CN118094336A - Underwater sound signal identification method based on integrated learning and sample synthesis - Google Patents

Abstract

The application belongs to the technical field of underwater acoustic signal processing, and particularly relates to an underwater acoustic signal identification method based on integrated learning and sample synthesis. The method comprises the following steps: collecting underwater sound signal sample data and manufacturing an initial data set; collecting underwater sound background sample data and manufacturing a background noise data set; iteratively training a weak classifier; constructing a random initialized neural network to perform repeated training to obtain a weak classifier; calculating the data set by using a weak classifier, and calculating total errors and weights; adding a sample to be identified by a synthesis method to generate a new data set; combining the weak classifiers into a strong classifier Q in a weighting way; the application trains a plurality of weak classifiers for classifying underwater acoustic signals by adopting an integrated learning method, and integrates the weak classifiers in a weighting mode, thereby obtaining a classifier with more powerful functions. Especially, the recognition errors of all weak classifiers are considered in the process of designing weights, so that the recognition capability of the integrated classifier is further improved.

Description

Underwater sound signal identification method based on integrated learning and sample synthesis

Technical Field

The invention belongs to the technical field of underwater acoustic signal processing, and particularly relates to an underwater acoustic signal identification method based on integrated learning and sample synthesis.

Background

With the increasing activity of human exploration of the ocean, marine environmental awareness has become a popular research direction. Because seawater has strong attenuation on electromagnetic signals, the electromagnetic signals cannot be adopted to carry out long-distance information transmission underwater, and the activities of human beings in the ocean are seriously hindered. Fortunately, sound waves have better transmission capability in sea water as a mechanical wave, so that the sound waves become a main mode of current underwater information transmission and information perception.

In a seawater environment, different targets produce acoustic signals of different characteristics. The classification of underwater sound targets and clutter plays an important role in fishery production, marine information industry, scientific investigation and research and national defense military. However, the ocean has a lot of background noise at any time due to the wave, tide, ocean current, etc., and the recognition of the water supply sound signal is very difficult. Accurate identification of underwater acoustic signals is a very challenging task that has become an important issue to be addressed in marine environmental awareness. The underwater sound signal identification comprises the stages of data acquisition, feature extraction, classification identification and the like. Because of the complex environment and the very rare objects, the acquisition of the underwater acoustic signals is very difficult, and the obtained sample library also usually has the problem of unbalanced categories, namely, a plurality of samples in certain categories and a very small number of samples in certain rare categories. On the other hand, the background noise of the underwater sound signal is large, so that the signal characteristics are difficult to extract, the prior cognitive deviation exists in the traditional manual characteristic extraction method, and certain distinguishing characteristics are possibly omitted, so that the recognition accuracy is low.

Disclosure of Invention

Aiming at the existing problems, the invention provides an integrated learning based on boost, which is used as a basic recognition framework, and a plurality of individual learners are trained successively to improve the recognition accuracy of signals; the method is characterized in that the category with low recognition rate is effectively processed, the number of samples is increased by adopting a signal synthesis method with adjustable parameters, and the method is used for recognizing the underwater sound signals based on the characteristic of multiple sound signals as the input of an individual learner so as to improve the richness and the diversity of information.

In order to achieve the above purpose, the present invention adopts the following technical scheme.

1. An underwater sound signal identification method based on integrated learning and sample synthesis comprises the following steps:

Step 1: collecting underwater sound signal sample data and manufacturing an initial data set D; wherein the data set D contains C categories of underwater acoustic signals;

step 2: collecting underwater sound background sample data and manufacturing a background noise data set Z; wherein, the underwater sound background sample does not contain a target sound signal;

Step 3: iteratively training T weak classifiers; the method specifically comprises the following steps: initializing operation, namely constructing a data set D ₁ = D required by the 1 st weak classifier; training the 2 nd to the T th weak classifiers and corresponding training sets D ₂ to D _T in turn on the basis; wherein, the training process of the t weak classifier is as follows from 3.1 to 3.4:

