CN104732970B - A kind of ship-radiated noise recognition methods based on comprehensive characteristics - Google Patents

Abstract

The present invention relates to a method for identifying ship radiation noise based on comprehensive features, comprising: constructing a training set of known ship radiation noise signals; extracting features of ship radiation noise from the constructed training set; wherein the features include Auditory features and statistical features, the auditory features include spectral flux, highest spectral peak and time center of gravity, the statistical features include power spectrum peak, average power, frequency of power spectrum peak, average frequency, frequency bandwidth, frequency kurtosis, Power standard deviation, power slope, power slope on frequency, power kurtosis, power kurtosis on frequency; use the characteristics of ship radiation noise as the feature training classifier for target recognition; read the signal; extract the features of the signal to be identified; input the extracted features to the classifier, so as to classify and identify the signal to be identified.

Description

A method for identifying ship radiation noise based on comprehensive features

technical field

The invention relates to the field of noise identification, in particular to a ship radiation noise identification method based on comprehensive features.

Background technique

The rapid development of underwater acoustic and electronic information technology makes the use of ship radiation noise for target classification and recognition an important research topic. It is an important part of the underwater information system. Personnel and the military sector are of great concern.

Ship radiated noise is very complex and is closely related to the marine environment and the ship's own motion state. They are constantly changing with different sea areas and different times. All of these have a great impact on the propagation of sound and the type of ship radiated noise, which brings great difficulties to the classification and identification of ship radiated noise. However, due to differences in internal structures such as hull structure, ship type, propeller size, number of blades, and power plant, different ships radiate different noises. Therefore, the type of ship can be identified by classifying the radiated noise of the ship.

The theories and methods of high-order statistics, wavelet transform, fractal geometry, artificial neural network, information fusion and data mining in modern signal processing technology have been widely used in ship radiation noise identification. However, due to the limitations of various methods, the accuracy of target recognition is affected, and the recognition results in the actual environment are not satisfactory. Human beings can achieve a very high level of recognition of ship radiation noise. However, the current description of human auditory characteristics is qualitative and empirical, and lacks visual description and quantitative analysis.

The human auditory system has very good natural scale and robustness for the decomposition and processing of sound signals, has good selectivity for sound source characteristics, and has good adaptability to environmental noise. Therefore, starting from the auditory characteristics of the human ear, we can study new feature extraction techniques suitable for underwater acoustic signals, find effective feature quantities in the subjective auditory quantities of the human ear, and achieve the purpose of improving the recognition rate of underwater targets. Quantitative or visual description of auditory features is of great significance for exploring the mechanism of human recognition of ship noise and underwater target recognition.

In the prior art, there is still a lack of methods for identifying ship radiated noise based on the auditory mechanism of the human ear.

Contents of the invention

The purpose of the present invention is to overcome the defect in the prior art that there is no method for identifying ship radiated noise based on the human auditory mechanism, thereby providing a method for identifying ship radiated noise based on comprehensive features.

In order to achieve the above object, the present invention provides a method for identifying ship radiation noise based on comprehensive features, including:

Step 1), constructing a training set of known ship radiation noise signals;

Step 2), in the training set constructed in step 1), extract the features of ship radiation noise; wherein, the features include auditory features and statistical features, and the auditory features include spectral flux, highest spectral peak and time center of gravity, The statistical features include power spectrum peak, average power, frequency of power spectrum peak, average frequency, frequency bandwidth, frequency kurtosis, standard deviation of power, power slope, slope of power on frequency, power kurtosis, kurtosis of power over frequency;

Step 3), using the characteristics of ship radiation noise obtained in step 2) as the feature training classifier for target recognition;

Step 4), read the signal to be identified;

Step 5), extract the features of the signal to be identified, the features include auditory features and statistical features, the auditory features include spectral flux, the highest spectral peak and time center of gravity, the statistical features include power spectral peak, average power, power Frequency of spectral peak, average frequency, frequency bandwidth, frequency kurtosis, standard deviation of power, power slope, slope of power over frequency, power kurtosis, kurtosis of power over frequency;

Step 6), input the features extracted in step 5) to the classifier trained in step 3), so as to classify and recognize the signal to be recognized.

In the above technical solution, in step 3), it also includes regularizing the values of the features and converting them to 0 before using the features of ship radiation noise as the features of target recognition to train the classifier A value between -1.

