CN-121996983-A - Sea area signal investigation method and device based on deep learning

CN121996983ACN 121996983 ACN121996983 ACN 121996983ACN-121996983-A

Abstract

The invention belongs to the technical field of underwater acoustic communication reconnaissance and artificial intelligence edge calculation, and discloses a sea area signal reconnaissance method and device based on deep learning. The method utilizes a constructed lightweight LeNet-5 hierarchical convolutional neural network model to be embedded and deployed into an STM32H743 microcontroller through an X-Cube-AI tool to process underwater sound signals, automatically identifies and classifies underwater signals of active sonar signals and non-cooperative communication signals in real time, and is used for analyzing recovered sea area information. The invention can automatically learn deep features from the time-frequency diagram, has better robustness to noise and complex marine environment, and has higher recognition accuracy.

Inventors

JI XIAONAN
LI LI
YIN WENJIE
Cao ran
ZHANG XIAO

Assignees

哈尔滨工程大学

Dates

Publication Date: 20260508
Application Date: 20260212

Claims (10)

1. A sea area signal investigation method based on deep learning is characterized in that the method utilizes a constructed lightweight LeNet-5 hierarchical convolutional neural network model to be embedded and deployed into an STM32H743 microcontroller through an X-Cube-AI tool to process underwater sound signals, automatically identifies and classifies underwater signals of active sonar signals and non-cooperative communication signals in real time, and is used for sea area information analysis after recovery, and the method specifically comprises the following steps: S101, judging by a LA sub-network, and carrying out primary judgment on an input signal based on the spectrum concentration degree and the feature significance of the signal, and dividing the input signal into a first type signal and a second type signal, wherein the first type signal comprises an active sonar signal and a synchronous signal which are more significant in feature and easy to identify; S102, LA sub-network processing is carried out to identify the first type of signals identified in the first step and carry out secondary identification through an LD sub-network, wherein the method comprises the steps of firstly, based on the characteristic of fixed modulation mode of active sonar signals, preferentially identifying the active sonar signals in the signals; and S103, carrying out LC sub-network processing, and carrying out modulation identification and classification on the second type signals distinguished in the first step, namely, the non-cooperative communication signals.
2. The deep learning-based sea area signal investigation method of claim 1, wherein constructing a lightweight LeNet-5 hierarchical convolutional neural network model comprises the steps of: (1) Generating five non-cooperative communication signals of CW, LFM, 2FSK, 4FSK and BPSK and HFM synchronous signals by using simulation software MATLAB, simulating the actually received signals; (2) Introducing ten active sonar frequency modulation signals in a currently known database; (3) Receiving the processed time-frequency diagram through an input layer, and taking the time-frequency diagram as the input of the light LeNet-5 hierarchical convolutional neural network; (4) Constructing TensorFlow/Keras library based on TensorFlow/Keras framework by using Python language, defining LA sub-network as two-class model, and defining LB, LC and LD sub-network as multiple classes; (5) Carrying out light-weight design on LA, LB, LC and LD subnetworks; (6) Converting the characteristic channels into scalar quantities through an SE attention module by adopting global average pooling to obtain the weight of each characteristic channel, processing the scalar quantities through two full-connection layers, and then outputting a weight vector with the same number as the channel number, namely the attention weight, as the input of a ReLU activation function, and adjusting the weight of the characteristic channels in the training process; (7) Carrying out average pooling on each feature map by adopting global average pooling, and extracting feature information describing global features; (8) Classifying the extracted characteristic information through an output layer Softmax function to obtain probability distribution of five types of modulation signals, and selecting the category with the highest probability as a recognition result through decision; (9) Training, aiming at a lightweight LeNet-5 hierarchical convolutional neural network model constructed by an upper computer, training parameters of the lightweight LeNet-5 hierarchical convolutional neural network model by adopting an unsupervised training stage and a fine tuning stage, acquiring conditional probability distribution of each layer, and adjusting different sub-network parameters to realize parameter self-adaptive adjustment of the lightweight LeNet-5 hierarchical convolutional neural network.
