement is characterized by comprising the following steps:
step 1, constructing a variational self-encoder data enhancement model, wherein the model consists of a time delay convolution encoder and a transposition time delay convolution decoder; collecting a radiation noise sample containing ship driving information, taking the type of a ship emitting the radiation noise as a class label of the ship, obtaining an original radiation noise data set B1, and dividing a data set B1 into a training set B2 and a test set B3 according to the proportion;
step 2, performing windowing operation on the training set B2 to segment a plurality of fragment signals from the original radiation noise signals, wherein the ship class labels of the fragment signals are the same as the labels of the corresponding original radiation noise signals, and the Mel frequency cepstrum coefficient MFCC characteristics of the fragment signals are used for training a variational self-encoder data enhancement model to obtain a training set B4;
step 3, inputting the data I in the training set B4 into a time delay convolution encoder to obtain the probability distribution of the input data I in the depth characteristic space, wherein the probability distribution is normal distribution, and the time delay convolution encoder outputs the mean value M and the standard deviation S of the probability distribution;
step 4, randomly sampling by using the mean value M and the standard deviation S output by the time delay convolution encoder to obtain a new eigenvector V, inputting the eigenvector V into a transposed time delay convolution decoder, and decoding the eigenvector V layer by layer to obtain reconstructed Mel frequency cepstrum coefficient MFCC data O;
step 5, calculating a reconstruction error between the input data I and the reconstruction output data O and a deviation between the depth characteristic vector probability distribution and the standard normal distribution, calculating a parameter updating value of the variational self-encoder data enhancement model by using the errors and the deviation items, and updating corresponding parameters in the variational self-encoder data enhancement model by using the parameter updating value;
step 6, inputting the training set data into a trained variational self-encoder data enhancement model to generate more reconstruction data, wherein the class labels of the reconstruction data are consistent with the input data, and then expanding a training set B4 by using the reconstruction data to obtain an expanded training set B5;
step 7, training a ResNet-18 classifier by using the expanded training set B5, and saving the result file of the trained classifier as a result file F;
and 8, identifying each radiation noise signal in the test set B3 by using the result file F to obtain the ship category to which the test radiation noise belongs, and finishing the identification of the ship radiation noise based on data enhancement.
2. The data-based ship radiation noise identification method for enhancement is characterized in that in the step 1, passive sonar equipment is used for collecting radiation noise emitted by ships in ports, radiation noise raw data, namely radiation noise samples, is obtained, the type of the ship emitting the signal is taken as the category of the radiation noise signal, all the collected radiation noise samples and the category data of the radiation noise signal are combined into a raw radiation noise data set B1, and the data set B1 is divided into a training set B2 and a testing set B3 according to proportion.
3. The ship radiation noise identification method based on data enhancement according to claim 2, wherein the windowing operation in step 2 comprises: the method comprises the steps of using a windowing method, defining the size of a fixed window as W, dividing original data in a training set B2 into a plurality of windowed signals with the length of the segment as W, wherein the class label of each windowed signal is the class label of the original data to which the windowed signal belongs, extracting corresponding Mel frequency cepstrum coefficient MFCC characteristics from segment signals obtained by windowing, wherein the characteristics are two-dimensional time-frequency characteristics, and forming a training set B4 by utilizing the Mel frequency cepstrum coefficient MFCC characteristics of the segment signals of the window for training a variational self-encoder data enhancement model.
4. The data enhancement-based ship radiation noise identification method according to claim 3, wherein in step 3, a Mel frequency cepstrum coefficient I is selected from a training set B4 as an input of a variational self-encoder data enhancement model, the variational self-encoder data enhancement model is composed of a time delay convolution encoder and a transposed time delay convolution decoder, the input data is firstly learned through a time delay convolution encoder structure to obtain a characteristic space probability distribution corresponding to the input data, and then is decoded through a characteristic space sampling vector to obtain reconstructed data, so that a data enhancement task is completed;
the time delay convolution encoder is formed by cascading three time delay convolution units, wherein each time delay convolution unit comprises four parts of convolution, transposition operation, batch normalization and residual connection; the depth features of input data I are extracted through a time delay convolution encoder, the depth features are expanded into one-dimensional features, and the one-dimensional features are input into two independent fully-connected neural networks to obtain a mean value M and a standard deviation S of random distribution of the depth features.
