EP3152534A1

EP3152534A1 - Method for classifying a water object, device, sonar, and water vehicle or stationary platform

Info

Publication number: EP3152534A1
Application number: EP15737967.8A
Authority: EP
Inventors: Ralf Michael MEYER; Sebastian KNABE
Original assignee: Atlas Elektronik GmbH
Current assignee: Atlas Elektronik GmbH
Priority date: 2014-06-05
Filing date: 2015-04-20
Publication date: 2017-04-12
Also published as: WO2015185032A1; DE102014107979A1

Abstract

The invention relates to a method for classifying a water object, in particular a ship or a submarine, wherein the method comprises the following steps: ascertaining an underwater sound signal by means of a hydrophone, wherein the underwater sound signal comprises a sound signal of the water object, determining a frequency spectrum of the determined underwater sound signal, determining a cepstrum of the frequency spectrum, determining a parameter set of the cepstrum, wherein the parameter set comprises direct parameters in particular, and determining a water object class on the basis of the parameter set.

Description

Method for classifying a water object, device, sonar and watercraft or stationary

platform

The invention relates to a method for classifying a water object, in particular a ship or a submarine, using detected underwater sound signals, as well as a device for carrying out the method and a sonar and a watercraft or stationary platform.

[02] Watercraft such as submarines and ships are sent by frequency lines

Underwater sound signals classified. The to

Classification features are partially extracted automatically, in addition to which an operator of the sonar gives specific characteristics recognized by him to a continuously evolving model and accordingly classifies. Thus, the classification is especially on a manually created rules.

In addition, underwater signals emitted by water objects are characterized on the basis of their spectral lines, for example, in a DEMON (DEMON = Detection of Envelope Modulation on Noise) or LOFAR (LOFAR = Low Frequency Analysis and Recording) method. By way of example, the patterns determined by means of this method can be used

Propeller speeds and the number of leaves are analyzed. An unerring classification in which The underwater sound signals emitting object can be reliably categorized, is not automated or only partially possible according to the prior art.

[04] The object of the invention is to improve the state of the art.

[05] The object is achieved by a method for classifying a water object, in particular a ship or a submarine, the method comprising the following steps:

Detecting an underwater sound signal by means of a hydrophone, wherein the underwater sound signal comprises a sound signal of the water object,

Determining a frequency spectrum of the detected underwater sound signal,

Determining a cepstrum from the frequency spectrum,

Determining a parameter rate of the cepstrum, the parameter set having, in particular, direct parameters, and

- Determining a water object class based on the parameter rate.

[06] Thus, a water object can be classified automatically based on the determined Parameterat zes. In addition, additional or additional information from an operator of the sonar is not needed, so that the This classification is not based on a manually created set of rules.

[07] The following terminology is explained:

[08] "Classification" means, in particular, classification in a specific (vehicle) category Since the "water objects" are, in particular, ships, submarines, manned or unmanned vessels or sound-emitting buoys, these can be described in certain categories are taken. For example, fish trawlers or small motorized vessels can form a first category, submarines (submarines) and AUV ^'s (autonomous underwater vehicle) or ROV ^'s (remotely operated vehicle). Another class, for example, form large container ships. The classes may be arbitrarily finely divided, the summary criteria being generally speed and / or size and / or applications.

[09] A "hydrophone" is in particular a sound-sensitive sensor that converts underwater sound signals into electrical signals, which may be part of a passive or active sonar.

[10] An "underwater sound signal" is generally a signal emitted or reflected by the water object, in the present case particularly good results being obtained with original underwater sound signals emitted by the object. [11] The signals recorded by the hydrophones are recorded especially in the time domain. These underwater sound signals present in the time domain can be converted, for example, by a transformation into a frequency domain, so that a "frequency spectrum" is present in this respect

Underwater sound signal forming frequencies shown.

[12] In particular, a "cepstrum" comprises a renewed transformation of the frequency spectrum

Logarithmizing or other mathematical

Been influenced influences. By forming the cepstrum, a "set of parameters" can be extracted that adequately describes the cepstrum, and parameters directly derived from the cepstrum are called direct parameters in the present case.

[13] On the basis of this determined parameter set, a water object can be assigned to a water object class. This can be done, for example, by determining the water object class based on the parameter set by means of a comparison, a correlation or other comparative functions.

[14] A "water object class" is in particular a category into which a water object is classified.

[15] In one embodiment, the

Underwater sound signal weighted. Thereby can advantageously higher-frequency components of

Underwater signal will be highlighted.

[16] "Weighting" is understood to mean, in particular, that gains and / or attenuations of the signal occur in a linear or non-linear manner., The weighting can be done both in the time domain and in the frequency domain.