3.1: constructing a random initialized neural network N _t, wherein the neural network adopts a convolutional neural network, and parameters of the neural network N _t are randomly set;

3.2: randomly sampling samples from the dataset D _t to train the neural network N _t;

3.2.1: randomly extracting a sample x from the data set D _t, and calculating a Mel spectrum m of the x;

3.2.2: inputting m to a neural network N _t, and adjusting the parameter value of N _t by using a back propagation algorithm;

3.2.3: repeating training based on the steps 2.2.1-2.2.2 to obtain a neural network Nt, namely a t weak classifier F _t;

3.3: performing one-time operation on all samples in the data set D _t by using a weak classifier F _t, calculating the total error e _t of F _t on the data set D _t, and calculating the total weight a _t of F _t on the final strong classifier by using e _t;

e _t is calculated as follows:

Wherein, P () represents a desire, I () is an indication function, x _i represents an I-th sample in the data set D _t, and y _i represents a class label corresponding to the I-th sample;

the calculation formula of a _t is as follows:

Step 3.4: adding a sample to be identified by a synthesis method to generate a new data set D _t+1; specifically:

3.4.1: for one sample x in the data set D _t, carrying out reasoning and prediction operation on the sample x by using a classifier Ft, wherein the last layer of the neural network of F _t outputs a probability distribution g which represents the probability that the sample is judged to be of each category;

If sample x is classified correctly, sample x need not synthesize a new sample;

If sample x is not classified correctly, then a new sample needs to be synthesized using x; the calculation formula of the synthesized sample number n is as follows:

Wherein the degree of difficulty in recognition of the sample is evaluated with 1-max (g);

Detecting an effective fragment s in a sample x by adopting a double-threshold endpoint detection algorithm [2], cutting the effective fragment s out of the sample x, randomly extracting n pieces of background noise from a background noise library Z, and then overlapping the effective fragment s on the background noise to form n new samples;

3.4.2: the operations of step 2.4.1 are performed for each sample in the dataset D _t;

3.4.3: adding all new samples synthesized in step 2.4.2 to dataset D _t, generating new dataset D _t+1;D_t+1 to be used for training of the next weak classifier;

Step 4: combining the weak classifiers into a strong classifier Q in a weighting way; the algorithm for the weighted combination is as follows: Wherein a _t is the weight of the t weak classifier calculated in step 3.3, and F _t is the trained t weak classifier.

The beneficial effects are that:

A plurality of weak classifiers for classifying underwater acoustic signals are trained by adopting an ensemble learning method, and are integrated in a weighting mode, so that a classifier with more powerful functions is obtained. Especially, the recognition errors of all weak classifiers are considered when the weights are designed (the weak classifiers with smaller errors occupy larger weights), so that the recognition capability of the integrated classifier is further improved.

In the training process of the weak classifier, a sample repetition synthesis method is adopted, so that the problem of insufficient data of the underwater sound samples is solved, and especially, different numbers of new samples can be iteratively synthesized according to the recognition difficulty degree of different samples in the sample synthesis process (namely, samples which are successfully recognized do not synthesize new samples any more, and the more difficult samples are recognized, the more new samples are synthesized), so that the pertinence and the effectiveness of sample synthesis are improved.

Drawings

Fig. 1 is a schematic flow diagram of a method for identifying underwater acoustic signals based on ensemble learning and sample synthesis.

Detailed Description

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

The underwater sound signal identification method based on integrated learning and sample synthesis is designed aiming at the problem of insufficient samples in underwater sound data, and achieves the purposes of reducing classifier errors, expanding data sample capacity and improving underwater sound signal identification efficiency and accuracy by means of integrated learning classification, sample repeated synthesis and the like.

