A method, device and equipment for underwater passive target detection

Abstract

The invention relates to the technical field of underwater acoustic target detection, and discloses an underwater acoustic passive target detection method, device and equipment, comprising dividing an underwater acoustic passive noise signal collected by a preset vertical array in different sea areas into a plurality of frame signals, and constructing a source domain and a target domain data set; for each frame signal in the source domain and the target domain data set, extracting a plurality of low-order features, calculating the first-order differential form of each low-order feature, obtaining a plurality of statistical function values ​​corresponding to the low-order feature and its first-order differential form, and constructing an acoustic signal feature of each frame signal; based on a transfer learning method, using a source domain target detection model obtained by using the source domain data set as a target domain initial detection model, and fine-tuning using the target domain data set to obtain a target domain target detection model; inputting the acoustic signal features corresponding to the real-time underwater acoustic passive noise signal collected in real time in the sea area to be identified into the target domain target detection model, and obtaining an underwater acoustic passive target detection result.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic target detection, and in particular to an underwater acoustic passive target detection method, device and equipment. Background Art

[0002] Vertical array underwater acoustic target passive detection technology analyzes the passive sonar signals received by the underwater vertical array and detects underwater acoustic targets within a certain range based on the noise information radiated by the underwater acoustic targets. As the signal-to-noise ratio of underwater acoustic targets in complex marine environments becomes lower and lower, how to improve the performance of underwater acoustic target detection has become a core issue in the field of underwater acoustics such as target monitoring and intelligence collection, and has been the focus of domestic and foreign researchers. The rapid development of artificial intelligence technologies represented by deep learning and big data has also provided an opportunity for the development of underwater target recognition technology in the direction of intelligence.

[0003] Traditional target detection technology relies on the observer's vision and hearing. The observer observes the collected signals in real time to determine whether there is an underwater acoustic target. Later, with the continuous advancement of digital signal processing technology, the ability to extract, analyze and identify underwater acoustic signals has been significantly enhanced. For example, the patent CN202210217666, named an underwater acoustic detection device, system, target tracking method and device, uses an underwater acoustic detection device to emit and receive lasers; when the sound wave causes the refractive index of the medium in the underwater acoustic detection area to change, the light intensity signal received by the laser receiving module changes, and then the phase change of the underwater acoustic signal is determined according to the change in the light intensity signal; the underwater acoustic detection is carried out according to the acousto-optic effect, which effectively avoids the problem of ambiguity of the starting and ending points caused by mechanical vibration of traditional acoustic transducers and improves the detection accuracy. However, when using the change in light intensity signal to track underwater acoustic targets, due to the large scattering of light waves in water, it is difficult to propagate over long distances, so it cannot be used for long-distance underwater acoustic target detection.

[0004] In recent years, machine learning and artificial intelligence technologies have been widely used in the field of underwater acoustic target detection. Through big data analysis and pattern recognition, automatic detection of underwater acoustic targets can be achieved, and underwater operation efficiency and autonomous control capabilities can be improved. For example, patent CN202010427316, named a method for underwater acoustic target recognition based on weighted support vector machine, extracts frame features from the original underwater acoustic target signal, builds a weighted support vector machine recognition model, uses the grid search method to find the optimal kernel function parameters and penalty factors, and trains the weighted support vector machine recognition model to obtain a support vector machine model with underwater target recognition function, with a recognition accuracy of more than 80%; however, its feature differentiation of underwater acoustic targets in single-stage feature extraction is poor, and the recognition accuracy is low.

[0005] The current mainstream machine learning method is fully supervised learning, which manually extracts labels of a large number of hydroacoustic target signals and uses machine learning algorithms to detect hydroacoustic targets; the trained model can extract the acoustic feature differences between hydroacoustic target signals, which has advantages in complex environments. For example, the patent CN201910661350, entitled Intelligent Recognition Method for Hydroacoustic Targets Based on Big Data, collects a large number of hydroacoustic signals emitted by hydroacoustic targets, extracts appropriate features, and then marks the data with or without targets, and establishes a training sample set to train the artificial intelligence model; in actual use, the surrounding hydroacoustic signals are collected and features are extracted, and input into the trained artificial intelligence model to determine the presence and type of hydroacoustic targets, and can identify hydroacoustic targets at a longer distance; however, this method requires first collecting a large number of hydroacoustic signals emitted by hydroacoustic targets, and then marking the data, which has high manpower and material costs, and the recognition accuracy is significantly reduced under small sample data. Moreover, the model obtained by training based on a single data set is difficult to be applied to other waters by direct transplantation, and the network is difficult to generalize and apply directly. Therefore, how to quickly and accurately migrate the pre-trained model to new waters in the complex and changeable underwater acoustic environment requires further research and exploration.

