Sound signal feature optimization method for analysis and recognition of underwater small target object

Abstract

The invention belongs to the technical field of underwater small target object analysis and recognition methods, and particularly relates to a sound signal feature optimization method for underwater small target object analysis and recognition. Comprising the following steps: 1, acquiring and processing original data; 2, eliminating interference line spectrums, and intercepting and splicing by adopting a time window to obtain a time recombination sound signal; using range transformation to optimize signal amplitude characteristics; processing the sound signal by using an FIR digital high-pass filter; obtaining an optimized signal by using a homomorphic filter; 3, extracting structural features, energy features, frequency features and sound intensity features of the optimized digital signals; step 4, sound signal feature fusion: carrying out underwater target sound signal feature fusion processing based on a canonical correlation analysis method; the invention provides a method for extracting effective sound signal data of an underwater target under the background of optimizing a complex marine environment, enhancing signal effectiveness during long-distance detection, and constructing effective sound signal features capable of being used for classification and recognition of the underwater target at the same time.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of underwater small target object analysis and recognition methods, and particularly relates to a sound signal feature optimization method for underwater small target object analysis and recognition. BACKGROUND

[0002] Underwater target sound signal analysis is an important means for marine resource development and utilization and underwater target detection analysis. Under the influence of underwater complex acoustic signals and underwater acoustic environment, underwater small target objects emit underwater acoustic signals which are easily disturbed and affected by a large amount of underwater noise. The sound signal of a long-distance underwater target attenuates rapidly and is weak and difficult to identify, and is easily submerged by environmental noise. It is difficult to obtain effective sound information data by using a traditional underwater target signal detection method. At the same time, the underwater target is affected by ocean currents and changes in its own motion state, and the difference in hydrological environment under the marine environment causes great uncertainty in the superposition of sound signals and environmental factors. The traditional sound information sorting scheme is difficult to effectively analyze and identify, and it is difficult to effectively obtain effective sound signal features that can be used for underwater target analysis and identification. SUMMARY

[0003] The purpose of the present application is to provide a method for optimizing the extraction of effective sound signal data of underwater targets under complex marine environment background, enhancing the effectiveness of signals during long-distance detection, and constructing effective sound signal features that can be used for underwater target classification and identification.

[0004] To achieve the above purpose, the technical scheme is as follows.

[0005] A sound signal feature optimization method for underwater small target object analysis and recognition, characterized by steps (1) to (4).

[0006] Step 1, sound signal detection and collection; specifically:

[0007] Based on the optical fiber hydrophone array, a water acoustic signal collection system is established. For the collected analog sound signals, an analog amplification filter is used for pre-processing, and an AD converter is used to convert the collected analog sound signals into digital signals for digital filtering processing.

[0008] Step 2, water acoustic signal pre-processing, specifically including:

[0009] (2.1) Eliminate interference line spectrum in the power spectrum of the water acoustic signal

[0010] (2.2) For the sound signal acquired based on the collection time , in order to avoid the interference of the motion change characteristics of the target sound signal caused by the change of the target motion, a time window is used to process the sound signal based on the time sequence The time-reconstructed audio signal is obtained by truncating and splicing the data. ;

[0011] (2.3) Due to the varying distances between the fiber optic hydrophone array and the small underwater target, as well as the influence of the target's motion, time-reconstructed acoustic signals may be generated. The amplitude is abnormal. In order to control the abnormal amplitude, the range transform is used to perform a linear transformation on the reconstructed acoustic signal to optimize the signal amplitude characteristics.

[0012] (2.4) During the underwater transmission of acoustic signals, the attenuation of the high-frequency characteristics of the acoustic signal will far exceed the attenuation rate of the low-frequency characteristics due to the influence of the acoustic transmission characteristics. In order to balance the frequency characteristic distribution of the acoustic signal and improve the effectiveness of the acoustic signal, an Nth-order FIR digital high-pass filter is used to process the acoustic signal.

[0013] (2.5) For underwater acoustic signal data of small target objects, underwater environmental noise can be suppressed by optimizing the signal-to-noise ratio; for underwater acoustic signals of small targets, they can be represented as the actual underwater acoustic signals of the small target objects. and ambient noise signals The convolution function is represented Logarithmic representation of convolution function based on Fast Fourier Transform Further analysis of the logarithmic expression signal Perform a Fourier transform to convert it to the frequency domain to obtain its spectral signal. Using homomorphic filters to convert the spectral signal in the frequency domain Multiplying the two results in the spectrum after homomorphic filtering. ; the filtered spectrum Transforming back to the spatial domain yields the filtered logarithmic expression. For the filtered logarithmic expression signal Perform exponential operations to obtain the optimized signal. ;in Indicates Fourier transform, Indicates inverse transformation;

[0014] Step 3: Construction of acoustic signal features, specifically including:

[0015] (3.1) Extracting the structural features of the optimized digital signal:

[0016] The structural features refer to the multiple time-domain characteristic attributes of digital signals, which can be expressed as multidimensional feature vectors. ,in The optimized digital signal represents the first... A time domain characteristic attribute corresponds to an attribute vector, the time domain characteristic attribute refers to a plurality of characteristic attributes related to the wavelength, amplitude of the digital signal, and the corresponding attribute vector includes but is not limited to one or more of the following attribute vectors:

