MEMS underwater acoustic transducer array and underwater frogman target detection method

Abstract

The present application provides a MEMS hydroacoustic transducer array, the MEMS hydroacoustic transducer array includes MEMS hydroacoustic transducers arranged in an array; the MEMS hydroacoustic transducer includes a signal generator, a power amplifier, a hydroacoustic transducer and a signal receiver; the signal generator is used to generate an electrical signal; the power amplifier is used to amplify the power of the electrical signal; the hydroacoustic transducer is used to convert the amplified electrical signal into original sound pressure data, and transmit the original sound pressure data to the area to be measured; the signal receiver is used to receive the echo signal generated by the original sound pressure data hitting the target object, and obtain the echo sound pressure data. The present application also provides an underwater frogman target detection method, which is realized by the MEMS hydroacoustic transducer array.

Description

Technical Field

[0001] The present application relates to the technical field of underwater target detection and identification, and in particular to a MEMS underwater acoustic transducer array and an underwater frogman target detection method. Background Art

[0002] With the application of the "asymmetric warfare" mode, underwater frogmen have gradually attracted attention due to their small size and low noise intensity. Therefore, both hardware and software algorithms need to meet the detection and identification of small targets such as underwater frogmen. Since electromagnetic wave signals propagate poorly underwater, underwater acoustic detection is currently the only reliable means of monitoring and discovering underwater frogmen, and sonar is usually used as the core equipment for underwater detection. Traditional sonar equipment includes passive sonar and active sonar. Although passive underwater acoustic monitoring has low energy consumption, strong concealment and high robustness, it relies on the sound waves emitted by the target object itself for target identification. For small targets with low noise intensity such as frogmen, the detection and identification effect of passive sonar is not ideal. Due to the fact that underwater frogmen often sneak into complex waters such as docks, ports, ships and offshore platforms, the optimal center frequency, detection distance, detection angle and coverage of active sonar will be limited by the complex underwater acoustic environment. At the same time, due to the small size of frogmen, when only a single active sonar device is used, the frogmen's echo signal is mixed with the background noise signal and is difficult to identify. In terms of software algorithms, there are currently matched filters, line spectrum enhancers or improved methods based on them to process ocean sonar data signals, but these algorithms are easily affected by background noise when processing small targets such as frogmen. If the resolution accuracy is forcibly improved, the false alarm rate will increase. Summary of the invention

[0003] In view of this, the present invention combines MEMS sonar transducers with neural networks to establish a MEMS sonar transducer array and underwater frogman target detection method that can be widely used in complex underwater scenes, so as to achieve accurate detection and identification of underwater frogman targets.

[0004] In a first aspect, the present invention provides a MEMS underwater acoustic transducer array, wherein the MEMS underwater acoustic transducer array comprises MEMS underwater acoustic transducers arranged in an array form;

[0005] The MEMS underwater acoustic transducer includes a signal generator, a power amplifier, an underwater acoustic transducer and a signal receiver;

[0006] The signal generator is used to generate an electrical signal; the power amplifier is used to amplify the power of the electrical signal; the underwater acoustic transducer is used to convert the amplified electrical signal into original sound pressure data, and transmit the original sound pressure data to the area to be tested; the signal receiver is used to receive the echo signal generated by the original sound pressure data hitting the target object, and obtain the echo sound pressure data.

[0007] In the MEMS underwater acoustic transducer array described in the present invention,

[0008] Each of the MEMS hydroacoustic transducer arrays includes m rows, each row has n MEMS hydroacoustic transducers, the MEMS hydroacoustic transducers are arranged in an array, and a spacing between two adjacent rows of MEMS hydroacoustic transducers and a spacing between two adjacent MEMS hydroacoustic transducers in each row are set.

[0009] In a second aspect, the present invention further provides an underwater frogman target detection method, which is implemented by the MEMS hydroacoustic transducer array described in any one of the above items, and the underwater frogman target detection method comprises the following steps:

[0010] S11, obtaining original echo sound pressure data collected by the MEMS underwater acoustic transducer array;

[0011] S12, processing the original echo sound pressure data to obtain first processed data;

[0012] S13. Put the first processed data into a pre-trained deep learning network model for identification, identify the first processed data through three types of discriminant feature data of the frogman target's bubble group, gas cylinder and frogman's body, and detect and identify the underwater frogman target based on the identification result.