[0006] In summary, due to the changeable and complex underwater acoustic environment, it is difficult for the target detection method based on feature combination model to use a model trained with water data to detect targets in multiple different waters. In addition, it is difficult to collect underwater acoustic signal data, and data labeling requires a lot of manpower and material resources. The amount of labeled data is small, and it is not enough to re-acquire the supervised training model for target detection in new waters. Therefore, the existing underwater acoustic detection method cannot be applied to different waters and has poor versatility. When facing a variety of different waters, it is necessary to repeat data collection and model training, resulting in low target detection efficiency and poor recognition accuracy. Summary of the invention

[0007] To this end, the technical problem to be solved by the present invention is to overcome the problem that the existing technology cannot be applied to different waters. When facing a variety of different waters, data collection and model training need to be repeated, resulting in low target detection efficiency and poor recognition accuracy.

[0008] In order to solve the above technical problems, the present invention provides a method for underwater acoustic passive target detection, comprising:

[0009] Using a preset vertical array, in a sea area with the same detection task as the sea area to be identified, underwater passive noise signals of a preset duration are collected, and divided into multiple sub-frame signals according to a preset frame length, and a true label is marked for each sub-frame signal to construct a source domain dataset;

[0010] Using a preset vertical array, in the sea area to be identified, underwater passive noise signals of preset duration are collected, and divided into multiple frame signals according to preset frame lengths. Some frame signals are selected, and real labels are marked for them to construct the target domain dataset.

[0011] For each frame signal in the source domain data set and the target domain data set, extract multiple low-order features of the frame signal;

[0012] For each low-order feature in each frame signal, multiple statistical function values ​​corresponding to each low-order feature are calculated;

[0013] Calculate the first-order difference form corresponding to each low-order feature in each frame signal, and multiple statistical function values ​​corresponding to each first-order difference form;

[0014] Based on multiple statistical function values ​​corresponding to each low-order feature and its first-order differential form in each frame signal, the acoustic signal feature of each frame signal is constructed;

[0015] Using the acoustic signal features of the framed signals in the source domain data set, the preset neural network model is trained until the preset number of training times is reached to obtain a source domain target detection model;

[0016] The source domain target detection model is used as the target domain initial detection model, and the target domain initial detection model is fine-tuned using the acoustic signal features of the framed signals in the target domain dataset until a preset number of training times is reached to obtain the target domain target detection model;

[0017] The real-time underwater acoustic passive noise signal collected in the sea area to be identified is divided into multiple sub-frame signals according to the preset frame length, the acoustic signal characteristics of each sub-frame signal are obtained, and the target detection model in the target domain is input to obtain the underwater acoustic passive target detection result corresponding to each sub-frame signal in the real-time underwater acoustic passive noise signal.

[0018] Preferably, the various low-order features of the framed signal include: sum of auditory spectra, sum of RASTA filtered auditory spectra, RMS energy, signal zero-crossing rate, RASTA type auditory spectrum, frequency bands 1-26, Mel-cephalometric coefficients 1-14, spectral energy, spectral attenuation points, spectral flux, spectral centroid, spectral entropy, spectral slope, acoustic sharpness, acoustic harmony, spectral variance, spectral skewness, spectral kurtosis, fundamental frequency, sound probability detection, inter-frame jitter, differential inter-frame jitter and inter-frame flicker.

[0019] Preferably, multiple statistical function values ​​include: quartiles 1-3, tertile range, 1% percentile, 99% percentile, percentile range 1-99%, minimum / maximum position, minimum / maximum range, arithmetic mean, root quadratic mean, contour centroid, flatness, standard deviation, skewness, kurtosis, ratio of relative LLD duration above 25%, 50%, 75%, 90%, continuous rise in relative LLD, duration of positive curvature of LLD curve, standard deviation of mean, standard deviation of maximum, standard deviation of minimum, standard deviation of paragraph length, peak mean, arithmetic mean of peak mean, mean of peak distance, standard deviation of peak distance, mean of peak amplitude, mean of minimum amplitude, amplitude range of peak, mean and standard deviation of rising slope, mean and standard deviation of falling slope, slope, intercept and quadratic error of linear regression, coefficient, intercept and quadratic error of quadratic regression, and percentage of non-zero frames.

[0020] Preferably, when constructing the target domain data set, the selected partial frame signals account for 20% of all frame signals.

[0021] Preferably, the same detection mission includes the same hydroacoustic target type; the hydroacoustic target type includes submarines and surface ships.

[0022] Preferably, the preset neural network model includes one of a BP neural network model, a deep learning network model, and a support vector machine model.

[0023] Preferably, the underwater acoustic passive target detection result is a binary classification result, including: including the underwater acoustic target and not including the underwater acoustic target.

[0024] Preferably, the array elements of the preset vertical array are evenly distributed in a preset area according to a preset spacing.