The method mainly comprises the following steps:

Step 1: collecting underwater sound signal sample data and manufacturing an initial data set D; wherein the data set D contains C categories of underwater acoustic signals;

step 2: collecting underwater sound background sample data and manufacturing a background noise data set Z; wherein, the underwater sound background sample does not contain a target sound signal;

Step 3: iteratively training T weak classifiers; the method specifically comprises the following steps: initializing operation, namely constructing a data set D ₁ = D required by the 1 st weak classifier; training the 2 nd to the T th weak classifiers and corresponding training sets D ₂ to D _T in turn on the basis; wherein, the training process of the t weak classifier is as follows from 3.1 to 3.4:

3.1: constructing a random initialized neural network N _t, wherein the neural network adopts a convolutional neural network, and parameters of the neural network N _t are randomly set;

3.2: randomly sampling samples from the dataset D _t to train the neural network N _t;

3.2.1: randomly extracting a sample x from the data set D _t, and calculating a Mel spectrum m of the x;

3.2.2: inputting m to a neural network N _t, and adjusting the parameter value of N _t by using a back propagation algorithm;

3.2.3: repeating training based on the steps 2.2.1-2.2.2 to obtain a neural network Nt, namely a t weak classifier F _t;

3.3: performing one-time operation on all samples in the data set D _t by using a weak classifier F _t, calculating the total error e _t of F _t on the data set D _t, and calculating the total weight a _t of F _t on the final strong classifier by using e _t;

e _t is calculated as follows:

Wherein, P () represents a desire, I () is an indication function, x _i represents an I-th sample in the data set D _t, and y _i represents a class label corresponding to the I-th sample;

the calculation formula of a _t is as follows:

Step 3.4: adding a sample to be identified by a synthesis method to generate a new data set D _t+1; specifically:

3.4.1: for one sample x in the dataset Dt, carrying out reasoning and prediction operation on the sample x by using a classifier F _t, wherein the last layer of the neural network of F _t outputs a probability distribution g which represents the probability that the sample is judged to be of each category;

If sample x is classified correctly, sample x need not synthesize a new sample;

If sample x is not classified correctly, then a new sample needs to be synthesized using x; the calculation formula of the synthesized sample number n is as follows:

Wherein the degree of difficulty in recognition of the sample is evaluated with 1-max (g);

Detecting an effective fragment s in a sample x by adopting a double-threshold endpoint detection algorithm [2], cutting the effective fragment s out of the sample x, randomly extracting n pieces of background noise from a background noise library Z, and then overlapping the effective fragment s on the background noise to form n new samples;

3.4.2: the operations of step 2.4.1 are performed for each sample in the dataset D _t;

3.4.3: adding all new samples synthesized in step 2.4.2 to dataset D _t, generating new dataset D _t+1;D_t+1 to be used for training of the next weak classifier;

Step 4: combining the weak classifiers into a strong classifier Q in a weighting way; the algorithm for the weighted combination is as follows: Wherein a _t is the weight of the t weak classifier calculated in step 3.3, and F _t is the trained t weak classifier.

The following is a description of specific embodiments.

Firstly, a water sound classification data set is collected and constructed, and mainly comprises water sound data of six categories of ships, namely, a tanker, a tug, a passenger ship, a cargo ship, a yacht and a submarine. The first five ships collect 100 underwater sound samples respectively, the occurrence rate of the submarine is low, signals are difficult to capture, and therefore 20 underwater sound samples are collected.

Based on the principle and the steps, and the classification processing scheme design comparison experiment commonly used at present, the recognition conditions of the classifier under different conditions are compared.

The convolutional neural network suitable for classification adopts Resnet, 3 weak classifiers (i.e. t=3) are adopted in boosting ensemble learning, and the training round of each weak classifier is 100 (i.e. e=100).

The identification accuracy of six kinds of ships under different conditions is shown in table 1:

TABLE 1 identification accuracy of six ships under different conditions

Oil tanker

Tugboat

Passenger ship

Cargo ship

Yacht

Submarine craft

Single weak classifier

63％

68％

59％

66％

62％

35％

The 3 weak classifiers are integrated without adopting a data synthesis algorithm

81％

84％

78％

85％

84％

65％

3 Weak classifiers are integrated and a data synthesis algorithm is adopted

93％

95％

91％

96％

93％

90％

As can be seen from table 1:

(1) The identification accuracy of a single weak classifier on five common ships (oil tankers, tugboats, passenger ships, cargo ships and yachts) is about 60%, and the identification accuracy of a submarine is only 35% because of few samples.