5. The method for identifying the ship radiation noise based on the data enhancement as claimed in claim 4, wherein in the step 4, random sampling is performed in the multidimensional normal distribution with the mean value M and the standard deviation S obtained from the time delay convolution encoder to obtain a new feature vector V, and the following re-parameter sampling strategy is adopted:
V’～N(0,1)
V＝V’*S+M
wherein N (0,1) represents a standard normal distribution with a mean value of 0 and a standard deviation of 1; firstly, sampling from standard normal distribution N (0,1) to obtain a random vector V' by a multiparameter sampling strategy, and then generating a characteristic vector V which accords with multidimensional normal distribution with a mean value M and a standard deviation S through the characteristic of normal distribution;
the transposed time delay convolution decoder is another component of the variational self-encoder data enhancement model, and is formed by cascading three transposed time delay convolution units, wherein each transposed time delay convolution unit comprises four parts of transposition operation, diffusion convolution, batch normalization and residual connection; and decoding the feature vector V into a feature map by a transposed time delay convolution decoder, and finally obtaining a reconstructed output O with the same size as the input Mel frequency cepstrum coefficient MFCC feature.
6. The method for recognizing the radiation noise of the ship based on the data enhancement as claimed in claim 5, wherein in the step 5, the reconstruction error between the input data I and the reconstruction output data O and the deviation between the depth eigenvector probability distribution and the standard normal distribution are calculated to obtain the loss function L for evaluating the performance of the data enhancement model, and the method comprises the following steps:
L＝mse_loss(O,I)-KL(N(M,S^2)|N(0,1))
the mean square error between the output O and the input I is calculated by a reconstruction error term mse _ loss (O, I) in the loss function, and the closer the output O and the input I are, the smaller the reconstruction error is and the larger the anti-regularization is; the regularization term KL (N (M, S ^2) | N (0,1)) in the loss function calculates the KL divergence between the depth feature vector probability distribution N (M, S ^2) and the standard normal distribution N (0, 1);
an error back-propagation algorithm is used to calculate the parameter update values of the model using these error values, and then the variational self-encoder data enhancement model is parameter updated by a random small batch gradient descent algorithm.
7. The method as claimed in claim 6, wherein in step 6, each piece of data in the training set B4 is input into a variational self-encoder data enhancement model, a time-delay convolutional encoder of the model outputs a probability distribution of the data in a feature space, a feature vector is obtained by sampling in the feature space based on the probability distribution, the feature vector is input into a transposed time-delay convolutional decoder, reconstructed data of the same category as the original data is generated, and the training set B4 is augmented by the reconstructed data, so that an augmented training set B5 is obtained.
8. The ship radiation noise identification method based on data enhancement as claimed in claim 7, characterized in that, in step 7, the augmented training set B5 is used, the data in the training set B5 is input into a ResNet-18 classifier in combination with the corresponding sample class label for training, a cross entropy loss function is used to calculate the deviation between the predicted class label output in the training process and the real sample class label, the updated value of the ResNet-18 model parameter is calculated through a back propagation algorithm, the parameter of the ResNet-18 model is updated through a random gradient descent algorithm, and multiple rounds of training are repeated until the variation trend of the cross entropy loss function is converged; and after the trained ResNet-18 classifier is obtained, saving the training result as a result file F.
9. The method for recognizing the radiation noise of the ship based on the data enhancement as claimed in claim 8, wherein in step 8, the ResNet-18 classifier obtained in step 7 is used for recognizing the test signals in the test set B3, the original test signals are firstly subjected to windowing operation to obtain segmented radiation noise signals identical to the training data, then mel-frequency cepstrum coefficients MFCC features of the segmented signals are extracted and input into the ResNet-18 classifier for prediction, and the class of the ship to which the segmented radiation noise signals belong is obtained.