[0013] In the underwater frogman target detection method of the present invention,

[0014] The step S12 comprises:

[0015] Performing analog-to-digital conversion on the original echo sound pressure data to obtain a discrete digital signal;

[0016] The discrete digital signals are sorted, and the sorted digital signals are processed using a wavelet transform algorithm to obtain first processed data.

[0017] In the underwater frogman target detection method of the present invention,

[0018] The step S13 comprises:

[0019] S131, putting the first processed data into a pre-trained deep learning network model, wherein the deep learning network model includes a backbone network and a head network; the head network is three discriminant features of a bubble group, a gas cylinder, and a frogman's body, respectively, for identifying a frogman target;

[0020] S132, the backbone network extracts features of the first processed data initially input into the deep learning network model, and inputs the extracted feature data into the head network respectively;

[0021] S133, identifying three types of discriminative feature data of the bubble group, the gas cylinder, and the body of the frogman target through the head network;

[0022] S134, judging the recognition result, if the three discrimination features of the frogman target, namely, the bubble group, the gas cylinder, and the frogman's body, are all judged to be true, then the recognition object is the target frogman.

[0023] In the underwater frogman target detection method of the present invention,

[0024] The pre-trained deep learning network model in step S13 is constructed in the following manner:

[0025] Construct a backbone network and a head network; the backbone network includes a convolution layer and a pooling layer, there are s convolution layers and pooling layers in total, p convolution layers and q pooling layers appear alternately, the convolution kernel size of the convolution layer is k×k, and the pooling layer adopts the form of a maximum pooling layer;

[0026] The head network is a fully connected layer, which first converts the feature data extracted by the backbone network into a one-dimensional series, and then puts the one-dimensional series into the fully connected layer for target recognition.

[0027] In the underwater frogman target detection method of the present invention,

[0028] The backbone network in the deep learning network model is iteratively learned in the following way:

[0029] S41, generating a convolutional neural network backbone network, wherein the input layer, hidden layer, output layer, and activation function used in each layer of the backbone network are consistent with the convolutional neural network backbone model;

[0030] S42, use the backbone network as the generative network of the adversarial neural network, and then construct a discriminative network, and combine them as a domain adversarial neural network;

[0031] S43, performing iterative learning on the generator network and the discriminant network respectively with the discriminant network accuracy and the difference between 1 and the discriminant network accuracy as the loss function;

[0032] S44, after iterative learning reaches a preset number of times, the backbone network and parameters are returned.

[0033] In the underwater frogman target detection method of the present invention,

[0034] Training the deep learning network model in step S13 includes the following steps:

[0035] S51, generating a training sample set;

[0036] S52, preprocessing the data in the training sample set;

[0037] S53, using the preprocessed data, training the backbone network based on the adversarial neural network and the transfer learning network model;

[0038] S54. Train the head network and output the deep learning network model.

[0039] In the underwater frogman target detection method of the present invention,

[0040] The step S53 comprises:

[0041] S531, extracting data features through a feature extraction network, and adding domain labels to the data features according to data sources;

[0042] S532, determine whether the maximum number of iterations has been reached, if yes, jump to step S54, otherwise jump to step S533;

[0043] S533, discriminating the domain label of the data through a domain adversarial neural network and calculating the judgment error loss;

[0044] S534, adjust the parameters of the feature extraction network and the domain adversarial network according to the error loss back propagation, and jump to step S531.

[0045] In the underwater frogman target detection method of the present invention,

[0046] The step S54 comprises:

[0047] S541, adjusting the head network parameters according to the error loss;

[0048] S542, inputting the training data samples with extracted features into the head network for training;

[0049] S543, determine whether the maximum number of iterations has been reached, if so, jump to step S544, otherwise jump to step S541;

[0050] S544. Output the deep learning network model.