[0025] This embodiment provides an underwater acoustic passive target detection device, including:

[0026] The data set construction module uses a preset vertical array to collect underwater passive noise signals of a preset duration in a sea area with the same detection task as the sea area to be identified, and divides them into multiple sub-frame signals according to a preset frame length, marks the true label for each sub-frame signal, and constructs a source domain data set; uses a preset vertical array to collect underwater passive noise signals of a preset duration in the sea area to be identified, and divides them into multiple sub-frame signals according to a preset frame length, selects some sub-frame signals, marks their true labels, and constructs a target domain data set;

[0027] A low-order feature extraction module extracts a variety of low-order features of each frame signal in the source domain data set and the target domain data set;

[0028] The function value calculation module calculates, for each low-order feature in each frame signal, multiple statistical function values ​​corresponding to each low-order feature; calculates the first-order differential form corresponding to each low-order feature in each frame signal, and multiple statistical function values ​​corresponding to each first-order differential form;

[0029] An acoustic signal feature construction module, which constructs the acoustic signal feature of each frame signal based on multiple statistical function values ​​corresponding to each low-order feature and its first-order differential form in each frame signal;

[0030] The model training module uses the acoustic signal features of the framed signals in the source domain data set to train the preset neural network model until the preset number of training times is reached, and the source domain target detection model is obtained; the source domain target detection model is used as the target domain initial detection model, and the target domain initial detection model is trained using the acoustic signal features of the framed signals in the target domain data set until the preset number of training times is reached, and the target domain target detection model is obtained;

[0031] The detection module divides the real-time underwater acoustic passive noise signal collected in the sea area to be identified into multiple sub-frame signals according to a preset frame length, obtains the acoustic signal characteristics of each sub-frame signal, inputs the target detection model in the target domain, and obtains the underwater acoustic passive target detection result corresponding to each sub-frame signal in the real-time underwater acoustic passive noise signal.

[0032] This embodiment provides an underwater acoustic passive target detection device, including:

[0033] A signal acquisition device, comprising a plurality of sensors distributed in a preset vertical array, for real-time acquisition of real-time underwater passive noise signals in the sea area to be identified;

[0034] An underwater acoustic passive target detection device is communicatively connected with the signal acquisition device, and acquires an underwater acoustic passive target detection result corresponding to each frame signal in the real-time underwater acoustic passive noise signal based on the real-time underwater acoustic passive noise signal;

[0035] A display is communicatively connected to the underwater acoustic passive target detection device and is used to display the underwater acoustic passive target detection results.

[0036] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0037] The underwater acoustic passive target detection method described in the present invention collects underwater acoustic passive noise signals using a preset vertical array, and after dividing them into multiple frame signals, extracts multiple low-order features of each frame signal, as well as the first-order differential form of each low-order feature; calculates multiple statistical function values ​​of each low-order feature and its first-order differential form, constructs acoustic signal features, analyzes the time-frequency characteristics of the underwater acoustic passive noise signal, and more comprehensively describes the time domain characteristics and frequency domain characteristics of the underwater acoustic passive noise signal, thereby improving the accuracy of underwater acoustic passive target detection based on acoustic signal features; and the present invention is based on the model parameter sharing method in transfer learning, fully utilizes the source domain data set to train the model, and at the same time, uses a small amount of labeled data to fine-tune the model in the target domain to improve the detection performance, reduce the demand for samples with labeled data in the target domain, and greatly saves the cost of manual labeling, thereby greatly reducing the training time of the target domain, and improving the training efficiency and generalization ability of the detection model under small sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0039] Figure 1 It is a flow chart of the steps of the underwater acoustic passive target detection method provided by the present invention;

[0040] Figure 2 It is a schematic diagram of the principle of underwater acoustic passive target detection;

[0041] Figure 3 It is a schematic diagram of transfer learning. DETAILED DESCRIPTION

[0042] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0043] Reference Figure 1 As shown, the process flow chart of the underwater acoustic passive target detection method of the present invention includes the following specific steps:

[0044] S101: using a preset vertical array, in a sea area with the same detection task as the sea area to be identified, collecting an underwater passive noise signal of a preset duration, dividing it into a plurality of sub-frame signals according to a preset frame length, marking a true label for each sub-frame signal, and constructing a source domain data set;

[0045] The same detection mission includes the same hydroacoustic target type; the hydroacoustic target type includes submarines and surface ships;

[0046] S102: using a preset vertical array, in the sea area to be identified, collecting underwater passive noise signals of a preset duration, dividing them into a plurality of sub-frame signals according to a preset frame length, selecting some of the sub-frame signals, marking them with true labels, and constructing a target domain data set;

[0047] S103: for each frame signal in the source domain data set and the target domain data set, extract multiple low-order features of the frame signal;

[0048] S104: for each low-order feature in each frame signal, calculating multiple statistical function values ​​corresponding to each low-order feature;

[0049] S105: Calculating a first-order differential form corresponding to each low-order feature in each frame signal, and multiple statistical function values ​​corresponding to each first-order differential form;

[0050] S106: constructing an acoustic signal feature of each frame signal based on multiple statistical function values ​​corresponding to each low-order feature and its first-order differential form in each frame signal;

[0051] S107: using the acoustic signal features of the framed signals in the source domain data set, training the preset neural network model until a preset number of training times is reached, and obtaining a source domain target detection model;

[0052] The preset neural network model includes one of a BP neural network model, a deep learning network model, and a support vector machine model;

[0053] S108: using the source domain target detection model as the target domain initial detection model, and fine-tuning the target domain initial detection model using the acoustic signal features of the framed signals in the target domain dataset until a preset number of training times is reached, thereby obtaining the target domain target detection model;

[0054] S109: dividing the real-time underwater acoustic passive noise signal collected in real time in the sea area to be identified into a plurality of sub-frame signals according to a preset frame length, obtaining the acoustic signal feature of each sub-frame signal, inputting the target domain target detection model, and obtaining the underwater acoustic passive target detection result corresponding to each sub-frame signal in the real-time underwater acoustic passive noise signal;

[0055] The underwater acoustic passive target detection result is a binary classification result, including: including the underwater acoustic target and not including the underwater acoustic target.