3. The deep learning-based sea area signal investigation method of claim 2, wherein obtaining the data sample set in step (1) specifically comprises: generating five non-cooperative communication signals of CW, LFM, 2FSK, 4FSK and BPSK by using simulation software MATLAB, and simulating an actually received signal; an active sonar signal data set which uses ten active sonar frequency modulation signals in a known database; The synchronous signal data set is synchronous signals of underwater acoustic communication commonly use HFM signals, so that simulation software MATLAB is used for generating a plurality of groups of HFM signals, and the actually received signals are simulated; And carrying out short-time Fourier transform (STFT) on the signals, generating a time-frequency diagram, carrying out time dimension translation by using SpecAugment data enhancement technology, and carrying out data expansion by using a mask of the time-frequency dimension, and expanding a data set to obtain a complete data set sample set.
4. The sea area signal investigation method based on deep learning of claim 2 is characterized in that in the step (4), a lightweight LeNet-5 hierarchical convolutional neural network model is adopted, and the method comprises the steps of adopting depth separable convolution for a convolutional layer, reducing the quantity and the computational complexity of model parameters, adopting a global average pooling module to replace flat flattening to prevent overfitting, adopting a ReLU function to replace Sigmoid function as an activation function, introducing nonlinear characteristics, and avoiding gradient disappearance; And (5) carrying out light weight design on the LA, LB, LC and LD subnetworks, wherein the light weight design comprises the steps of adopting depth separable convolution, independently calculating a plurality of characteristic images of a time-frequency image in an independent channel, carrying out calculation on the channel by using point convolution with the size of 1*1 to finish convolution calculation, reducing the upper and lower sizes of the characteristic images by two pixels respectively, carrying out depth separable convolution again after halving the size of the characteristic images through a second pooling layer, and halving the size of the characteristic images again through a fourth pooling layer.
5. The deep learning-based sea area signal investigation method of claim 1, wherein the built lightweight LeNet-5 hierarchical convolutional neural network model is embedded and deployed into the STM32H743 microcontroller by an X-Cube-AI tool, comprising the steps of: (a) An X-Cube-AI tool deployed in STM32CubeMX is adopted to guide the trained lightweight LeNet-5 hierarchical convolutional neural network model into an embedded microcontroller system, and model compatibility and resource demand verification are carried out on the lightweight LeNet-5 hierarchical convolutional neural network model; (b) Converting the lightweight LeNet-5 hierarchical convolutional neural network model into a pure C code through the code generation function of an X-Cube-AI tool; (c) The tool generates a core file network.c and a core file_data.c, and opens the two files in a Keil MDK development environment; (d) Opening the converted project through Keil MDK, compiling a modified code, checking code information through a serial port debugging assistant, and burning the project into an STM32H743 microcontroller connected to a PC end to complete embedded deployment of the lightweight LeNet-5 hierarchical convolutional neural network model; (e) Returning to CubeMX, click Validate on target, run a validation program on board, and if working, successfully deploy in the STM32H743 microcontroller.
6. A sea area signal detection device based on deep learning, which is characterized by implementing the sea area signal detection method based on deep learning as claimed in any one of claims 1-5, wherein the device comprises a core processor module, an AD signal acquisition module, an underwater filter circuit module, a serial port communication module, a lightweight sea area signal processing module, a power management module and an SD storage module; The AD signal acquisition module, the underwater filter circuit module, the serial port communication module, the lightweight sea area signal processing module, the power management module and the SD storage module are embedded into the core processor module, and the core processor module adopts a microcontroller taking STM32H743 as a core; The marine aquatic signal processing system comprises a hydrophone capsule, an AD signal acquisition module, an underwater filter circuit module, a lightweight marine signal processing module, an SD storage module, a serial port communication module, a power management module and a power management module, wherein the hydrophone capsule is arranged in the marine signal acquisition module, the AD signal acquisition module acquires marine signals and converts the signals, the underwater filter circuit module filters noise and then inputs the marine signals into the lightweight marine signal processing module to process and classify the marine signals, the result is finally input into the SD storage module, the serial port communication module is used for the interaction communication between an STM32H743 microcontroller and each module, and the power management module is used for normal power supply of each module during working.