In the above technical solution, in the step 2) or step 4), extracting the spectral flux includes:

Step 2-1-1-1), first normalize the time domain waveform;

Step 2-1-1-2), determine the number of points in each frame of the sampling sequence, add a Hamming window to each frame sequence and perform a discrete FFT, and take the square of the modulus to obtain a discrete power spectrum;

Step 2-1-1-3), calculate the spectrum vectors of two adjacent frames according to the discrete power spectra of two adjacent frames, and then calculate the Pearson correlation coefficient between the spectrum vectors of two adjacent frames; the calculation formula is as follows:

Among them, X and Y are the _n -dimensional spectrum vectors of two adjacent frames, Xi and Yi are respectively the _i -th value in the two vectors, are the mean values of the two vectors;

Step 2-1-1-4), according to the Pearson correlation coefficient between two adjacent frames of spectrum vectors, calculate the spectral flux of the noise signal; the calculation formula is as follows:

Among them, M is the number of time frames that the noise signal is divided into; r _k , r _k-1 are the Pearson correlation coefficients between the spectrum vectors of two adjacent time frames.

In the above technical solution, in step 2) or step 4), extracting the highest spectral peak includes:

Step 2-1-2-1), first normalize the time domain waveform;

Step 2-1-2-2), add Hamming window to the time domain sequence and perform discrete FFT, take the square of the modulus to obtain the discrete power spectrum of the signal, and convert the power unit into decibels;

Step 2-1-2-3), finally find out the frequency value F _m corresponding to the maximum peak value of the noise signal power spectrum.

In the above technical solution, in the step 2) or step 4), extracting the time center of gravity includes:

Step 2-1-3-1), first normalize the time domain waveform;

Step 2-1-3-2), calculate the signal energy at each moment;

Step 2-1-3-3), calculate the time domain center of gravity of the noise signal, the calculation formula is as follows:

Among them, E(t) is the energy value corresponding to time t on the time domain diagram.

In the above technical solution, the classifier adopts SVM support vector machine.

The advantages of the present invention are:

The present invention comprehensively utilizes the subjective characteristics of simulating human hearing and the objective characteristics based on statistics, quantitatively expresses the subjective feelings of human beings about sound through three characteristics, and reflects the objective information of sound through statistical characteristics, comprehensively utilizes these two types of characteristics, and Optimizing the combination of features can effectively improve the recognition performance, making the recognition process closer to the human recognition process.

Description of drawings

Fig. 1 is a flow chart of the inventive method;

Figure 2(a) is a schematic diagram of identifying different types of ship radiated noise by using spectral flux;

Figure 2(b) is a schematic diagram of identifying different types of ship radiated noise by using time center of gravity;

Figure 3(a) shows the distribution of the eight types of ships under the st2-st5 two-dimensional map;

Figure 3(b) shows the distribution of eight types of ships under the st3-st5 two-dimensional map.

Detailed ways

The present invention will be further described now in conjunction with accompanying drawing.

The ship radiation noise identification method of the present invention is realized based on passive sonar. After the passive sonar passively receives radiation noise generated by underwater targets such as ships and signals emitted by underwater acoustic equipment, for these noises and signals, referring to Fig. 1, the method of the present invention adopts the following steps to realize the identification of ship radiation noise.

Step 1) Construct a training set of known ship radiation noise signals.

Firstly, preprocessing such as normalization is performed on the known types of ship radiation noise signals, and then the normalized known types of ship radiation noise are constructed as a training set.

Step 2), in the training set constructed in step 1), extract the characteristics of ship radiation noise.

In the present invention, the features of the ship radiation noise extracted include two categories, namely auditory features and statistical features. The following describes the extraction process of these two types of features.

Step 2-1), extracting auditory features.

Feature extraction is to extract some parameters that can characterize the physical characteristics of the target, so as to obtain the essential features that can best characterize the target features. The auditory features to be extracted in this step include three categories: spectral flux, highest spectral peak, and temporal center of gravity.

Step 2-1-1), extract spectral flux

The spectral flux of a sound signal describes the human ear's perception of sound over time, that is, the frequency transition characteristics on the time axis. It simulates the nonlinear discrimination characteristics of human hearing, reflects a large number of important characteristic information of sound signals, strongly affects subjective timbre, and is closer to human hearing experience. The Pearson correlation coefficient reflects the degree to which two variables are linearly related. According to the calculation method of spectral flux, the extraction of spectral flux features is carried out by frame, and the specific extraction steps are as follows:

Step 2-1-1-1), first preprocess the time-domain waveform (that is, the noise signal), that is, normalize the waveform;

Step 2-1-1-2), determine the number of points in each frame of the sampling sequence (for example, when the sampling rate is 8000Hz, 600 points per frame, 300 points overlapping), add a Hamming window to each frame sequence and do a 2048-point discrete FFT , taking the square of the modulus to obtain the discrete power spectrum;

Step 2-1-1-3), calculate the spectrum vectors of two adjacent frames according to the discrete power spectra of two adjacent frames, and then calculate the Pearson correlation coefficient between the spectrum vectors of two adjacent frames; the calculation formula is as follows:

Among them, X and Y are the _n -dimensional spectrum vectors of two adjacent frames, Xi and Yi are respectively the _i -th value in the two vectors, are the mean values of the two vectors, respectively.