[0056] Specifically, the vertical array underwater acoustic target passive detection technology realizes the detection of underwater acoustic targets within a certain range by analyzing the passive sonar signals received by the underwater vertical array and based on the noise information radiated by the underwater acoustic target. In this embodiment, the array elements of the preset vertical array are evenly distributed in the preset area according to the preset spacing.

[0057] Specifically, various low-order features of the framed signal include: sum of auditory spectrum, sum of RASTA filtered auditory spectrum, RMS energy, signal zero crossing rate, RASTA type auditory spectrum, frequency band 1-26, Mel-frequency cepstral coefficients (MFCC) 1-14, spectral energy, spectral decay point, spectral flux, spectral centroid, spectral entropy, spectral slope, acoustic sharpness, acoustic harmony, spectral variance, spectral skewness, spectral kurtosis, fundamental frequency, sound probability detection, inter-frame jitter, differential inter-frame jitter and inter-frame flicker.

[0058] Specifically, various statistical function values ​​include: quartiles 1-3, tertile range, 1% percentile, 99% percentile, percentile range 1-99%, minimum / maximum position, minimum / maximum range, arithmetic mean, root quadratic mean, contour centroid, flatness, standard deviation, skewness, kurtosis, ratio of relative LLD duration above 25%, 50%, 75%, 90%, continuous rise in relative LLD, duration of positive curvature of LLD curve, standard deviation of mean, standard deviation of maximum, standard deviation of minimum, standard deviation of paragraph length, peak mean, arithmetic mean of peak mean, mean of peak distance, standard deviation of peak distance, mean of peak amplitude, mean of minimum amplitude, amplitude range of peak, mean and standard deviation of rising slope, mean and standard deviation of falling slope, slope, intercept and quadratic error of linear regression, coefficient, intercept and quadratic error of quadratic regression, and percentage of non-zero frames.

[0059] Based on the above description, in an embodiment of the present invention, for each frame signal, a total of 64 low-order features are counted, and 64 first-order differential forms are obtained accordingly; 50 statistical function values ​​of each low-order feature and 50 statistical function values ​​of each first-order differential form are calculated. At this time, theoretically, 64*2*50=6400-dimensional features are obtained. In actual calculation, some descriptors do not need to calculate some statistical functions, totaling 27, that is, 6400-27=6373-dimensional features are finally calculated, which are used as the acoustic signal features corresponding to each frame signal, and have good distinguishing ability for the detection of actual underwater acoustic target signals.

[0060] Specifically, in an embodiment of the present invention, when constructing a target domain data set, the selected partial frame signals account for 20% of all frame signals; after obtaining the source domain target detection model trained using the source domain data set, it is used as the target domain initial detection model, and the model is fine-tuned using the 20% labeled frame signals to obtain the target domain target detection model. Using the transfer learning method, the knowledge and experience of a source domain are transferred to the target domain, thereby accelerating and improving the learning and problem-solving capabilities of the target domain. In the field of underwater acoustic target detection with little labeled data, transfer learning can make full use of the model pre-trained using a large amount of data in the source domain, and use the target domain's less labeled data to fine-tune the pre-trained model to improve the detection performance, thereby greatly reducing the data requirements and training time of the target domain, and improving the training efficiency and generalization ability of the underwater acoustic target detection model in the case of small samples.

[0061] Based on the above embodiments, in the embodiments of the present invention, the underwater acoustic passive target detection method provided by the present invention is used to perform target detection. Figure 2 The figure shows a schematic diagram of the principle of underwater acoustic passive target detection, which specifically includes: underwater acoustic target acoustic feature extraction and model migration; after processing the underwater acoustic passive noise signal collected by the vertical array through underwater acoustic target acoustic feature extraction, the source domain labeled data set and the target domain labeled data set are obtained, and the labeled data set is used as the input for subsequent model training and fine-tuning; model migration is divided into source domain model pre-training and target domain model migration fine-tuning; specifically includes:

[0062] S201: Acoustic feature extraction of underwater acoustic targets;

[0063] S201-1: Use a vertical array to collect underwater passive noise signals in the source domain and the target domain respectively as the original underwater acoustic data, including a large number of source domain signals and a small number of target domain signals;

[0064] S201-2: Extract the acoustic features of underwater acoustic targets from the collected underwater acoustic passive noise signals, including the low-order descriptive features of the signals and the statistical function values ​​of these features;

[0065] The 64 low-level descriptors LLD (Low-Level Descriptors) of the signal are shown in Table 1:

[0066] Table 1 Low-level descriptors

[0067] name Number of low-level descriptors Auditory Spectrum Sum (Loudness) 1 RASTA filtered auditory spectrum sum 1 RMS Energy 1 Zero crossing rate 1 RASTA-type auditory spectrum, frequency bands 1-26 (0-8kHz) 26 MFCC 1–14 14 Spectral energy (250-650Hz, 1k-4kHz) 2 Spectrum attenuation points (0.25, 0.50, 0.75, 0.90) 4 Spectral flux, spectral centroid, spectral entropy, spectral slope 4 Acoustic sharpness, acoustic harmony 2 Spectral variance, spectral skewness, spectral kurtosis 3 Baseband 1 Sound Probability Detection 1 Inter-frame jitter, differential inter-frame jitter, inter-frame flicker 3 total 64

[0068] Since the spectrum of the passive noise signal radiated by the underwater acoustic target includes two parts: the line spectrum and the low-frequency continuous spectrum. The line spectrum has better phase stability and higher energy. Therefore, in the analysis of the spectrum characteristics of the underwater acoustic signal, the main focus is on analyzing the line spectrum signal radiated by the underwater acoustic target. However, the underwater acoustic environment is complex and changeable, and may change significantly in time and space. Therefore, it is not enough to rely solely on a single time domain waveform analysis or spectrum feature analysis. In order to more comprehensively describe the time domain and frequency domain characteristics of the signal, it is also necessary to analyze the extracted time-frequency characteristics of the acoustic signal.

[0069] This embodiment uses a variety of statistical functions to analyze the time-frequency characteristics of acoustic signals. The values ​​of the various statistical functions are shown in Table 2:

[0070] Table 2 Function statistics

[0071] name Counting the number of functions Quartile 1-3, three-point range 6 1% percentile, 99% percentile 2 Percentile range 1-99% 1 Min / max position, range 4 Arithmetic mean, root quadratic mean 2 Contour centroid, flatness 2 Standard Deviation, Skewness, Kurtosis 3 Relative LLD duration is above 25%, 50%, 75%, 90% 4 Relative LLD continues to rise 1 The duration of the positive curvature of the LLD curve 1 Mean, maximum, minimum, and standard deviation of paragraph length 4 Peak Mean 1 Peak mean - arithmetic mean 1 Peak-to-peak distance mean and standard deviation 2 Amplitude mean of peak / minimum values 2 Peak amplitude range 2 Mean / standard deviation of the upward / downward slope 4 Linear regression slope, intercept, quadratic error 3 Quadratic regression a, b, intercept, quadratic error 4 Non-zero frame percentage 1 total 50

[0072] This embodiment calculates the statistical function of the obtained low-order descriptor and its first-order differential form, and finally obtains the 6373-dimensional acoustic features of the underwater acoustic target. The classification model is trained using traditional methods to verify the quality of the extracted acoustic features. The extracted features will be used as input data sets to train and fine-tune the detection models of the source domain and the target domain.

[0073] S202: Model migration method;

[0074] The present invention uses a small sample target detection method based on model parameter sharing in the model migration part. The method uses deep learning or machine learning technology to process target detection with a small amount of sample data. The core idea of ​​the method is to effectively utilize sample data by sharing parameters in the model, thereby achieving accurate detection of the target. In traditional deep learning methods, a large amount of sample data is usually required to train the model in order to obtain good generalization ability. However, in the field of underwater acoustic target detection, it may be difficult or even impossible to obtain a large amount of labeled sample data, which requires a lot of manpower and material resources. Therefore, the small sample target detection method based on model parameter sharing attempts to overcome the problem of data scarcity by sharing model parameters.

[0075] Reference Figure 3As shown in the figure, it is a schematic diagram of transfer learning. In the field of transfer learning, source domain and target domain are two important concepts, which are usually used to describe different data distributions or tasks. The source domain refers to the existing data distribution used as a reference in transfer learning, and has large-scale data. In model migration, the source domain detection model has been trained on the source domain labeled data set, which can provide rich feature representation and detection knowledge. The data in the target domain is very limited, and it is difficult to independently train the detection model to complete the target detection task. The water conditions and detection targets in the target domain are different from those in the source domain, so the target detection model pre-trained in the source domain cannot be directly applied to the target domain. Therefore, it is necessary to use the transfer learning method to transfer the knowledge of the source domain to the target domain, and learn a model with good generalization performance in the target domain, so as to solve the underwater acoustic target detection task in the target domain.

[0076] In transfer learning in the image field, the similarity between the source domain and the target domain is an important consideration, but it is not absolutely necessary. The passive noise signal radiated by the underwater acoustic target has special characteristics in the spectrum, including better phase stability and higher energy line spectrum. In the process of transfer learning, adding this knowledge can achieve the use of a small amount of target domain signal data to train a detection model with excellent performance. Therefore, the underwater acoustic passive noise signal with similar detection tasks as the target domain is selected as the source domain. Similar data distribution and task type can further optimize the detection model.