[17] In order to improve the digital signal processing, a window function, in particular a Hamming window function, can be applied to the underwater sound signal and / or the weighted underwater sound signal. In particular, this window function can be used to determine the weighting with which samples of a signal are sampled within a section

(Window) in subsequent calculations. In particular, in a frequency analysis so-called leak effects can be avoided. In the present case, the window function can be applied both in the time domain and in the frequency domain.

[18] In a further embodiment, the determination of the frequency spectrum is effected by means of a fast Fourier transformation. Thus, an effective method for transforming into the frequency domain can be provided. In particular, such functions can be realized by means of computer or separately designed for building blocks, such as ^FPGAs (Field Programmable Gate Arrays).

[19] To obtain a manageable set of parameters, determining the cepstrum can be logarithmic Absolute value and / or a logarithm of several absolute values and / or a logarithm of all absolute values of the frequency spectrum.

[20] "Logarithm" can be applied to all bases such as base 2 or base 10.

[21] The "absolute values" are calculated in particular by amount formation from the complex frequency spectrum

In addition, the improvement of the parameter set is that when determining the cepstrum a discrete

Cosine transformation is performed.

[23] The modified discrete cosine transformation is also included. The discreet

Cosine transformation is one of the real, discrete, linear, orthogonal ones

Transformations that generally transform a discrete-time signal from the time domain or a spatial domain into a frequency domain, similar to the discrete Fourier transform.

[24] In order to increase the information content of the parameter set, parameters derived from the direct parameters of the parameter set can be formed so that the parameter set includes derived parameters (with). It should be noted at this point that, of course, only derived parameters can be used. [25] A "derived parameter" may be, for example, a delta value or a difference between two direct parameters or a direct parameter with a derived parameter, the direct parameters being, in particular, real valued numerical values.

[26] Since the significance of the individual parameters of the parameter can be different, the direct parameters and / or the derived parameters can be weighted. Such a weighting function may be, for example, a linear or a nonlinear function, which amplifies and / or attenuates the parameters.

In order to realize the determination of a water object class on the basis of the parameter object and the assignment of the water object to this water object class, the water object class can be determined by means of a comparison of the parameter data with a comparison parameter set, wherein in particular a correlation and / or a detection by means of a Gaussian mixing distribution takes place and / or the comparison takes place by means of a trained neural network.

In particular, already known water objects were assigned to an object class and the known underwater sound signals and the resulting "comparison parameter set ze" for the water object class were determined.

[29] In particular, all individual (comparison) parameter sets that apply to a water object class can be averaged, so that a comparison with these averaged Parameters takes place. If the determined parameter z is correlated with the averaged (comparison) parameter set of the object class, a probability is determined with which the water object of this object class is associated.

When using the Gaussian mixing distribution or the neural network, the corresponding computer or computer models by means of known

Water object data and thus the associated Parametersät ze trains. The more test data available, the better the water objects can be reliably assigned to the individual water objects in these implementations.

[31] A mixed distribution is understood to mean in particular a composite distribution. The principles are the same as in probability theory, so that after applying the Gaussian Mixture (GMM = Gaussian Mixture Model) the basic function is

fei Λ | 2 ^ΤΓ Σ ^ Ι

[32] where K is the number of mixed components, _{k is} the mean of a component E _k , the covariance matrix and c _{k is} the weight of the mixture of a component.

[33] In particular, the GMM is trained for each individual class. [34] The water object class that has reached the highest probability is assigned the corresponding water object.

[35] The neural networks present here are, in particular, artificial neural networks which are simulated, for example, by means of a computer. These neural networks are also trained on the individual water object classes, wherein a trained neural network, for example, in a hardware structure such as an FPGA can be transferred.

[36] In a further embodiment, the object is achieved by a device which is set up such that a previously described method can be carried out.

[37] Such a device may be, for example, a computer with corresponding input and output means. An FPGA is also included in this device.

[38] In another embodiment, the object is achieved by a sonar, in particular a passive sonar, which has a device described above.

Finally, the problem is solved by a watercraft or a stationary platform, in particular ship, buoy, submarine, AUV, ROV,

Underwater hull, Hovercraft, which or which has a previously described sonar or a device described above. [40] In the following, the invention will be explained in more detail with reference to an exemplary embodiment. It shows

Figure 1 is a schematic representation of a sent out for a water object

Underwater signal detected

Probability with a determined

Belonging to a (water object

) Class .

[41] A container ship is powered by propellers at sea. The propeller introduces an underwater sound signal into the seawater. This underwater sound signal is detected in the present case by a passive sonar and processed accordingly electronically.