[0017] A, wavelength time domain vector ;

[0018] B, zero-crossing wave distribution vector ;

[0019] C, main wavelength vector ;

[0020] D, wavelength distribution vector ;

[0021] E, wavelength difference distribution vector ;

[0022] F, amplitude distribution vector ;

[0023] Joint attribute vector, time domain characteristic attribute feature vector ;

[0024] Wherein, is the average period corresponding to the wavelength, represents the detection period; represents the number of zero-crossing waves of the wavelength in the interval , represents the first wavelength interval; represents the number of adjacent zero-crossing wave wavelength differences in the interval , represents the first wavelength difference interval; represents the number of wave peaks with interpeak amplitude in the amplitude interval , represents the first interpeak amplitude interval, which is obtained by equally dividing the maximum interpeak amplitude of the optimized digital signal;

[0025] (3.2) Extract the energy characteristics of the optimized digital signal:

[0026] By wavelet packet decomposition of the optimized digital signal, the wavelet packet coefficients of each layer from low frequency to high frequency are extracted , represents the wavelet coefficient of the jth layer and the kth node, and the energy characteristics of each frequency band are obtained by reconstructing the digital signal according to the wavelet packet coefficient , and further obtaining the subband energy coefficient ; the energy characteristic attribute vector is obtained ;

[0027] (3.3) Extract the frequency feature of the optimized digital signal:

[0028] The obtained digital signal is converted by a Mel filter bank to convert the frequency domain of the digital signal to the Mel scale frequency domain; the frequency domain signal after the Mel filter bank is subjected to logarithmic operation to enhance the frequency domain feature intensity of the speech signal; the logarithmic spectrum obtained in the previous step is subjected to discrete cosine transform (DCT) to convert it to a cepstrum space; the cepstrum coefficients are normalized to enhance the stability of the signal; the cepstrum coefficients of adjacent frames are spliced to form the final frequency feature vector;

[0029] (3.4) Extract the intensity feature of the optimized digital signal:

[0030] The optimized digital signal is obtained, and the frequency domain signal is obtained by discrete Fourier transform, and the square sum of the frequency spectrum imaginary part and the real part is obtained to obtain the power spectrum of the signal , wherein represents the real part extraction operation, represents the imaginary part extraction operation;

[0031] Based on the critical band analysis method, the frequency axis of the power spectrum is mapped to the critical frequency to obtain the critical spectrum , and the frequency axis of the critical spectrum can be represented as ;

[0032] The center frequency of each frequency band of the critical spectrum is extracted, the high-end frequency , the low-end frequency and the frequency band weighting coefficient of each frequency band are calculated, and the power spectrum is multiplied by the frequency band weighting coefficient to obtain the critical bandwidth spectrum ; wherein:

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] wherein, is an N-point frequency;

[0038] The signal is reconstructed using an analog isometric curve to obtain a reconstructed signal The reconstructed signal is subjected to loudness amplitude compression to obtain a sound intensity characteristic attribute vector

[0039]

[0040] Step four, acoustic signal feature fusion, based on canonical correlation analysis method for underwater target acoustic signal feature fusion processing; Specifically includes:

[0041] (4.1) Based on the attribute vector determined in step three, the corresponding feature vector space is established, and for the underwater target signal , assuming its sample space is ,

[0042] (4.2) For the underwater target signal, the feature vectors and in any two different feature vector spaces and are solved respectively, and the canonical correlation feature between the two feature vectors is determined based on the canonical correlation analysis algorithm , wherein represents the vector space group;

[0043] (4.3) The space correlation feature vector is constructed, ; then the combined feature composed of the foregoing attribute vectors can be represented as .

[0044] Further improvement or preferred implementation steps of the foregoing acoustic signal feature optimization method for underwater small target object analysis and recognition, the step (2.2) specifically includes: for the acoustic signal , assuming that the time window function adopted is , then the sufficient time recombination acoustic signal can be represented as: ; wherein represents the conjugate function, refers to the angular frequency of the acoustic signal , is the imaginary part symbol.

[0045] Further improvement or preferred implementation steps of the foregoing acoustic signal feature optimization method for underwater small target object analysis and recognition, in the step (2.3), for the recombination acoustic signal , the maximum amplitude and the minimum amplitude amplitude are obtained; for the recombination acoustic signal​​​​ Observed amplitude The adjusted amplitude can be expressed as .

[0046] A further improvement or preferred implementation of the aforementioned acoustic signal feature optimization method for underwater small target analysis and identification, wherein the FIR digital high-pass filter output in step (2.4) can be expressed as:

[0047]

[0048] in This indicates the signal output. Indicates the filter order. Indicates the delay period. This indicates the transfer function corresponding to the filter. This represents the filter coefficients.

[0049] Further improvements or optimizations to the aforementioned acoustic signal feature optimization method for underwater small target analysis and identification can be achieved by fusing multi-sensor data using the following algorithm, depending on actual needs:

[0050] A. Data fusion method based on weighted average

[0051] This method involves weighted averaging of data from different types of sensors to obtain a more accurate and reliable fusion result. It is simple, intuitive, and easy to implement. The fused data processed using the weighted averaging method can reduce the errors and uncertainties that may exist in data from a single sensor, improving the stability and robustness of the entire system. This can be expressed as follows: ;

[0052] in It refers to the fused signal. This refers to the number of signals. It refers to the k-th signal. This refers to the weight of the k-th signal.