[0051] Beneficial effects of the present invention:

[0052] The MEMS hydroacoustic transducer array and underwater frogman target detection method of the present application have two advantages over the prior art: establishing a layout method based on MEMS hydroacoustic transducers to realize active detection of multiple underwater scenes and high-precision identification of small targets such as underwater frogmen. MEMS hydroacoustic transducers have the functions of both hydrophones and transducers, and have good adaptability to full-band signals; the underwater frogman target detection method can effectively learn the corresponding features of underwater frogman targets from existing samples through a neural network algorithm, avoiding the work of manually extracting the weak signal features of frogmen from the overall signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 This is a schematic diagram of the MEMS underwater acoustic transducer array structure. Icon: 1-MEMS underwater acoustic transducer;

[0055] Figure 2 A flow chart of an underwater frogman target detection method provided by an embodiment of the present invention;

[0056] Figure 3 is a sub-flow chart of step S13 in an embodiment of the present invention;

[0057] Figure 4 A model and implementation flow chart based on adversarial neural network and transfer learning according to an embodiment of the present invention;

[0058] Figure 5 This is a training flowchart of a deep learning model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this specification. Instead, they are merely examples of devices and methods consistent with some aspects of this specification as detailed in the appended claims.

[0060] The terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit this specification. The singular forms "a", "the" and "the" used in this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0061] It should be understood that although the terms first, second, third, etc. may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this specification, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0063] Example 1

[0064] like Figure 1 As shown, in a first aspect, an embodiment of the present invention provides a MEMS underwater acoustic transducer array, wherein the MEMS underwater acoustic transducer array includes MEMS underwater acoustic transducers 1 arranged in an array form.

[0065] Optionally, each of the MEMS hydroacoustic transducer arrays includes m rows, each row has n MEMS hydroacoustic transducers 1, the MEMS hydroacoustic transducers 1 are arranged in an array form, and a spacing a between two adjacent rows of MEMS hydroacoustic transducers 1 and a spacing b between two adjacent MEMS hydroacoustic transducers 1 in each row are set.

[0066] The MEMS underwater acoustic transducer 1 includes a signal generator, a power amplifier, an underwater acoustic transducer and a signal receiver.

[0067] The signal generator is used to generate an electrical signal; the power amplifier is used to amplify the power of the electrical signal; the underwater acoustic transducer is used to convert the amplified electrical signal into original sound pressure data, and transmit the original sound pressure data to the area to be tested; the signal receiver is used to receive the echo signal generated by the original sound pressure data hitting the target object, and obtain the echo sound pressure data.

[0068] The MEMS hydroacoustic transducers 1 in the MEMS hydroacoustic transducer array are numbered in sequence, and the original sound pressure data collected by the MEMS hydroacoustic transducer 1 are numbered according to the number of the MEMS hydroacoustic transducer 1 as input samples of the deep learning network model.

[0069] In the embodiment of the present invention, the MEMS underwater acoustic transducer array integrates the functions of a signal generator, a power amplifier, an underwater acoustic transducer (hydrophone) and a signal receiver, that is, it has the functions of both sending and receiving signals.

[0070] Second, as Figure 2 As shown, an embodiment of the present invention further provides an underwater frogman target detection method, which is implemented by the MEMS hydroacoustic transducer array described in any one of the above items, and the underwater frogman target detection method includes the following steps:

[0071] S11. Obtaining original echo sound pressure data collected by the MEMS underwater acoustic transducer array.

[0072] The MEMS underwater acoustic transducer array first generates an electrical signal and amplifies the power of the electrical signal, then converts the electrical signal into an original sound wave signal and transmits it to the area to be tested; the echo signal generated by the original sound wave signal hitting the target object is collected by the MEMS underwater acoustic transducer 1 to obtain the original echo sound pressure data.

[0073] Preferably, the generated electrical signal includes but is not limited to a triangular wave, a sawtooth wave, and a rectangular wave; preferably, the initial electrical signal can be set to a fixed frequency mode or a sweep frequency mode. According to the body size of the frogman, the lower limit of the frequency can be calculated to be no less than 1kHz and the upper limit is no more than 10kHz. If the frequency is too high, it will decay quickly in the water and it will be difficult to detect distant targets; if the frequency is too low, the sound wave wavelength will be too long and will bypass the target, causing misjudgment.

[0074] S12: Process the original echo sound pressure data to obtain first processed data.