[0077] In specific use, select source domain signals with similar detection tasks to the target domain, obtain source domain labeled data sets through passive noise signal feature extraction, and use the source domain labeled data sets to train the source domain detection model. Share the pre-trained model parameters to the target domain to provide initialization model parameters for the target domain detection model. Then, perform passive noise signal feature extraction on the collected target domain signals to obtain a small amount of target domain labeled data. Use a small amount of target domain labeled data to fine-tune the model parameters migrated from the source domain according to the requirements of the target detection task, and make full use of the knowledge and features learned by the source domain detection model to make the target domain detection model have better performance. After fine-tuning is completed, use the target domain verification to evaluate the performance of the model on the underwater acoustic target detection task.

[0078] Specifically, based on the above embodiment, in the embodiment of the present invention, a marine test was carried out in a certain sea area using the underwater acoustic passive target detection method provided by the present invention, and the real ship radiated noise data under the actual sea conditions was collected. The data was collected in 3 different sea areas for 3 times. The specific data description is shown in Table 3:

[0079] Table 3 Experimental data

[0080]

[0081] Based on this data, specific target recognition experimental tests were carried out.

[0082] S301: Extracting acoustic signal features, mainly including:

[0083] (1) Split the long sound signal data into one-second segments;

[0084] (2) Divide each segment of sound signal data into frames with a frame length of 50ms;

[0085] (3) Calculate 64 low-level descriptors (LLD) and their first-order difference forms, including zero-crossing rate (ZCR), root mean square energy (E), fundamental frequency (F0), Mel-frequency cepstral coefficients 1-14 (MFCC1-14), spectral centroid, spectral sharpness, spectral entropy, and spectral skewness, on a frame-by-frame basis.

[0086] (4) Calculate statistical functions of the obtained low-order descriptors and their first-order difference forms, including mean, variance, standard deviation, maximum value, minimum value, kurtosis, skewness, maximum and minimum value range, quartiles, etc., to form 6373-dimensional features.

[0087] S302: Randomly select 2000 data with / without targets in the source domain to form a source domain labeled data set, and divide the training data set and the test data set into a ratio of 3:1.

[0088] S303: Train the source domain target detection model, use the BP neural network model as the classifier, and set the parameters as follows: 6373 input neurons, 1 hidden layer, 80 hidden neurons, S-type transfer function as activation function, 2 output neurons, gradient descent BP algorithm training function, loss function as mean square error MSE, required training accuracy of 10-9, maximum number of training times of 1000, and learning rate of 0.01.

[0089] S304: Input the source domain training data set for training, and stop training when the required training accuracy or the maximum number of training times is reached to obtain the source domain pre-training model. Input the source domain test data set for testing, and output the source domain detection accuracy.

[0090] S305: 400 data are randomly selected from the target domain and labeled with / without target, and the remaining data are not labeled to form a target domain labeled data set, which is divided into a training data set and a test data set in a ratio of 3:1.

[0091] S306: Build a target domain target detection model. The model structure and training parameters are consistent with those of the source domain. Extract model parameters from the source domain pre-trained model and share them with the target domain detection model as the target domain model initialization parameters.

[0092] S307: Input the target domain training data set to fine-tune the model. When the training accuracy requirement is reached or the maximum number of training times is reached, stop training to obtain the target domain object detection model. Input the target domain test data set to test and output the source domain detection accuracy.

[0093] S308: Using sea area A as the source domain and sea area B as the target domain: the source domain detection accuracy is 75.4%, and the target domain detection accuracy is 98.08%; using sea area C as the source domain and sea area B as the target domain: the source domain detection accuracy is 82.1%, and the target domain detection accuracy is 97.53%.

[0094] Based on the above data, it can be seen that the underwater acoustic passive target detection method provided by the present invention has a high model detection accuracy after migration; and based on the model parameter sharing method, a small amount of labeled data can be used to train a model detection accuracy equivalent to that of a large amount of labeled data, thereby improving the training efficiency of the detection model and being able to quickly match the model to the target water environment in practical applications.

[0095] Based on the above embodiments, an embodiment of the present invention further provides an underwater acoustic passive target detection device, which may include:

[0096] The data set construction module uses a preset vertical array to collect underwater passive noise signals of a preset duration in a sea area with the same detection task as the sea area to be identified, and divides them into multiple sub-frame signals according to a preset frame length, marks the true label for each sub-frame signal, and constructs a source domain data set; uses a preset vertical array to collect underwater passive noise signals of a preset duration in the sea area to be identified, and divides them into multiple sub-frame signals according to a preset frame length, selects some sub-frame signals, marks their true labels, and constructs a target domain data set;

[0097] A low-order feature extraction module extracts a variety of low-order features of each frame signal in the source domain data set and the target domain data set;

[0098] The function value calculation module calculates, for each low-order feature in each frame signal, multiple statistical function values ​​corresponding to each low-order feature; calculates the first-order differential form corresponding to each low-order feature in each frame signal, and multiple statistical function values ​​corresponding to each first-order differential form;

[0099] An acoustic signal feature construction module, which constructs the acoustic signal feature of each frame signal based on multiple statistical function values ​​corresponding to each low-order feature and its first-order differential form in each frame signal;