10. The method for identifying the ship radiation noise based on the data enhancement as claimed in claim 9, wherein the transposed time delay convolution process in the step 4 comprises:
transposition operation: this operation exchanges the channel dimensions and high dimensions of the input feature vectors;
diffusion convolution: diffusing each element of the transposed feature vector into a matrix with the same size of a transposed convolution kernel by the diffusion convolution operation; then calculating convolution results between the diffusion matrix and the transposed convolution kernel, repeating the operation on each element in the input characteristics to obtain a series of intermediate results of convolution operation, and then performing matrix addition on the intermediate results to obtain a result of diffusion convolution;
batch normalization: the vector distribution drift caused by multilayer nonlinear operation is relieved, a batch normalization layer is accessed after each diffusion convolution, and the characteristic distribution of batch training data is dynamically adjusted;
residual connection: and a residual connecting branch is added between the input of the transposition time delay convolution and the output of the batch normalization layer, so that the gradient disappearance and gradient explosion of a deep model are relieved.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210362029.9A CN114818789A (en) | 2022-04-07 | 2022-04-07 | Ship radiation noise identification method based on data enhancement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210362029.9A CN114818789A (en) | 2022-04-07 | 2022-04-07 | Ship radiation noise identification method based on data enhancement |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114818789A true CN114818789A (en) | 2022-07-29 |
Family
ID=82534857
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210362029.9A Pending CN114818789A (en) | 2022-04-07 | 2022-04-07 | Ship radiation noise identification method based on data enhancement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114818789A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117271969A (en) * | 2023-09-28 | 2023-12-22 | 中国人民解放军国防科技大学 | Online learning method, system, equipment and medium for individual fingerprint characteristics of radiation source |
-
2022
- 2022-04-07 CN CN202210362029.9A patent/CN114818789A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117271969A (en) * | 2023-09-28 | 2023-12-22 | 中国人民解放军国防科技大学 | Online learning method, system, equipment and medium for individual fingerprint characteristics of radiation source |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110245608B (en) | Underwater target identification method based on half tensor product neural network | |
| CN110491416B (en) | Telephone voice emotion analysis and identification method based on LSTM and SAE | |
| CN112364779B (en) | Underwater sound target identification method based on signal processing and deep-shallow network multi-model fusion | |
| CN111709315A (en) | Underwater acoustic target radiation noise identification method based on field adaptation | |
| CN110853630B (en) | Lightweight speech recognition method facing edge calculation | |
| CN113361636B (en) | Image classification method, system, medium and electronic device | |
| CN109581339B (en) | Sonar identification method based on automatic adjustment self-coding network of brainstorming storm | |
| CN111627419A (en) | Sound generation method based on underwater target and environmental information characteristics | |
| Wei et al. | A method of underwater acoustic signal classification based on deep neural network | |
| CN113111786B (en) | Underwater target identification method based on small sample training diagram convolutional network | |
| CN112559797A (en) | Deep learning-based audio multi-label classification method | |
| CN114863938A (en) | Bird language identification method and system based on attention residual error and feature fusion | |
| CN114818789A (en) | Ship radiation noise identification method based on data enhancement | |
| Mustika et al. | Comparison of keras optimizers for earthquake signal classification based on deep neural networks | |
| Xu et al. | Self-supervised learning–based underwater acoustical signal classification via mask modeling | |
| CN106448660A (en) | Natural language fuzzy boundary determining method with introduction of big data analysis | |
| Jiang et al. | Speech emotion recognition method based on improved long short-term memory networks | |
| CN112052880A (en) | Underwater sound target identification method based on weight updating support vector machine | |
| CN117310668A (en) | Underwater sound target identification method integrating attention mechanism and depth residual error shrinkage network | |
| Ju et al. | A new low SNR underwater acoustic signal classification method based on intrinsic modal features maintaining dimensionality reduction | |
| CN116343825A (en) | Sample mapping-based few-sample underwater sound target identification method | |
| Xie et al. | Data augmentation and deep neural network classification based on ship radiated noise | |
| Li et al. | A position weighted information based word embedding model for machine translation | |
| CN113488069A (en) | Method and device for quickly extracting high-dimensional voice features based on generative countermeasure network | |
| Ashok et al. | An improving recognition accuracy of underwater acoustic targets based on gated recurrent unit (GRU) neural network method |
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 |