[0075] In a specific implementation, the step S12 includes:

[0076] The original echo sound pressure data is subjected to analog-to-digital conversion processing to obtain discrete digital signals, and the digital signals are numbered and sorted according to the number of the MEMS underwater acoustic transducer 1 that acquired the data.

[0077] The discrete digital signals are sorted, and the sorted digital signals are processed using a wavelet transform algorithm to obtain first processed data.

[0078] In a more specific implementation, the sorted digital signal is processed using a wavelet transform algorithm, where wavelet transform is performed on the data acquired by each MEMS underwater acoustic transducer 1. To facilitate subsequent processing, the data obtained after the wavelet transform is normalized to obtain first processed data.

[0079] Optionally, the wavelet transform algorithm may be any one of Haar wavelet transform, Daubechies wavelet transform, Biorthogonal wavelet transform, Coiflets wavelet transform, Symlets wavelet transform, or other algorithms.

[0080] S13. Put the first processed data into a pre-trained deep learning network model for identification, identify the first processed data through three types of discriminant feature data of the frogman target's bubble group, gas cylinder and frogman's body, and detect and identify the underwater frogman target based on the identification result.

[0081] In order to increase the early warning capability after identification, when the identification results of the three types of target discrimination are all true, it can be judged that a frogman is approaching and an early warning can be issued to the ground personnel; otherwise, no early warning will be issued to the ground personnel.

[0082] like Figure 3 As shown, in a specific implementation, step S13 includes:

[0083] S131. Put the first processed data into a pre-trained deep learning network model, wherein the deep learning network model includes a backbone network and a head network; the head network is composed of three discriminant features of a bubble group, a gas cylinder, and a frogman's body, which are respectively used to identify the frogman target.

[0084] S132. The backbone network extracts features of the first processed data initially input into the deep learning network model, and inputs the extracted feature data into the head network respectively.

[0085] S133, identifying three types of discriminative feature data of the bubble group, the gas cylinder, and the frogman's body of the frogman target through the head network.

[0086] The feature data of the bubble group is set to be recognized, mainly for the characteristics of a large number of bubbles discharged by open breathing.

[0087] S134, judging the recognition result, if the three discrimination features of the frogman target, namely, the bubble group, the gas cylinder, and the frogman's body, are all judged to be true, then the recognition object is the target frogman.

[0088] Optionally, in the underwater frogman target detection method described in the embodiment of the present invention, step S13 also includes: when the identified object is a target frogman, issuing a warning to ground personnel.

[0089] Optionally, in the underwater frogman target detection method described in the embodiment of the present invention,

[0090] The pre-trained deep learning network model in step S13 is constructed in the following manner:

[0091] Construct a backbone network and a head network; the backbone network includes a convolution layer and a pooling layer, there are s convolution layers and pooling layers in total, p convolution layers and q pooling layers appear alternately, the convolution kernel size of the convolution layer is k×k, and the pooling layer adopts the form of a maximum pooling layer.

[0092] The head network is a fully connected layer, which first converts the feature data extracted by the backbone network into a one-dimensional series, and then puts the one-dimensional series into the fully connected layer for target recognition.

[0093] Alternatively, if Figure 4 As shown, in the underwater frogman target detection method described in the embodiment of the present invention,

[0094] The backbone network in the deep learning network model is iteratively learned in the following way:

[0095] S41, generating a convolutional neural network backbone network, wherein the input layer, hidden layer, output layer, and activation function used in each layer of the backbone network are consistent with the convolutional neural network backbone model;

[0096] S42, use the backbone network as the generative network of the adversarial neural network, and then construct a discriminative network, and combine them as a domain adversarial neural network;

[0097] S43, performing iterative learning on the generator network and the discriminant network respectively with the discriminant network accuracy and the difference between 1 and the discriminant network accuracy as the loss function;

[0098] S44, after iterative learning reaches a preset number of times, the backbone network and parameters are returned.

[0099] Optionally, in the underwater frogman target detection method described in the embodiment of the present invention,

[0100] like Figure 5 As shown, the training of the deep learning network model in step S13 includes the following steps:

[0101] S51: Generate a training sample set.