[0053] B. Data fusion method based on Kalman filtering

[0054] Specifically, the system state is estimated iteratively through two steps: prediction and update. In the prediction step, the dynamic model of the system is used to predict the state at the next moment. In the update step, the observed data is used to correct the predicted value. Its advantages are that it can handle noisy and uncertain data, reduce the impact of noise by fusing data from different sensors, improve the accuracy and stability of the system, and has a small memory footprint and fast computing speed. It is suitable for systems with high real-time requirements and is often used for the fusion of low-level real-time dynamic multi-sensor data.

[0055] C. Bayesian estimation based data fusion method;

[0056] Specifically: the uncertainty of the observation data and the uncertainty of the prior probability are combined together to obtain a more accurate state estimation. Prior probability distribution of the system needs to be given as accurately as possible before use. It is a fusion algorithm based on probability statistics, which uses prior probability and new observation data to update posterior probability.

[0057] In single sensor detection, under certain assumptions, the conditional probability function of the single sensor decision result is the likelihood function p(z|Hi), and the concept of the fusion decision result output of multiple sensors is the likelihood function, i.e. p(u|Hi), where u=(u1, u2,..., uN). In engineering applications, the consequences of various errors are not equally serious, i.e. the losses or costs caused by different types of errors are different. In order to reflect these differences, different costs should be specified for each type of error probability, i.e. the so-called cost function to reflect the difference in loss. In the Bayesian estimation based data fusion process, the Bayesian fusion detection criterion allocates a corresponding cost value to each decision result, and the average total cost is obtained based on the assumption probability. The detection strategy is to minimize the average total cost. The specific steps include:

[0058] Step 1: initialize the conditions, determine the observation distribution of each sensor and assume the Bayesian decision threshold of each sensor:

[0059]

[0060]

[0061] From Calculate ;

[0062] Step 2: set the loop variable and the termination control variable , prepare for the loop; estimate the Bayesian fusion detection criterion:

[0063]

[0064] Calculate the corresponding Bayesian fusion risk

[0065]

[0066] Step 3: calculate the new decision threshold based on the Bayesian decision threshold of each sensor calculated in the last step:

[0067]

[0068]

[0069]

[0070] According to the new threshold of each sensor, the corresponding sensor is calculated ,

[0071] The Bayesian decision threshold of each sensor is calculated based on: ;

[0072] The Bayesian fusion detection criterion is estimated:

[0073]

[0074] Step 4: Calculate the Bayesian risk RB(K+1) of this iteration k+1:

[0075]

[0076] Step 5: After each sensor makes a decision, the decision result is sent to the fusion center to realize data fusion;

[0077] D. Fuzzy logic reasoning method refers to using a real number between 0 and 1 to represent the degree of truth or membership. In the process of multi-sensor fusion, fuzzy logic is used to handle uncertainty, and these uncertain factors are included in the reasoning process. Through systematic methods, the uncertainty in the fusion process is modeled, and consistent reasoning is performed based on fuzzy logic, so as to obtain more accurate and reliable fusion results.

[0078] E. Artificial neural network method refers to using deep learning models (such as CNN, LSTM) to learn the nonlinear mapping relationship of multi-sensor data. Through continuous training of sample data, high-efficiency logical reasoning ability is gradually formed. By utilizing its advantages in signal processing and automatic reasoning function, precise fusion of multi-sensor data is realized. Neural network algorithm has excellent fault tolerance, self-adaptability, self-learning ability and self-organization ability, and can simulate extremely complex nonlinear mapping relationship. In a multi-sensor system, the information provided by each sensor has certain uncertainty, so the fusion of these uncertain information is essentially an uncertainty reasoning.

[0079] F. Extended Kalman filter fusion method performs Taylor expansion linearization on nonlinear systems, and then applies Kalman filter framework. Extended Kalman Filter (EKF) solves the problem of nonlinearity through local linearization. The prediction equation and observation equation of nonlinearity are differentiated to linearize in the form of tangent replacement.

[0080] The prediction model and the measurement model of the extended Kalman are nonlinear, in order to simplify the calculation, the motion equation and the observation equation are linearized through first-order Taylor decomposition, the posterior probability density is described in the form of Gauss, and when the variance is calculated, the state transition matrix and the observation matrix are the Jacobian matrix of the state information. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 is a flowchart of the acoustic signal feature optimization method for underwater small target object analysis and recognition;

[0082] Figure 2 is a schematic diagram of a significant interference line spectrum that may occur in the power spectrum;

[0083] Figure 3 is a schematic diagram of an optical fiber hydrophone acquisition system;

[0084] Figure 4 is a schematic diagram of an optical fiber hydrophone array establishing an underwater acoustic signal acquisition system; DETAILED DESCRIPTION

[0085] The application will be described in detail below in combination with specific embodiments.