7. The sea area signal investigation device based on deep learning of claim 6, wherein the core processor module is based on Arm Cortex-M7 kernel, built-in FPU, and supports light LeNet-5 hierarchical convolutional neural network model deployment through X-Cube-AI tool; The STM32H743 microprocessor collects underwater sound signals through an internal ADC collecting module, filters noise through an underwater filtering circuit module, processes the underwater sound signals through short-time Fourier transform to obtain a time-frequency diagram, and specifically comprises the following steps: (i) The ADC acquisition module performs signal sampling through an ADC of an STM32H743 microprocessor, the sampling rate is set to 96kHz, and a double-buffer technology is used to obtain real-time underwater sound signal input; (ii) Filtering low-frequency and high-frequency noise of an input signal by using an underwater filter circuit module to obtain an underwater sound signal with noise removed; (iii) Calculating energy judgment statistics of the signals by utilizing CMSIS-DSP library functions, performing modulo and square sum calculation on the received signals, and comparing the energy judgment statistics with a judgment threshold to obtain whether the signals are effective or not; (iv) Unifying the amplitude characteristics of the detected signals to be within a standard range through normalization, and windowing the signals; (v) Performing Fast Fourier Transform (FFT) by using a DSP library function to obtain a complex number array; (vi) Performing amplitude calculation on the complex columns by using the DSP library function to obtain amplitude information on each frequency, and outputting a time-frequency diagram of the signal; (vii) After the signal STFT calculation is completed, a one-dimensional array with the length of 512 is obtained, the data size scale is rearranged, and the array is converted into 321 multiplied by 70 with the same output size scale as the input size scale of the lightweight LeNet-5 hierarchical convolutional neural network model.
8. The deep learning-based sea area signal detection device according to claim 6, wherein the underwater filtering circuit module adopts an operational amplifier to construct a band-pass filter composed of a high-pass filter of a fourth-order butterworth design and a low-pass filter of an eighth-order butterworth design in cascade; The AD signal acquisition module specifically comprises an ADC acquisition module and a bias circuit module, wherein the ADC acquisition module adopts an STM32H743 self-contained ADC function to carry out analog-to-digital conversion on signals, adopts a double-buffer technology to write data into one buffer area when the data is read in the other buffer area, and realizes the real-time processing of the data; The SD storage module uses the SD card module to store data, adopts the TF card seat to fix the SD card module and provides electric connection to carry out parallel data transmission through a communication protocol of an SD mode; the power supply management module adopts an external battery to supply power, and the integrated voltage reduction module is used for voltage reduction output; The serial port communication module adopts a universal asynchronous transceiver as a communication protocol, adopts an RS232 serial port communication interface standard, realizes conversion between a logic level and an RS232 standard level through a MAX232 level conversion chip, and realizes serial port communication of a core processor module, an AD signal acquisition module, an underwater filter circuit module, a lightweight sea area signal processing module, a power management module, an SD storage module and a PC end.
9. The sea area signal investigation device based on deep learning according to claim 6, wherein the lightweight LeNet-5 hierarchical convolutional neural network model of the lightweight sea area signal processing module is embedded and deployed into an STM32H743 microcontroller through an X-Cube-AI tool to perform hierarchical processing on characteristics of the sea area underwater sound signals, and when the sea area underwater sound signals are input into the sea area, underwater signals including active sonar signals, synchronous signals and non-cooperative communication signals are automatically identified, classified and modulated in real time and used for analysis of the recovered sea area information.
10. The deep learning-based sea area signal detection device according to claim 6, wherein the SD memory module comprises an SD card module, a memory external expansion module and a memory management module; The SD card module is driven in a SDMMC mode by adopting a large SD card interface and is used for storing mass data, the memory external expansion module adopts a synchronous dynamic random access memory, and initializes and allocates addresses to the synchronous dynamic random access memory through an external memory controller and is used for storing data generated in the operation process of the system; the power management module specifically comprises an integrated voltage reduction module and a linear voltage stabilizer, wherein the integrated voltage reduction module reduces the higher input voltage of an external battery to an intermediate voltage level, and the linear voltage stabilizer reduces fluctuation and noise of the output voltage while further reducing the voltage.