[0102] Optionally, it is planned to select two typical underwater scenes of representative ports and ships suitable for shallow water and deep sea environments. The port underwater scenes include shore bases, shelters, water surfaces, and seabed; the ship underwater scenes include the underwater parts of ship platforms and water surfaces. On the basis of building a typical underwater scene, the present invention intends to model the underwater frogman target in three parts, namely, a bubble group, a gas cylinder and a frogman's body. By controlling an external air compressor connected to a hose, a bubble group with the same flow, period and duration as the frogman's exhalation is generated in an underwater anechoic environment. An open breathing gas cylinder and soft materials similar to the human body structure are selected, and the bubble group, gas cylinder and human body model are tested for echo signals through a MEMS hydroacoustic transducer array. The acoustic signals collected by the array are used as experimental samples for the input of subsequent deep learning of acoustic recognition of underwater frogman targets, and the label "0" is added to these samples. The experiment obtains the same model and the same situation, and the MEMS hydroacoustic transducer array is used to test the bubble group, gas cylinder and human body model for echo signals. The acoustic signals collected by the array are used as real samples for the input of subsequent deep learning of acoustic recognition of underwater frogman targets, and the label "1" is added to these samples.

[0103] S52, preprocessing the data in the training sample set; the objects of preprocessing include preprocessing the experimental sample data and the real sample data.

[0104] S53. Using the preprocessed data, the backbone network is trained based on the adversarial neural network and transfer learning network model.

[0105] Optionally, the step S53 includes:

[0106] S531, extracting data features through a feature extraction network, and adding domain labels to the data features according to data sources;

[0107] S532, determine whether the maximum number of iterations has been reached, if yes, jump to step S54, otherwise jump to step S533;

[0108] S533, discriminating the domain label of the data through a domain adversarial neural network and calculating the judgment error loss;

[0109] S534, adjust the parameters of the feature extraction network and the domain adversarial network according to the error loss back propagation, and jump to step S531.

[0110] Optionally, in a more specific implementation, step S53 includes: building a deep learning network model, selecting Adam as the optimizer, and setting the learning rate to η1. 80% of the collected target echo signal samples are used as the training set, and the remaining 20% ​​are used as the validation set. The number of training times is set to Nmax1 times, and the data is put into the domain adversarial network for learning, so that the convolutional neural network can extract features that satisfy both the training data and the actual data. The loss function of the domain adversarial neural network is:

[0111]

[0112] Here, the discriminant network D is trained to accurately identify the labels of the training samples with the greatest probability, that is, to maximize logD(x) and log(1-D(G(x))), and the generative network G is trained to minimize log(1-D(G(x))), that is, to maximize the loss of D.

[0113] S54. Train the head network and output the deep learning network model.

[0114] Optionally, the step S54 may include:

[0115] S541, adjusting the head network parameters according to the error loss;

[0116] S542, inputting the training data samples with extracted features into the head network for training;

[0117] S543, determine whether the maximum number of iterations has been reached, if so, jump to step S544, otherwise jump to step S541;

[0118] S544. Output the deep learning network model.

[0119] Optionally, in a more specific implementation, step S54 may include:

[0120] The trained backbone network and head network are combined into a new network, namely the convolutional neural network in S32. The parameters in the backbone network are fixed and only the parameters in the head network are changed. The head network is trained for different recognition targets. The optimizer is Adam and the learning rate is set to η2. 80% of the collected target echo signal samples are used as the training set and the remaining 20% ​​as the validation set. The number of training times is set to Nmax2 times. The head network loss function is set to the cross entropy loss function, and the expression is as follows:

[0121] Loss=-ylog(f(x))-(1-y)log(1-f(x))

[0122] Where f(x) is the predicted value of the neural network and y is the sample label. The head network parameters are obtained by training in this way.

[0123] The beneficial effects of the embodiments of the present invention are as follows:

[0124] The MEMS hydroacoustic transducer array and underwater frogman target detection method of the present application have two advantages over the prior art: the arrangement method based on the MEMS hydroacoustic transducer 1 is established to realize active detection of multiple underwater scenes and high-precision identification of small targets such as underwater frogmen. The MEMS hydroacoustic transducer 1 has the functions of both a hydrophone and a transducer, and has good adaptability to full-band signals; the underwater frogman target detection method can effectively learn the corresponding features of underwater frogman targets from existing samples through a neural network algorithm, avoiding the work of manually extracting the weak signal features of frogmen from the overall signal.