Step 2-1-1-4), according to the Pearson correlation coefficient between two adjacent frames of spectrum vectors, calculate the spectral flux of the noise signal; the calculation formula is as follows:

Among them, M is the number of time frames that the noise signal is divided into; r _k , r _k-1 are the Pearson correlation coefficients between the spectrum vectors of two adjacent time frames.

Step 2-1-2), extract the highest spectral peak.

The signal power spectrum reflects the random distribution of signal energy, and the frequency value corresponding to the maximum peak value of the noise signal power spectrum represents the frequency value of the maximum signal energy.

The extraction steps of the highest spectral peak feature are as follows:

Step 2-1-2-1), first preprocess the time-domain waveform, that is, normalize the waveform;

Step 2-1-2-2), do 2048-point discrete FFT after adding Hamming window to the time domain sequence, take the square of the modulus to obtain the discrete power spectrum of the signal, and convert the unit of power into decibels;

Step 2-1-2-3), finally find out the frequency value F _m corresponding to the maximum peak value of the noise signal power spectrum.

Step 2-1-3), extract time center of gravity.

The time center of gravity of the noise signal is also the center of gravity of the time domain envelope, which reflects the time domain characteristics of the signal. The specific extraction steps are as follows:

Step 2-1-3-1), first preprocess the time-domain waveform, that is, normalize the waveform;

Step 2-1-3-2), calculate the signal energy at each moment;

Step 2-1-3-3), calculate the time domain center of gravity of the noise signal, the calculation formula is as follows:

Among them, E(t) is the energy value corresponding to time t on the time domain diagram.

Step 2-2), extract statistical features.

In order to facilitate the extraction of statistical features of ship radiation noise, the noise signal is first divided into T time frames, and the 2F-point Short Fourier Transform (STFT) of each frame is calculated. Thus, the signal will be represented by F spectral coefficients of T frames. In the formula below, p _t,f represents the power of the signal at frequency f at time t.

The statistical features to be extracted include:

The power spectrum peak, i.e. the maximum value of the total power over all timeframes:

st ₂ =M=max(p _t )

Average power, which is the average of the total power over all timeframes:

The frequency of the peak of the power spectrum, that is, the frequency corresponding to the peak of the power spectrum:

Average frequency, that is, the average frequency of the signal, where P represents the total power of the signal:

RMS bandwidth, that is, frequency bandwidth:

Frequency kurtosis, that is, the average frequency kurtosis:

The power SD is the standard deviation of the power, where F is the number of spectral coefficients and T is the number of time frames:

Power slope, that is, the slope of power:

Power slope, that is, the slope of power over frequency:

Power kurtosis, that is, the kurtosis of power:

Power kurtosis F, that is, the kurtosis of power at frequency:

The above 11 statistical features are the features that can best reflect the radiation noise of ships, so the following 11-dimensional statistical features are obtained from the above statistical features (st ₂ , st ₃ , st ₅ , st ₆ , st ₇ , st ₉ , st ₁₄ , st ₁₇ , st ₁₉ , st ₂₀ , st ₂₂ ).

Step 2-3), the three auditory features obtained from step 2-1) and the statistical features obtained from step 2-2) construct the following vectors used to represent the characteristics of ship radiated noise:

v={SF, F _m , TC, st ₂ , st ₃ , st ₅ , st ₆ , st ₇ , st ₉ , st ₁₄ , st ₁₇ , st ₁₉ , st ₂₀ , st ₂₂ }.

Step 3), the characteristics of ship radiation noise obtained in the previous step are input into the classifier as the characteristics of target recognition, so as to train the classifier.

As a preferred implementation, in order to prevent the contribution of other features from being overwhelmed by a certain feature value being too large, the feature values are regularized before the features of ship radiation noise are input into the classifier, and they are converted into values between 0 and 1. value.

The classifier in this step can use SVM support vector machine. SVM support vector machine is developed from the optimal classification hyperplane in the case of linear separability. Its mechanism and processing process are equivalent to transforming the original input space into a new one. The feature space of , and solve the optimal linear classification plane in the new space.

Wherein, the main form of kernel function has following four kinds: linear kernel function, polynomial kernel function, radial kernel function and Sigmoid kernel function, what this application adopts is the polynomial kernel function of 6th order, in other embodiments, also can adopt Other types of kernel functions.

The classification function formed by SVM training data has the following properties: SVM is a linear combination of a set of nonlinear functions with support vectors as parameters, so the expression of the classification function is only related to the number of support vectors, but independent of the dimension of the space. SVMs are especially effective when dealing with classification in high-dimensional input spaces.

Step 4), read the signal to be identified;

Step 5), extracting the features of the signal to be identified; the features include the aforementioned 3 auditory features and 11 statistical features.