[0086] The present application mainly realizes the analysis and recognition of long-distance underwater small target objects based on underwater acoustic signal analysis. Affected by complex underwater acoustic signals and underwater acoustic environments, the underwater acoustic signals emitted by underwater small target objects have problems such as poor recognizability, low signal strength, and strong interference noise. It is difficult to obtain effective analysis and recognition results by using traditional underwater acoustic target signal detection methods. To solve this problem, the present application proposes an underwater small target analysis and recognition method for enhancing and stripping the recognition of small target acoustic signals through multi-dimensional feature fusion after optimizing the underwater acoustic signals.

[0087] To achieve the above purpose, as shown in Figure 1 The acoustic signal feature optimization method for underwater small target object analysis and recognition of the present application mainly includes the following steps

[0088] Step one, raw data acquisition and processing;

[0089] Based on the optical fiber hydrophone array establishing an underwater acoustic signal acquisition system (as shown in Figure 3 , Figure 4 For the collected analog acoustic signals, an analog amplification filter is used for pre-processing the signals, and an AD converter is used to convert the collected analog acoustic signals into digital signals for digital filtering processing.

[0090] Step two, processing and analysis of the original digital signal

[0091] (2.1) During the acquisition of underwater acoustic signals, the presence of abnormal signals from ships, equipment, and the marine environment can lead to abnormal signals in the underwater acoustic target signal. These abnormal signals will produce significant interference lines in the power spectrum of the underwater acoustic signal, such as... Figure 2 As shown, the underwater acoustic signal corresponding to the interference line spectrum (such as...) Figure 2 The three most prominent extreme points in the data are eliminated.

[0092] (2.2) For data collection time-based data collection Acquired sound signals To avoid interference from motion-related changes in the acoustic signal caused by the target's motion, a time window is used to process the time-series-based acoustic signal. The time-reconstructed audio signal is obtained by truncating and splicing the data. ;

[0093] Specifically: for sound signals Assuming it uses a time window function as follows: Sufficient time is given to reconstruct the sound signal. It can be represented as: ;in Denotes the conjugate function. refers to sound signals angular frequency, It is the symbol for the imaginary part;

[0094] (2.3) Due to the varying distances between the fiber optic hydrophone array and the small underwater target, as well as the influence of the target's motion, time-reconstructed acoustic signals may be generated. The amplitude is abnormal. In order to control the abnormal amplitude, the range transform is used to perform a linear transformation on the reconstructed acoustic signal to optimize the signal amplitude characteristics.

[0095] Specifically: for reconstructed acoustic signals Obtain its maximum amplitude. and minimum amplitude Amplitude; for reconstructed acoustic signals Observed amplitude The adjusted amplitude can be expressed as ;

[0096] (2.4) During the underwater transmission of acoustic signals, the attenuation of the high-frequency characteristics of the acoustic signal will far exceed the attenuation rate of the low-frequency characteristics due to the influence of the acoustic transmission characteristics. In order to balance the frequency characteristic distribution of the acoustic signal and improve the effectiveness of the acoustic signal, an Nth-order FIR digital high-pass filter is used to process the acoustic signal.

[0097] The output of the FIR digital high-pass filter can be expressed as:

[0098]

[0099] wherein represents a signal output, represents a filter order, represents a delay period, represents a filter corresponding transfer function, represents a filter coefficient;

[0100] (2.5) For small target object underwater acoustic signal data, the underwater environmental noise can be suppressed by signal-to-noise ratio optimization; for underwater small target underwater acoustic signal, it is expressed as small target object underwater acoustic actual signal and environmental noise signal convolution function , the logarithmic expression of the convolution function is obtained based on fast Fourier transform ; further Fourier transform is performed on the logarithmic expression signal , and the frequency spectrum signal is obtained by converting to the frequency domain; the homomorphic filter is used to multiply the frequency spectrum signal in the frequency domain to obtain the frequency spectrum after homomorphic filtering ; the filtered frequency spectrum is converted back to the spatial domain to obtain the filtered logarithmic expression , the exponential operation is performed on the filtered logarithmic expression signal to obtain the optimized signal ; wherein represents Fourier transform, represents inverse transform;

[0101] Step three, acoustic signal feature construction

[0102] (3.1) Extract the structural features of the optimized digital signal:

[0103] The structural features refer to a plurality of digital signal time domain feature attributes, which can be expressed as a multi-dimensional feature vector , wherein represents the time domain feature attribute corresponding attribute vector of the optimized digital signal, the time domain feature attribute refers to a plurality of feature attributes related to the wavelength and amplitude of the digital signal, and the corresponding attribute vector includes but is not limited to one or more of the following attribute vectors:

[0104] A, wavelength time domain vector ;

[0105] B, zero-crossing wave distribution vector ;

[0106] C, main wavelength vector ;

[0107] D, wavelength distribution vector ;

[0108] E, wavelength difference distribution vector ;

[0109] F, amplitude distribution vector ;

[0110] Joint attribute vector, time domain feature attribute feature vector ;

[0111] wherein, refers to the average period corresponding to the wavelength, represents the detection period; represents the number of zero-crossing waves whose wavelength is in the interval , represents the jth wavelength interval; represents the number of adjacent zero-crossing wave wavelength differences whose wavelength difference is in the interval , represents the jth wavelength difference interval; represents the number of peak-to-peak amplitudes whose peak-to-peak amplitude is in the amplitude interval , represents the jth peak-to-peak amplitude interval, which is obtained by equally dividing the maximum peak-to-peak amplitude of the optimized digital signal; The time domain feature attribute feature vector based on the multi-dimensional structure feature vector can fully express the basic feature attributes of the underwater target acoustic signal waveform, and is convenient for further analysis and processing.