[0125] The concept and principle of the embodiments of the present invention are further described below:

[0126] When small targets such as frogmen are detected by active sonar, it is difficult for traditional sonar to accurately identify them in the complex underwater acoustic environment. The MEMS underwater acoustic transducer 1 used in the present invention has the advantages of small size, low power consumption and single-unit orientation, and can achieve high detection accuracy and efficiency for small targets such as frogmen. The present invention is based on the wavelet transform algorithm to represent, process and analyze the data after analog-to-digital conversion, and can effectively process the underwater frogman target echo signal that is a mixture of oscillating signals and transient non-oscillating signals with oscillation, non-stationarity and nonlinearity. Processing data with a wavelet transform algorithm can increase the signal-to-noise ratio of the signal, reduce the difficulty of subsequent data processing and target identification, and improve the credibility of the results. The adversarial neural network and transfer learning model used in the training of the convolutional neural network of the present invention can solve the problems of difficulty in converging neural network model parameters due to insufficient real data and difficulty in reflecting the real ocean reverberation environment by simulating the real environment in the simulation software. The convolutional neural network model adopted by the present invention establishes training and testing sets for different targets and their combinations by designing its unique hidden layer, and selects underwater acoustic signal samples that are not predicted by the training set for evaluation on the validation set, further optimizing and forming a design method that can accurately detect and identify underwater frogmen, avoiding interference of underwater low-frequency noise and underwater strong reverberation environment on target detection. The present invention identifies three types of features, namely, a large number of bubbles discharged by open breathing, breathing cylinders and frogmen's bodies. When the detection target reaches the warning value, an alarm will be issued to ground personnel, effectively dealing with the problem of high false alarm rate of traditional sonar in complex marine environments with strong reverberation background.

[0127] The algorithm and display provided herein are not inherently related to any particular computer, virtual system or other device. Various general purpose systems can also be used together with the teachings based on this. According to the above description, it is obvious that the structure required for constructing such systems. In addition, the present invention is not directed to any specific programming language either. It should be understood that various programming languages ​​can be utilized to realize the content of the present invention described herein, and the description of the above specific languages ​​is for disclosing the best mode of the present invention.

[0128] In addition, those skilled in the art will appreciate that, although some embodiments herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present invention and form different embodiments. For example, in the claims below, any one of the claimed embodiments may be used in any combination.

[0129] The various component embodiments of the present invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or a digital signal processor (DSP) may be used in practice to implement some or all functions of some or all components of the thermal simulation device for the aluminum substrate and the electronic device according to the embodiment of the present invention. The present invention may also be implemented as a device or device program (e.g., a computer program and a computer program product) for executing part or all of the methods described herein. Such a program for implementing the present invention may be stored on a computer-readable medium, or may be in the form of one or more signals. Such a signal may be downloaded from the Internet, or provided on a carrier signal, or provided in any other form.

[0130] The above is only an embodiment of the present application. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present application, which should also be regarded as the protection scope of the present application, and these will not affect the effect of the implementation of the present application and the practicality of the patent. The protection scope required by the present application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for underwater frogman target detection, implemented by a MEMS hydroacoustic transducer array, wherein the MEMS hydroacoustic transducer array comprises MEMS hydroacoustic transducers arranged in an array; The MEMS underwater acoustic transducer includes a signal generator, a power amplifier, an underwater acoustic transducer and a signal receiver; The signal generator is used to generate an electrical signal; the power amplifier is used to amplify the power of the electrical signal; the underwater acoustic transducer is used to convert the amplified electrical signal into original sound pressure data, and transmit the original sound pressure data to the area to be tested; the signal receiver is used to receive the echo signal generated by the original sound pressure data hitting the target object, and obtain the echo sound pressure data; it is characterized in that The underwater frogman target detection method comprises the following steps: S11, obtaining original echo sound pressure data collected by the MEMS underwater acoustic transducer array; S12, processing the original echo sound pressure data to obtain first processed data; S13. Put the first processed data into a pre-trained deep learning network model for identification, identify the first processed data through three types of discriminant feature data of the frogman target's bubble group, gas cylinder and frogman's body, and detect and identify the underwater frogman target based on the identification result.