[42] The processed time signal is weighted by the function S _n = Si - a · Si -1, where S _{n is} the weighted signal, Si is the signal at time step i, a is the weighting parameter, and Si -1 is the signal at time step i-1 is. The range of values of a is between 0 and 1, typically close to one.

[43] Subsequently, this weighted input signal S _{n is transferred} into two overlapping segments by means of a Hamming window. Thus, a reasonable time resolution is achieved for the downstream parameters.

[44] For each segment, the associated frequency spectrum is determined by means of a Fast Fourier Transform. [45] Subsequently, the absolute values of each individual frequency spectrum and the associated logarithms to the base 10 are formed.

[46] Subsequently, this logarithmic spectrum is adapted by means of bandpass filters, so that in particular a frequency axis adapted to human frequency perception is present. The advantage here is especially in a dimensional reduction.

[47] Subsequently, the filtered spectra are converted by means of a discrete cosine transformation, resulting in thirteen direct parameters.

[48] These thirteen direct parameters are amplified by a linear function, so that higher-order parameters experience greater amplification.

In addition, delta parameters are formed so that there are further (derived) parameters for the thirteen direct parameters through difference formation. In this case, both adjacent parameters and the first with the third, the second with the fourth, the third with the fifth, etc. derived parameters are formed.

[50] Four water object classes were formed in the present case. The first class includes submarines 105, the second class includes trawlers and small craft 109, the third class includes speedboats 113 and the fourth class includes container ships 117. [51] For ten different members of the respective class 105, 109, 113, 117, the known underwater sound signals were used and the associated direct parameters and derived parameters were formed. The arithmetic mean was then calculated from these parameters.

[52] The measured subsonic sound signal and the associated direct and derived parameters are correlated with the arithmetic-mean value parameter sets of the associated water object classes and the probability in percent is determined.

The water object classes for submarines 105, and the trawler class 109 and the speed boat class 113 give a lower probability (<30%) of belonging to this class again (107, 111, 115), with the probability 119 belonging to the

Water object class of container ships 117 is over 90% very high.

[54] This is shown graphically in FIG. 1, wherein the probability axis 103 indicates a value between 0 and 100% and four areas are indicated on the abscissa, which are each associated with one of the water object classes. The respectively associated probability 107 for the submarine class 105, the probability 111 for the trawlers 109 and the probability 115 for the speed boats 113 is so small that it can be ruled out in the present case that the water object which has transmitted the subsea sound signal belongs to these classes is associated. Due to the high probability 119, the ship mentioned above is classified as a container ship.

Reference sign list

101 bar chart

103 Correlation Coordinate in%

105 submarine

107 Correlation result to the submarine 105

109 trawlers

111 Correlation result to trawler 109

113 speedboat

115 Correlation result for Speedboat 113

117 container ship

119 Correlation result for the container ship 117

Claims

Claims:

A method for classifying a water object, in particular a ship or a submarine, the method comprising the following steps:

Determining a frequency spectrum of the detected underwater sound signal,

Determining a cepstrum from the frequency spectrum,

Determining a parameter rate of the cepstrum, wherein the parameter z has, in particular, direct parameters, and

- Determining a water object class based on the parameter rate.

2. The method according to claim 1, characterized in that the underwater sound signal is weighted.

3. The method according to any one of the preceding claims, characterized in that a window function, in particular a Hamming window function, is applied to the underwater sound signal or the weighted underwater sound signal.

4. The method according to any one of the preceding claims, characterized in that the determination of the frequency spectrum by means of a fast Fourier transform.

5. The method according to any one of the preceding claims, characterized in that determining the cepstrum is a logarithm of an absolute value and / or more Absolute values and / or all absolute values of the frequency spectrum includes.

6. The method according to any one of the preceding claims, characterized in that when determining the cepstrum a discrete cosine transformation is performed.

7. The method according to any one of the preceding claims, characterized in that from the direct parameters of the Parameterat zes derived parameters are formed, so that the parameter set comprises derived parameters.

8. The method according to any one of the preceding claims, characterized in that the direct parameters and / or the derived parameters are weighted.

9. Method according to one of the preceding claims, characterized in that the determination of the water object class by means of a comparison of the parameter set with a Vergleichsparametersat z takes place, in particular a correlation and / or detection by means of a Gaussian mixing distribution and / or

Comparison by means of a trained neural network takes place.

10. Device, which is arranged such that a method according to one of the preceding claims is feasible.

11. Sonar comprising a device according to claim 10.

12. Watercraft or stationary platform, in particular ship, submarine, buoy, which or which has a sonar according to claim 11 or an apparatus according to claim 10