Description

Sea area signal investigation method and device based on deep learning Technical Field The invention belongs to the technical field of underwater acoustic communication reconnaissance and artificial intelligence edge calculation, and particularly relates to a sea area signal reconnaissance method and device based on deep learning. Background With the continuous increase of ocean resource development and ocean safety requirements, the ocean area signal detection technology plays an increasingly important role in the fields of national defense safety, ocean resource exploration, ocean environment monitoring and the like. The sea area signal detection is mainly used for collecting, processing and identifying underwater sonar signals, communication signals and the like, and provides technical support for sea area monitoring and target identification. The conventional underwater signal recognition method mainly relies on a conventional algorithm based on feature extraction and pattern recognition, for example, features such as frequency spectrum, zero crossing rate, instantaneous frequency and the like of a signal are calculated, and the recognition is performed by combining a classifier such as a support vector machine or a decision tree. The method is effective in a specific and stable environment, but the identification performance of the method is seriously dependent on the completeness and accuracy of artificial design features, and for non-cooperative signals with low signal-to-noise ratio in complex and changeable marine environments, the generalization capability and the robustness of the method are often insufficient, and the identification accuracy is obviously reduced. In recent years, deep learning, particularly convolutional neural networks, has achieved great success in the fields of image and voice recognition, and the strong end-to-end characteristic learning capability provides a new idea for underwater sound signal recognition. There have been attempts to automatically classify the underwater acoustic signal using a deep learning model using a time-frequency plot (e.g., obtained by short-time fourier transform) as input. However, deploying large deep learning models onto embedded front-end devices (such as microcontrollers placed in buoys or submarines) where computational resources, storage resources, and power consumption are extremely limited presents a significant challenge. The existing general deep learning model is high in computational complexity and large in parameter quantity, and real-time reasoning is difficult to realize on embedded microcontrollers such as STM 32. If the data is transmitted back to the cloud for processing, the real-time requirement cannot be met, and the power consumption and the communication complexity of the system are increased. Therefore, a method and a device capable of accurately completing automatic identification of underwater acoustic signals in sea areas in real time for a long time are needed, and the method should be capable of adopting a layered processing strategy for different types of signals, improving the computing efficiency, and being capable of efficiently running on embedded equipment with limited resources, and providing instant information support for underwater non-cooperative target reconnaissance. Disclosure of Invention In order to overcome the problems in the related art, the embodiment of the invention discloses a sea area signal detection method and device based on deep learning. The invention provides a sea area signal detection method and device based on deep learning, which aims to solve the technical problems that an existing underwater detection system lacks a self-contained hydrophone capable of automatically identifying sea area underwater acoustic signals in a long-term, real-time and accurate manner under water, can not provide instant information support for underwater non-cooperative target detection, is low in traditional neural network calculation efficiency, is insufficient in robustness in noise and complex marine environments and the like in a traditional artificial feature-based method. The technical scheme is that the sea area signal investigation method based on deep learning utilizes a constructed lightweight LeNet-5 hierarchical convolutional neural network model to be embedded and deployed into an STM32H743 microcontroller through an X-Cube-AI tool to process underwater sound signals, automatically identifies and classifies underwater signals of active sonar signals and non-cooperative communication signals in real time, and is used for analyzing recovered sea area information, and the method specifically comprises the following steps: S101, judging by a LA sub-network, and carrying out primary judgment on an input signal based on the spectrum concentration degree and the feature significance of the signal, and dividing the input signal into a first type signal and a second type signal, wherein the first type