2. The underwater frogman target detection method according to claim 1, characterized in that: Each of the MEMS hydroacoustic transducer arrays includes m rows, each row has n MEMS hydroacoustic transducers, the MEMS hydroacoustic transducers are arranged in an array, and a spacing between two adjacent rows of MEMS hydroacoustic transducers and a spacing between two adjacent MEMS hydroacoustic transducers in each row are set.

3. The underwater frogman target detection method according to claim 1, characterized in that: The step S12 comprises: Performing analog-to-digital conversion on the original echo sound pressure data to obtain a discrete digital signal; The discrete digital signals are sorted, and the sorted digital signals are processed using a wavelet transform algorithm to obtain first processed data.

4. The underwater frogman target detection method according to claim 3, characterized in that: The step S13 comprises: S131, putting the first processed data into a pre-trained deep learning network model, wherein the deep learning network model includes a backbone network and a head network; the head network is three discriminant features of a bubble group, a gas cylinder, and a frogman's body, respectively, for identifying a frogman target; S132, the backbone network extracts features of the first processed data initially input into the deep learning network model, and inputs the extracted feature data into the head network respectively; S133, identifying three types of discriminative feature data of the bubble group, the gas cylinder, and the body of the frogman target through the head network; S134, judging the recognition result, if the three discrimination features of the frogman target, namely, the bubble group, the gas cylinder, and the frogman's body, are all judged to be true, then the recognition object is the target frogman.

5. The underwater frogman target detection method according to claim 4, characterized in that: The pre-trained deep learning network model in step S13 is constructed in the following manner: Construct a backbone network and a head network; the backbone network includes a convolution layer and a pooling layer, there are s convolution layers and pooling layers in total, p convolution layers and q pooling layers appear alternately, the convolution kernel size of the convolution layer is k×k, and the pooling layer adopts the form of a maximum pooling layer; The head network is a fully connected layer, which first converts the feature data extracted by the backbone network into a one-dimensional series, and then puts the one-dimensional series into the fully connected layer for target recognition.

6. The underwater frogman target detection method according to claim 5, characterized in that: The backbone network in the deep learning network model is iteratively learned in the following way: S41, generating a convolutional neural network backbone network, wherein the input layer, hidden layer, output layer, and activation function used in each layer of the backbone network are consistent with the convolutional neural network backbone model; S42, use the backbone network as the generative network of the adversarial neural network, and then construct a discriminative network, and combine them as a domain adversarial neural network; S43, performing iterative learning on the generator network and the discriminant network respectively with the discriminant network accuracy and the difference between 1 and the discriminant network accuracy as the loss function; S44, after iterative learning reaches a preset number of times, the backbone network and parameters are returned.

7. The underwater frogman target detection method according to claim 6, characterized in that: Training the deep learning network model in step S13 includes the following steps: S51, generating a training sample set; S52, preprocessing the data in the training sample set; S53, using the preprocessed data, training the backbone network based on the adversarial neural network and the transfer learning network model; S54. Train the head network and output the deep learning network model.

8. The underwater frogman target detection method according to claim 7, characterized in that: The step S53 comprises: S531, extracting data features through a feature extraction network, and adding domain labels to the data features according to data sources; S532, determine whether the maximum number of iterations has been reached, if yes, jump to step S54, otherwise jump to step S533; S533, discriminating the domain label of the data through a domain adversarial neural network and calculating the judgment error loss; S534, adjust the parameters of the feature extraction network and the domain adversarial network according to the error loss back propagation, and jump to step S531.

9. The underwater frogman target detection method according to claim 8, characterized in that: The step S54 comprises: S541, adjusting the head network parameters according to the error loss; S542, inputting the training data samples with extracted features into the head network for training; S543, determine whether the maximum number of iterations has been reached, if so, jump to step S544, otherwise jump to step S541; S544. Output the deep learning network model.

Images