[0112] (3.2) Extract the energy feature of the optimized digital signal:

[0113] Through wavelet packet decomposition of the optimized digital signal, the wavelet packet coefficients of each layer from low frequency to high frequency are extracted ,

[0114] represents the wavelet coefficient of the jth layer and the kth node, and the energy feature of each frequency band is obtained by reconstructing the digital signal according to the wavelet packet coefficient, and further obtaining the subband energy coefficient ; the energy feature attribute vector is obtained;

[0115] The wavelet packet energy decomposition method can visualize the recognizable ability of the underwater target acoustic signal in the unstable state under the influence of movement and variable environment.

[0116] (3.3) Extract the frequency feature of the optimized digital signal: ​​​

[0117] The obtained digital signal is converted by a Mel filter bank to convert the frequency domain of the digital signal to a Mel scale frequency domain; the frequency domain signal after the Mel filter bank is subjected to logarithmic operation to enhance the frequency domain feature intensity of the speech signal; the logarithmic spectrum obtained in the previous step is subjected to discrete cosine transform (DCT) to convert it to a cepstrum space; the cepstrum coefficients are subjected to normalization processing to enhance the stability of the signal; the cepstrum coefficients of adjacent frames are spliced to form a final frequency feature vector;

[0118] The sound signal attribute features obtained based on the Mel filter can better preserve the basic attribute features of the sound signal, make it have more intuitive recognizable ability, and further optimize the recognition ability of the underwater target sound signal in a complex environment.

[0119] (3.4) Extract the sound intensity features of the optimized digital signal:

[0120] Obtain the optimized digital signal, obtain the frequency domain signal through discrete Fourier transform, obtain the square sum of the frequency spectrum imaginary part and real part, and obtain the power spectrum of the signal , wherein represents a real part extraction operation, represents an imaginary part extraction operation;

[0121] Based on the critical band analysis method, the frequency axis of the power spectrum is mapped to the critical frequency to obtain the critical spectrum , and the frequency axis of the critical spectrum can be represented as ;

[0122] Extract the center frequency of each frequency band of the critical spectrum, calculate the high-end frequency , the low-end frequency and the frequency band weighting coefficient of each frequency band, multiply the power spectrum and the frequency band weighting coefficient to obtain the critical bandwidth spectrum ; wherein:

[0123] ;

[0124] ;

[0125] ;

[0126] ;

[0127] wherein, is an N-point frequency;

[0128] Reconstructing the signal using the analog isometric curve to obtain a reconstructed signal , compressing the loudness amplitude of the reconstructed signal to obtain an intensity feature attribute vector ;

[0129] Wherein ; ;

[0130] Based on the anti-noise characteristics of the sound intensity signal, it is integrated into the sound signal feature vector, which can make the noise signal further visible, facilitate further identification of the noise signal, and improve the identifiability of the target sound signal.

[0131] Step four, feature fusion, underwater target sound signal feature fusion processing based on canonical correlation analysis method;

[0132] Based on the attribute vectors determined in step three, the corresponding feature vector space is established, and for the underwater target signal , assuming its sample space is , ;

[0133] For the underwater target signal, the feature vectors and in any two different feature vector spaces and are solved respectively, and the canonical correlation features between the two feature vectors are determined based on the canonical correlation analysis algorithm , wherein represents a vector space group;

[0134] Construct a space correlation feature vector , ; then the combined feature composed of the foregoing attribute vectors can be represented as .

[0135] Further, the multi-sensor data fusion processing can be performed according to the actual requirements as follows:

[0136] A, data fusion method based on weighted average

[0137] That is, the data of different types of sensors are processed by weighted average to obtain a more accurate and reliable fusion result. This method is simple and intuitive, easy to implement, and the fusion data processed by the weighted average method can reduce the error and uncertainty that may exist in single sensor data, improve the stability and robustness of the entire system, and can be expressed as ;

[0138] Wherein is the fusion signal, is referred to as the number of signals, is referred to as the kth signal, is referred to as the weight of the kth signal

[0139] B, Kalman filter-based data fusion method

[0140] Specifically, the system state is estimated iteratively through two steps of prediction and update. In the prediction step, the state at the next time is predicted using the dynamic model of the system, and in the update step, the predicted value is corrected using the observation data. The advantage is that it can handle data with noise and uncertainty, reduce noise influence by fusing different sensor data, improve system accuracy and stability, and has small memory occupation and fast operation speed, which is suitable for systems with high real-time requirements and is commonly used for low-level real-time dynamic multi-sensor data fusion.

[0141] C, Bayesian estimation-based data fusion method.

[0142] Specifically, the uncertainty of the observation data and the uncertainty of the prior probability are combined to obtain a more accurate state estimation. Before use, the system prior probability distribution needs to be given as accurately as possible. It is a fusion algorithm based on probability statistics, which uses prior probability and new observation data to update posterior probability.