(2) After the integrated learning is adopted, the recognition accuracy of all ships is obviously improved, the recognition accuracy of five common ships reaches about 80%, and the recognition accuracy of the submarine is improved to 65%.

(3) When the integrated learning is adopted and the data synthesis algorithm is adopted, the recognition accuracy of five common ships is improved to more than 90%, and the recognition accuracy of the submarine is also improved to 90%. The improvement of the accuracy of the underwater vehicle is most obvious mainly because the sample recognition difficulty of the underwater vehicle is the greatest, and the data synthesis algorithm synthesizes the most new samples aiming at the samples of the underwater vehicle, so that the recognition accuracy of the underwater vehicle is remarkably improved.

Finally, it should be noted that the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the scope of the present invention, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions can be made to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention.

Claims (1)

1. The underwater sound signal identification method based on integrated learning and sample synthesis is characterized by comprising the following steps of:

Step 1: collecting underwater sound signal sample data and manufacturing an initial data set D; wherein the data set D contains C categories of underwater acoustic signals;

step 2: collecting underwater sound background sample data and manufacturing a background noise data set Z; wherein, the underwater sound background sample does not contain a target sound signal;

Step 3: iteratively training T weak classifiers; the method specifically comprises the following steps: initializing operation, namely constructing a data set D ₁ = D required by the 1 st weak classifier; training the 2 nd to the T th weak classifiers in turn on the basis of the training, and constructing corresponding training sets D ₂ to D _T; wherein, the training process of the t weak classifier is as follows from 3.1 to 3.4:

3.1: constructing a random initialized neural network N _t, wherein the neural network adopts a convolutional neural network, and parameters of the neural network N _t are randomly set;

3.2: randomly sampling samples from the dataset D _t to train the neural network N _t;

3.2.1: randomly extracting a sample x from the data set D _t, and calculating a Mel spectrum m of the x;

3.2.2: inputting m to a neural network N _t, and adjusting the parameter value of N _t by using a back propagation algorithm;

3.2.3: repeating training based on the steps 3.2.1-3.2.2 to obtain a neural network N _t, namely a t weak classifier F _t;

3.3: performing one-time operation on all samples in the data set D _t by using a weak classifier F _t, calculating the total error e _t of F _t on the data set D _t, and calculating the total weight a _t of F _t on the final strong classifier by using e _t;

e _t is calculated as follows:

Wherein, P () represents a desire, I () is an indication function, x _i represents an I-th sample in the data set D _t, and y _i represents a class label corresponding to the I-th sample;

the calculation formula of a _t is as follows:

Step 3.4: adding a sample to be identified by a synthesis method to generate a new data set D _t+1; specifically:

3.4.1: for one sample x in the data set D _t, carrying out reasoning and prediction operation on the sample x by using a classifier F _t, wherein the last layer of the neural network of F _t outputs a probability distribution g which represents the probability that the sample is judged to be of each category;

If sample x is classified correctly, sample x need not synthesize a new sample;

If sample x is not classified correctly, then a new sample needs to be synthesized using x; the calculation formula of the synthesized sample number n is as follows:

Wherein the degree of difficulty in recognition of the sample is evaluated with 1-max (g);

Detecting an effective fragment s in a sample x by adopting a double-threshold endpoint detection algorithm [2], cutting the effective fragment s out of the sample x, randomly extracting n pieces of background noise from a background noise library Z, and then overlapping the effective fragment s on the background noise to form n new samples;

3.4.2: the operations of step 2.4.1 are performed for each sample in the dataset D _t;

3.4.3: adding all new samples synthesized in step 2.4.2 to dataset D _t, generating new dataset D _t+1;D_t+1 to be used for training of the next weak classifier;

Step 4: combining the weak classifiers into a strong classifier Q in a weighting way; the algorithm for the weighted combination is as follows: Wherein a _t is the weight of the t weak classifier calculated in step 3.3, and F _t is the trained t weak classifier.