Step 6), input the features extracted in step 5) to the trained classifier, so as to classify and recognize the signal to be recognized.

Prove the effect of the method of the present invention by experiment below. In an experiment, the radiation noise of eight types of ships was detected, as shown in Figure 2, where 'asterisk' represents active sonar signals; 'square' represents 10,000-ton ships, and 'circle' represents cargo ships; The 'plus sign' represents another large ship; the 'diamond' represents a small fishing vessel; the 'right triangle' represents a medium-sized surface vessel; the 'down triangle' represents ocean noise; and the 'pentagram' represents a small craft. It can be seen from the figure that both the spectral flux (Fig. 2(a)) and the time center of gravity (Fig. 2(b)) in the acoustic features can be used to distinguish different types of ship radiation noise. The discrimination of different types of ship radiated noise is better than the time center of gravity.

Figure 3(a) shows the distribution of eight types of ships under the st2-st5 two-dimensional map, and Figure 3(b) shows the distribution of eight types of ships under the st3-st5 two-dimensional map. These two figures show that the separability of these statistical features is good.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than limit them. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent replacements to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all of them should be included in the scope of the present invention. within the scope of the claims.

Claims (5)

1. A ship radiated noise identification method based on comprehensive features, comprising: Step 1), constructing a training set of known ship radiation noise signals; Step 2), in the training set constructed in step 1), extract the features of ship radiation noise; wherein, the features include auditory features and statistical features, and the auditory features include spectral flux, the highest spectral peak and time center of gravity, The statistical features include power spectrum peak, average power, frequency of power spectrum peak, average frequency, frequency bandwidth, frequency kurtosis, standard deviation of power, power slope, slope of power on frequency, power kurtosis, kurtosis of power over frequency; Step 3), utilizing the feature of the ship radiation noise obtained in step 2) as a feature training classifier for target recognition; Step 4), read the signal to be identified; Step 5), extracting the features of the signal to be identified, the features include auditory features and statistical features, the auditory features include spectral flux, the highest spectral peak and time center of gravity, the statistical features include power spectrum peak, average power, power Frequency of spectral peak, average frequency, frequency bandwidth, frequency kurtosis, standard deviation of power, power slope, slope of power over frequency, power kurtosis, kurtosis of power over frequency; Step 6), inputting the feature extracted in step 5) to the classifier trained in step 3), so as to classify and identify the signal to be identified; In the step 2) or step 4), extracting the spectral flux includes: Step 2-1-1-1), first normalize the time domain waveform; Step 2-1-1-2), determine the points of each frame sampling sequence, do discrete FFT after adding a Hamming window to each frame sequence, and obtain the discrete power spectrum by taking the square of the modulus; Step 2-1-1-3), calculate the spectrum vectors of two adjacent frames according to the discrete power spectra of two adjacent frames, and then calculate the Pearson correlation coefficient between the spectrum vectors of two adjacent frames; its calculation formula is as follows: Among them, X and Y are the _n -dimensional spectrum vectors of two adjacent frames, Xi and Yi are respectively the _i -th value in the two vectors, are the mean values of the two vectors; Step 2-1-1-4), calculate the spectral flux of the noise signal according to the Pearson correlation coefficient between each adjacent two frame spectrum vectors; its calculation formula is as follows: Among them, M is the number of time frames that the noise signal is divided into; r _k , r _k-1 are the Pearson correlation coefficients between the spectrum vectors of two adjacent time frames. 2. the ship radiation noise recognition method based on comprehensive feature according to claim 1, is characterized in that, in described step 3), also comprises in the feature training classification with the feature of ship radiation noise as target recognition Before the filter, it also includes regularizing the values of the features and converting them into values between 0 and 1. 3. according to claim 1 or 2 described ship radiation noise identification methods based on integrated features, it is characterized in that, in described step 2) or step 5), extracting the highest spectral peak comprises: Step 2-1-2-1), first normalize the time domain waveform; Step 2-1-2-2), do discrete FFT after adding Hamming window to the time domain sequence, take the square of the modulus to obtain the discrete power spectrum of the signal, and convert the unit of power into decibels; Step 2-1-2-3), finally find out the frequency value F _m corresponding to the maximum peak value of the noise signal power spectrum. 4. according to claim 1 and 2 described methods for identifying ship radiation noise based on comprehensive features, it is characterized in that, in described step 2) or step 5), extracting the time center of gravity comprises: Step 2-1-3-1), first normalize the time domain waveform; Step 2-1-3-2), obtain the signal energy at each moment; Step 2-1-3-3), calculate the time-domain center of gravity of the noise signal, and its calculation formula is as follows: Among them, E(t) is the energy value corresponding to time t on the time domain diagram. 5. The ship radiation noise recognition method based on comprehensive features according to claim 1 or 2, wherein the classifier adopts SVM support vector machine.