[0143] In single sensor detection, under certain assumptions, the conditional probability function of the single sensor decision result is the likelihood function p(z|Hi), and the concept of the output of the multi-sensor integrated fusion decision result is the likelihood function, that is, p(u|Hi), where u=(u1, u2,..., uN); In engineering applications, the consequences of various errors are not equally serious, that is, the losses or costs caused by different types of errors are different. In order to reflect these differences, different costs should be specified for each type of error probability, that is, the so-called cost function to reflect the difference in losses. In the data fusion process based on Bayesian estimation, the Bayesian fusion detection criterion assigns a corresponding cost value to each decision result, and the average total cost is obtained based on the assumption probability. The detection strategy is to minimize the average total cost. The specific steps include:

[0144] First, initialize the conditions, determine the observation distribution of each sensor, and assume the Bayesian decision threshold of each sensor:

[0145]

[0146]

[0147] From ; ;

[0148] Second, set the loop variable , and the termination control quantity , the preparation cycle. Estimate the Bayesian fusion detection criterion:

[0149]

[0150] Calculate the corresponding Bayesian fusion risk

[0151]

[0152] Third Step Calculate the new decision threshold based on the previously calculated Bayesian decision threshold of each sensor:

[0153]

[0154]

[0155]

[0156] Calculate the corresponding sensor based on the new threshold of each sensor ,

[0157] Calculate the Bayesian decision threshold of each sensor: .

[0158] Estimate the Bayesian fusion detection criterion:

[0159]

[0160] Fourth Step Calculate the Bayesian risk RB(K+1) of the k+1 iteration:

[0161]

[0162] Fifth Step After each sensor makes a decision, the decision result is sent to the fusion center to realize data fusion.

[0163] D. Fuzzy logic reasoning method refers to using a real number between 0 and 1 to represent the degree of truth or membership. In the process of multi-sensor fusion, fuzzy logic is used to handle uncertainty, and these uncertain factors are included in the reasoning process. Through the use of systematic methods to model the uncertainty in the fusion process and consistent reasoning based on fuzzy logic, more accurate and reliable fusion results are obtained.

[0164] E, Artificial neural network method refers to using deep learning model (such as CNN, LSTM) to learn the nonlinear mapping relationship of multi-sensor data, and gradually forming efficient logical reasoning ability by continuously training sample data, and using the advantages of signal processing and automatic reasoning function to realize accurate fusion of multi-sensor data. Neural network algorithm has excellent fault tolerance, self-adaptability, self-learning ability and self-organization ability, and can simulate extremely complex nonlinear mapping relationship. In a multi-sensor system, since the information provided by each sensor has certain uncertainty, therefore, the fusion of these uncertain information is essentially to carry out uncertainty reasoning.

[0165] F, Extended Kalman filter fusion method, Taylor expansion linearization is carried out on the nonlinear system, and then Kalman filter framework is applied, and Extended Kalman Filter (EKF) solves the problem of nonlinearity through local linearity. The derivative of the nonlinear prediction equation and the observation equation is calculated, and the tangent is replaced to linearize;

[0166] The prediction model and the measurement model of the extended Kalman are nonlinear, in order to simplify the calculation, the motion equation and the observation equation are linearized by first-order Taylor decomposition, the posterior probability density is described in Gaussian form, and when calculating the variance, the state transition matrix and the observation matrix are the Jacobian matrix of the state information.

[0167] The application can effectively retain the correlation between the feature values in the aforementioned multi-dimensional attribute vector based on typical correlation feature analysis, relative to the traditional feature superposition processing mode, without causing the feature vector to be too complex, the correlation features of the underwater target contained in the multi-dimensional data can be further strengthened and highlighted, so that the underwater target can better represent the characteristics of the underwater target in the motion state and in different marine environment conditions, and the recognizable ability in the motion state and in the changing environment is improved.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not limited to the scope of the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalent, without departing from the essence and scope of the technical solutions of the present application.