[0100] The model training module uses the acoustic signal features of the framed signals in the source domain data set to train the preset neural network model until the preset number of training times is reached, and the source domain target detection model is obtained; the source domain target detection model is used as the target domain initial detection model, and the target domain initial detection model is trained using the acoustic signal features of the framed signals in the target domain data set until the preset number of training times is reached, and the target domain target detection model is obtained;

[0101] The detection module divides the real-time underwater acoustic passive noise signal collected in the sea area to be identified into multiple sub-frame signals according to a preset frame length, obtains the acoustic signal characteristics of each sub-frame signal, inputs the target detection model in the target domain, and obtains the underwater acoustic passive target detection result corresponding to each sub-frame signal in the real-time underwater acoustic passive noise signal.

[0102] The underwater acoustic passive target detection device of this embodiment is used to implement the aforementioned underwater acoustic passive target detection method. Therefore, the specific implementation of the underwater acoustic passive target detection device can be seen in the embodiment part of the underwater acoustic passive target detection method in the above text. For example,

[0103] A data set construction module is used to implement steps S101 and S102 in the above-mentioned underwater acoustic passive target detection method; a low-order feature extraction module is used to implement step S103 in the above-mentioned underwater acoustic passive target detection method; a function value calculation module is used to implement steps S104 and S105 in the above-mentioned underwater acoustic passive target detection method; an acoustic signal feature construction module is used to implement step S106 in the above-mentioned underwater acoustic passive target detection method; a model training module is used to implement steps S107 and S108 in the above-mentioned underwater acoustic passive target detection method; a detection module is used to implement step S109 in the above-mentioned underwater acoustic passive target detection method, so its specific implementation method can refer to the description of the corresponding various partial embodiments, which will not be repeated here.

[0104] Based on the above embodiment, in an embodiment of the present invention, a hydroacoustic passive target detection device is also provided, including:

[0105] A signal acquisition device, comprising a plurality of sensors distributed in a preset vertical array, for real-time acquisition of real-time underwater passive noise signals in the sea area to be identified;

[0106] An underwater acoustic passive target detection device is communicatively connected with the signal acquisition device, and acquires an underwater acoustic passive target detection result corresponding to each frame signal in the real-time underwater acoustic passive noise signal based on the real-time underwater acoustic passive noise signal;

[0107] A display is communicatively connected to the underwater acoustic passive target detection device and is used to display the underwater acoustic passive target detection results.

[0108] The underwater acoustic passive target detection method described in the present invention collects underwater acoustic passive noise signals using a preset vertical array, and after dividing them into multiple frame signals, extracts multiple low-order features of each frame signal, as well as the first-order differential form of each low-order feature; calculates multiple statistical function values ​​of each low-order feature and its first-order differential form, constructs acoustic signal features, analyzes the time-frequency characteristics of the underwater acoustic passive noise signal, and more comprehensively describes the time domain characteristics and frequency domain characteristics of the underwater acoustic passive noise signal, thereby improving the accuracy of underwater acoustic passive target detection based on acoustic signal features; and the present invention is based on the model parameter sharing method in transfer learning, fully utilizes the source domain data set to train the model, and at the same time, uses a small amount of labeled data to fine-tune the model in the target domain to improve the detection performance, reduce the demand for samples with labeled data in the target domain, and greatly saves the cost of manual labeling, thereby greatly reducing the training time of the target domain, and improving the training efficiency and generalization ability of the detection model under small sample conditions.

[0109] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0110] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0111] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0112] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0113] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A method for underwater acoustic passive target detection, characterized in that: include: Using a preset vertical array, in a sea area with the same detection task as the sea area to be identified, underwater passive noise signals of a preset duration are collected, and divided into multiple sub-frame signals according to a preset frame length, and a true label is marked for each sub-frame signal to construct a source domain dataset; Using a preset vertical array, in the sea area to be identified, underwater passive noise signals of preset duration are collected, and divided into multiple frame signals according to preset frame lengths. Some frame signals are selected, and real labels are marked for them to construct the target domain dataset. For each frame signal in the source domain data set and the target domain data set, extract multiple low-order features of the frame signal; For each low-order feature in each frame signal, multiple statistical function values ​​corresponding to each low-order feature are calculated; Calculate the first-order difference form corresponding to each low-order feature in each frame signal, and multiple statistical function values ​​corresponding to each first-order difference form; Based on multiple statistical function values ​​corresponding to each low-order feature and its first-order differential form in each frame signal, the acoustic signal feature of each frame signal is constructed; Using the acoustic signal features of the framed signals in the source domain data set, the preset neural network model is trained until the preset number of training times is reached to obtain a source domain target detection model; The source domain target detection model is used as the target domain initial detection model, and the target domain initial detection model is fine-tuned using the acoustic signal features of the framed signals in the target domain dataset until a preset number of training times is reached to obtain the target domain target detection model; The real-time underwater acoustic passive noise signal collected in the sea area to be identified is divided into multiple sub-frame signals according to the preset frame length, the acoustic signal characteristics of each sub-frame signal are obtained, and the target detection model in the target domain is input to obtain the underwater acoustic passive target detection result corresponding to each sub-frame signal in the real-time underwater acoustic passive noise signal.