Claims

1. A method for optimizing acoustic signal features for underwater small target object analysis and identification, characterized in that, Including steps (i) to (iv); Step 1: Sound signal detection and acquisition; specifically, this refers to: An underwater acoustic signal acquisition system is established based on a fiber optic hydrophone array. For the acquired analog acoustic signals, an analog amplification filter is used to preprocess the signals, and an AD converter is used to convert the acquired analog acoustic signals into digital signals for digital filtering. Step 2: Underwater acoustic signal preprocessing, specifically including: (2.1) Remove interference lines from the power spectrum of the underwater acoustic signal (2.2) For data collection time-based data collection Acquired sound signals To avoid interference from motion-related changes in the acoustic signal caused by the target's motion, a time window is used to process the time-series-based acoustic signal. The time-reconstructed audio signal is obtained by truncating and splicing the data. ; (2.3) Due to the varying distances between the fiber optic hydrophone array and the small underwater target, as well as the influence of the target's motion, time-reconstructed acoustic signals may be generated. The amplitude is abnormal. In order to control the abnormal amplitude, the range transform is used to perform a linear transformation on the reconstructed acoustic signal to optimize the signal amplitude characteristics. (2.4) During the underwater transmission of acoustic signals, the attenuation of the high-frequency characteristics of the acoustic signal will far exceed the attenuation rate of the low-frequency characteristics due to the influence of the acoustic transmission characteristics. In order to balance the frequency characteristic distribution of the acoustic signal and improve the effectiveness of the acoustic signal, an Nth-order FIR digital high-pass filter is used to process the acoustic signal. (2.5) For underwater acoustic signal data of small target objects, underwater environmental noise can be suppressed by optimizing the signal-to-noise ratio; for underwater acoustic signals of small targets, they can be represented as the actual underwater acoustic signals of the small target objects. and ambient noise signals The convolution function is represented Logarithmic representation of convolution function based on Fast Fourier Transform Further analysis of the logarithmic expression signal Perform a Fourier transform to convert it to the frequency domain to obtain its spectral signal. Using homomorphic filters to convert the spectral signal in the frequency domain Multiplying the two results in the spectrum after homomorphic filtering. ; the filtered spectrum Transforming back to the spatial domain yields the filtered logarithmic expression. For the filtered logarithmic expression signal Perform exponential operations to obtain the optimized signal. ;in Indicates Fourier transform, Indicates inverse transformation; Step 3: Construction of acoustic signal features, specifically including: (3.1) Extracting the structural features of the optimized digital signal: The structural features refer to the multiple time-domain characteristic attributes of digital signals, which can be expressed as multidimensional feature vectors. ,in The optimized digital signal represents the first... Each time-domain feature attribute corresponds to an attribute vector. The time-domain feature attribute refers to several feature attributes related to the wavelength and amplitude of the digital signal. The corresponding attribute vector includes, but is not limited to, one or more of the following attribute vectors: A. Wavelength time-domain vector ; B. Zero-crossing wave distribution vector ; C. Dominant wavelength vector ; D. Wavelength distribution vector ; E, wavelength difference distribution vector ; F, amplitude distribution vector ; By combining the attribute vectors, we obtain the temporal feature attribute feature vectors. ; in, It refers to the average period corresponding to the wavelength. Indicates the testing cycle; Indicates that the wavelength is in the range The number of zero-crossing waves, Indicates the first One wavelength range; This indicates that the wavelength difference between adjacent zero-crossing waves lies within the interval Quantity, Indicates the first A wavelength difference interval; This indicates that the interpeak amplitude is within the amplitude range. The number of peaks, Indicates the first Each peak amplitude interval is obtained by dividing the maximum peak amplitude of the optimized digital signal into equal intervals. (3.2) Extracting the energy characteristics of the optimized digital signal: By performing wavelet packet decomposition on the optimized digital signal, wavelet packet coefficients at each level from low frequency to high frequency are extracted. , This represents the wavelet coefficients of the k-th node in the j-th layer. The digital signal is reconstructed based on the wavelet packet coefficients to obtain the energy characteristics of each frequency band. Further, the energy coefficients of each subband were obtained. ; Obtain the energy feature attribute vector ; (3.3) Extracting the frequency characteristics of the optimized digital signal: The acquired digital signal is transformed through a Mel filter bank to convert the frequency domain of the digital signal to the Mel-scale frequency domain; logarithmic operation is performed on the frequency domain signal after passing through the Mel filter bank to enhance the frequency domain feature intensity of the speech signal; Discrete cosine transform (DCT) is performed on the logarithmic spectrum obtained in the previous step to transform it to the cepstral space; the cepstral coefficients are normalized to enhance the stability of the signal; the cepstral coefficients of adjacent frames are concatenated to form the final frequency feature vector. (3.4) Extracting the acoustic intensity features of the optimized digital signal: The optimized digital signal is obtained, and the frequency domain signal is obtained through Discrete Fourier Transform. The sum of squares of the imaginary and real parts of the spectrum is obtained to obtain the power spectrum of the signal. ,in This indicates the real part extraction operation. This indicates the imaginary part extraction operation; Based on the critical frequency band analysis method, the power spectrum frequency The critical spectrum is obtained by mapping the axis to the critical frequency. Critical spectrum The frequency axis can be represented as ; Extracting each frequency band of the critical spectrum center frequency Calculate the high-end frequencies of each frequency band. low-end frequency and frequency band weighting coefficients The power spectrum With frequency band weighting coefficient Multiplying them yields the critical bandwidth spectrum. ;in: ； ； ； ； in, For N points, the frequency is... The reconstructed signal is obtained by reconstructing the signal using simulated equal-loudness curves. The reconstructed signal is compressed in loudness amplitude to obtain a sound intensity feature vector. ; ； ； Step 4: Acoustic signal feature fusion. Underwater target acoustic signal feature fusion processing is performed based on canonical correlation analysis; specifically including: (4.1) Based on the attribute vectors determined in step three, establish the corresponding feature vector space for underwater target signals. Assuming its sample space is , ; (4.2) For underwater target signals, solve for their values ​​in any two different eigenvector spaces. and eigenvectors in and Canonical correlation analysis algorithm is used to determine the canonical correlation features between pairwise feature vectors. ,in Represents a vector space group; (4.3) Constructing spatial correlation feature vectors , The combined feature formed by the aforementioned attribute vectors can then be expressed as: .

2. The acoustic signal feature optimization method for underwater small target object analysis and identification according to claim 1, characterized in that, Step (2.2) specifically includes: for the acoustic signal Assuming it uses a time window function as follows: Sufficient time is given to reconstruct the sound signal. It can be represented as: ;in Denotes the conjugate function. refers to sound signals angular frequency, It is the symbol for the imaginary part.

3. The acoustic signal feature optimization method for underwater small target object analysis and identification according to claim 1, characterized in that, In step (2.3), for the reconstructed acoustic signal Obtain its maximum amplitude. and minimum amplitude Amplitude; for reconstructed acoustic signals Observed amplitude The adjusted amplitude can be expressed as .

4. The acoustic signal feature optimization method for underwater small target object analysis and identification according to claim 1, characterized in that, The output of the FIR digital high-pass filter in step (2.4) can be expressed as: in This indicates the signal output. Indicates the filter order. Indicates the delay period. This indicates the transfer function corresponding to the filter. This represents the filter coefficients.

5. The acoustic signal feature optimization method for underwater small target object analysis and identification according to claim 1, characterized in that, The following algorithms can be used to fuse multi-sensor data according to actual needs: A. Data fusion method based on weighted average This method involves weighted averaging of data from different types of sensors to obtain a more accurate and reliable fusion result. It is simple, intuitive, and easy to implement. The fused data processed using the weighted averaging method can reduce the errors and uncertainties that may exist in data from a single sensor, improving the stability and robustness of the entire system. This can be expressed as follows: ; in It refers to the fused signal. This refers to the number of signals. It refers to the k-th signal. This refers to the weight of the k-th signal. B. Data fusion method based on Kalman filtering Specifically, the system state is estimated iteratively through two steps: prediction and update. In the prediction step, the dynamic model of the system is used to predict the state at the next moment, and in the update step, the observed data is used to correct the predicted value. That Its advantages lie in its ability to process noisy and uncertain data. By fusing data from different sensors, it can reduce the impact of noise and improve the accuracy and stability of the system. It also has a small memory footprint and fast processing speed, making it suitable for systems with high real-time requirements. It is often used for the fusion of low-level real-time dynamic multi-sensor data. C. Data fusion methods based on Bayesian estimation; Specifically, it combines the uncertainty of the observed data with the uncertainty of the prior probability to obtain a more accurate state estimate. Before use, the prior probability distribution of the system needs to be given as accurately as possible. It is a fusion algorithm based on probability statistics that uses the prior probability and new observed data to update the posterior probability. In single-sensor detection, under certain assumptions, the conditional probability function of the single sensor's decision result is the likelihood function p(z|Hi). The concept of the output of the fused decision result from multiple sensors is also the likelihood function, i.e., p(u|Hi), where u=(u1, u2, ..., uN). In engineering applications, the consequences of various errors are not equally severe; that is, the losses or costs caused by different types of errors are different. To reflect these differences, different costs should be specified for each type of error probability, i.e., the cost function, to reflect these differences in loss. In the data fusion process based on Bayesian estimation, the Bayesian fusion detection criterion assigns a corresponding cost to each decision result, obtains the average total cost based on the assumed probability, and the detection strategy is to minimize the average total cost. The specific steps include: The first step is to initialize the conditions, determine the observation distribution of each sensor, and assume the Bayesian decision threshold for each sensor: Depend on work out ; The second step is to set the loop variable. and termination control quantity Prepare for the loop; estimate the Bayesian fusion detection criteria: Calculate the corresponding Bayesian fusion risk The third step is to calculate a new decision threshold based on the Bayesian decision thresholds of each sensor calculated in the previous step: Calculate the corresponding sensor based on the new threshold of each sensor. , The calculation is based on the Bayesian decision thresholds of each sensor: ; Estimation of Bayesian fusion detection criteria: The fourth step is to calculate the Bayesian risk RB(K+1) for this iteration: The fifth step involves each sensor making its own decision and then sending the decision results to the fusion center to achieve data fusion. D. Fuzzy logic reasoning method refers to using a real number between 0 and 1 to represent the degree of truth or membership. In the process of multi-sensor fusion, fuzzy logic is used to handle uncertainty, and these uncertain factors are incorporated into the reasoning process. By adopting a systematic approach to model the uncertainty in the fusion process, and performing consistent reasoning based on fuzzy logic, more accurate and reliable fusion results can be obtained. E. Artificial neural network method refers to using deep learning models (such as CNN and LSTM) to learn the nonlinear mapping relationship of multi-sensor data. Through continuous training on sample data, it gradually forms an efficient logical reasoning ability. Utilizing its advantages in signal processing and automatic reasoning function, it achieves accurate fusion of multi-sensor data. Neural network algorithms have excellent fault tolerance, adaptability, self-learning ability, and self-organization ability, and can simulate extremely complex nonlinear mapping relationships. In multi-sensor systems, since the information provided by each sensor has a certain degree of uncertainty, the fusion of this uncertain information is essentially performing uncertainty reasoning. F. The Extended Kalman Filter (EKF) fusion method linearizes the nonlinear system using Taylor expansion, and then applies the Kalman filter framework. The EKF solves the nonlinear problem through local linearity; it linearizes the nonlinear prediction equation and observation equation by taking the derivatives and replacing them with tangents. The prediction and measurement models of the extended Kalman spectroscopy are nonlinear. To simplify the calculation, the motion and observation equations are linearized by first-order Taylor decomposition, and the posterior probability density is described in Gaussian form. When calculating the variance, the state transition matrix and the observation matrix are Jacobian matrices of the state information.

Images