2. The underwater acoustic passive target detection method according to claim 1, characterized in that: Various low-level features of framed signals, including: auditory spectrum sum, RASTA filtered auditory spectrum sum, RMS energy, signal zero crossing rate, RASTA type auditory spectrum, frequency bands 1-26, Mel-cephalometric coefficients 1-14, spectral energy, spectral decay point, spectral flux, spectral centroid, spectral entropy, spectral slope, acoustic sharpness, acoustic harmony, spectral variance, spectral skewness, spectral kurtosis, fundamental frequency, sound probability detection, inter-frame jitter, differential inter-frame jitter and inter-frame flicker.

3. The underwater acoustic passive target detection method according to claim 1, characterized in that: Various statistical function values, including: quartiles 1-3, tertile range, 1% percentile, 99% percentile, percentile range 1-99%, minimum / maximum position, minimum / maximum range, arithmetic mean, root quadratic mean, contour centroid, flatness, standard deviation, skewness, kurtosis, ratio of relative LLD duration above 25%, 50%, 75%, 90% range, relative LLD continuous rise, duration of positive curvature of LLD curve, standard deviation of mean, standard deviation of maximum, standard deviation of minimum, standard deviation of paragraph length, peak mean, arithmetic mean of peak mean, mean of peak distance, standard deviation of peak distance, mean of peak amplitude, mean of minimum amplitude, amplitude range of peaks, mean and standard deviation of rising slope, mean and standard deviation of falling slope, slope, intercept and quadratic error of linear regression, coefficient, intercept and quadratic error of quadratic regression, and percentage of non-zero frames.

4. The underwater acoustic passive target detection method according to claim 1, characterized in that: When constructing the target domain dataset, the selected partial frame signals account for 20% of all frame signals.

5. The underwater acoustic passive target detection method according to claim 1, characterized in that: The same detection mission includes the same hydroacoustic target type; the hydroacoustic target type includes submarines and surface ships.

6. The underwater acoustic passive target detection method according to claim 1, characterized in that: The preset neural network model includes one of a BP neural network model, a deep learning network model, and a support vector machine model.

7. The underwater acoustic passive target detection method according to claim 1, characterized in that: The underwater acoustic passive target detection result is a binary classification result, including: including the underwater acoustic target and not including the underwater acoustic target.

8. The underwater acoustic passive target detection method according to claim 1, characterized in that: The array elements of the preset vertical array are evenly distributed in the preset area according to the preset spacing.

9. An underwater acoustic passive target detection device, characterized in that: include: The dataset construction module uses a preset vertical array to collect underwater passive noise signals of a preset duration in a sea area with the same detection task as the sea area to be identified, and divides them into multiple sub-frame signals according to the preset frame length, marks the true label for each sub-frame signal, and constructs the source domain dataset; Using a preset vertical array, in the sea area to be identified, underwater passive noise signals of preset duration are collected, and divided into multiple frame signals according to preset frame lengths. Some frame signals are selected, and real labels are marked for them to construct the target domain dataset. A low-order feature extraction module extracts a variety of low-order features of each frame signal in the source domain data set and the target domain data set; The function value calculation module calculates, for each low-order feature in each frame signal, multiple statistical function values ​​corresponding to each low-order feature; calculates the first-order differential form corresponding to each low-order feature in each frame signal, and multiple statistical function values ​​corresponding to each first-order differential form; An acoustic signal feature construction module, which constructs the acoustic signal feature of each frame signal based on multiple statistical function values ​​corresponding to each low-order feature and its first-order differential form in each frame signal; The model training module uses the acoustic signal features of the framed signals in the source domain data set to train the preset neural network model until the preset number of training times is reached to obtain the source domain target detection model; The source domain target detection model is used as the target domain initial detection model, and the target domain initial detection model is trained using the acoustic signal features of the framed signals in the target domain dataset until a preset number of training times is reached to obtain the target domain target detection model; The detection module divides the real-time underwater acoustic passive noise signal collected in the sea area to be identified into multiple sub-frame signals according to a preset frame length, obtains the acoustic signal characteristics of each sub-frame signal, inputs the target detection model in the target domain, and obtains the underwater acoustic passive target detection result corresponding to each sub-frame signal in the real-time underwater acoustic passive noise signal.

10. An underwater acoustic passive target detection device, characterized in that: include: A signal acquisition device, comprising a plurality of sensors distributed in a preset vertical array, for real-time acquisition of real-time underwater passive noise signals in the sea area to be identified; The underwater acoustic passive target detection device according to claim 9 is communicatively connected to the signal acquisition device, and obtains the underwater acoustic passive target detection result corresponding to each frame signal in the real-time underwater acoustic passive noise signal based on the real-time underwater acoustic passive noise signal; A display is communicatively connected to the underwater acoustic passive target detection device and is used to display the underwater acoustic passive target detection results.

Images