阅读说明：本技术 一种基于cnn的多水下信标信号识别方法 (Multi-underwater beacon signal identification method based on CNN ) 是由 赵冬冬 毛威波 陈朋 蔡天诚 梁世慧 于 2021-06-16 设计创作，主要内容包括：一种基于CNN的多水下信标信号识别方法,每个信标发射信号的设计,运用扩频编码技术并结合相频同移混合键控方式调制,使不同的信标信号具有不同的相位特征和频率特征；将采样的信标信号序列转换为二维图片,通过短时傅里叶变换得到相位图和频谱图进而融合而成相频特征图,以此作为信号识别的依据；轻量化的卷积神经网络结构的搭建,共由1个输入层、4个卷积层、3个池化层、1个全连接层和1个softmax层组成,完成从相频特征图到信标序号的映射；不仅在相频特征图上进行数据扩充,而且在采样信号序列上进行数据扩充,使训练的模型具有较强鲁棒性。该方法可有效提升信标的作用距离和识别精度,轻量化的模型设计也适用于信号的实时识别。(A CNN-based multi-underwater beacon signal identification method is characterized in that each beacon emission signal is designed, and different beacon signals have different phase characteristics and frequency characteristics by applying a spread spectrum coding technology and combining a phase-frequency homoshift hybrid keying mode for modulation; converting the sampled beacon signal sequence into a two-dimensional picture, obtaining a phase diagram and a spectrogram through short-time Fourier transform, and further fusing the phase diagram and the spectrogram into a phase-frequency characteristic diagram which is used as a basis for signal identification; the method comprises the following steps of (1) building a lightweight convolutional neural network structure, wherein the lightweight convolutional neural network structure consists of 1 input layer, 4 convolutional layers, 3 pooling layers, 1 full-connection layer and 1 softmax layer, and mapping from a phase-frequency characteristic diagram to a beacon sequence number is completed; the method not only carries out data expansion on the phase-frequency characteristic diagram, but also carries out data expansion on the sampling signal sequence, so that the trained model has stronger robustness. The method can effectively improve the acting distance and the recognition precision of the beacon, and the lightweight model design is also suitable for real-time recognition of the signal.)

1. A CNN-based multi-underwater beacon signal identification method, comprising the steps of:

1) modulating beacon emission signals, adopting n-order spread spectrum coding technology to make each underwater beacon have only one corresponding coding sequence, and making phase-frequency homoshift keying mixed modulation on said coding sequence, different code elements have different phases and frequencies so as to form beacon emission signals, and every beacon is formed from 2ⁿ1 symbol component of 1 or-1, respectively 2^n-11 code element and 2^n-11-1 symbols, where 1 denotes m₁F with phase of 0 DEG₁Filling Hz sine waves; -1 represents by m₂F with 180 DEG phase₂Hz sine wave filling, and the modulated beacon analog signal is as follows:

when in useThen, the formula is simplified as follows (2):

where Y represents the beacon-modulated analog signal, T represents time, X [ i ]]Is shown in (2)ⁿ-1 [ i ] th of symbols]Sequence values, 1 or-1;

2) collecting various beacon signals;

3) performing data enhancement on the sampled beacon signal sequence set;

4) converting one-dimensional signal information into two-dimensional picture information through short-time Fourier transform, wherein the process is as follows: taking a beacon signal period t₁The sampling sequence of ms time is the conversion target of one frame, and t is₂ms timeThe length of the signal sample sequence between is the window length, and t₃The length of the signal sampling sequence of ms time is the translation step length to perform frame processing on the conversion target of one frame, and the number M of split sub-frames_fAs shown in formula (3):

wherein | represents integer division, |. represents non-integer division;

then, performing discrete fourier transform on each subframe sequence x (n) as shown in formula (4):

wherein X (k) is the sequence x (n) after DFT, f_sRepresenting the sampling rate (KHz), t₂Represents time (ms), f_st₂Number of sampling points representing the length of a window, the frequency resolution being 1/t₂；

Thereby obtaining 0-f according to the formula (5)_sThe component amplitude of the frequency, the offset phase with respect to the cosine according to equation (6);

P(k)＝real(X(k))²+imag(X(k))² (5)

calculating the component amplitude and offset phase of a subframe and plotting t₃Time-frequency diagram and time phase diagram of ms time, M_fThe time-frequency diagram and the phase diagram of each sub-frame are respectively spliced into a time length t₃*M_fA time-frequency diagram and a phase diagram of ms (one frame), and finally, two characteristic diagrams generated by the signal sequence of the frame are fused into a phase-frequency characteristic diagram with pixels of 100 × 3;

5) performing data enhancement on the picture data set again;

6) and training the data set by using a lightweight CNN network, and realizing real-time identification of various beacon signals by using an obtained network model.

2. The CNN-based multi-underwater beacon signal recognition method according to claim 1, wherein in the step (3), the sampled signal sequence subjected to data set expansion includes an actual sampled signal sequence and a theoretical sampled signal sequence, and the actual sampled signal sequence is subjected to amplitude size conversion; the following processing is performed for the theoretical sampling signal sequence: setting different amplitudes for the beacon signal itself; adding white noise with different amplitudes; and randomly carrying out distortion processing on the partial signals.

3. The CNN-based multi-underwater beacon signal identification method according to claim 1 or 2, wherein in the step (4), the window length of the short-time fourier transform is designed as an overlapping sliding window, and information of a longer time is mapped in a shorter time interval in the form of frequency component intensity, and a longer time (t) is used₂ms) as a window length; while using a shorter time (t)₃ms) as the step length of the sliding window, so that the information difference between adjacent sub-frames is only reflected in t at the head of the last sub-frame₃ms interval and t at the end of the next subframe₃ms interval.

4. The CNN-based multi-underwater beacon signal identification method according to claim 1 or 2, wherein in the step (5), the data set of the phase-frequency characteristic diagram is enhanced in the following three ways: firstly, setting different brightness, saturation and contrast for a phase frequency diagram by taking a beacon signal frequency band as a reference; secondly, performing Gaussian filtering processing and fuzzy processing of different degrees on the picture; thirdly, the two treatments are carried out simultaneously.

5. The CNN-based multi-underwater beacon signal identification method according to claim 1 or 2, wherein in the step (6), the convolutional neural network is used for mapping the frequency signature to the beacon sequence number, and the lightweight structure thereof is composed of many nonlinear structures, which are composed of 1 input layer, 4 convolutional layers, 3 pooling layers, 1 fully-connected layer and 1 softmax layer, wherein the first three convolutional layers have pooling layers.

Technical Field

The invention relates to the technical field of underwater acoustic navigation and positioning, in particular to a CNN-based multi-beacon signal identification method.

Background

With the progress of times, people have stronger development demand on ocean resources, and self-service underwater vehicles are produced. The method plays a very important role in resource exploration in the civil field and national defense deployment in the military field. The underwater acoustic navigation and positioning technology is one of the key technologies of an underwater vehicle, and currently, many problems exist, such as short transmission distance of acoustic signals and easy noise interference, so that research on the underwater acoustic navigation and positioning technology is not slow.

The above problems are mainly caused by two factors, one of which is the type of the beacon transmitting signal, and there are currently various acoustic communication modes in the field of underwater communication, such as single-frequency signals, chirp signals and some digital modulation technology signals, including Amplitude Shift Keying (ASK), Frequency Shift Keying (FSK) and Phase Shift Keying (PSK), etc. However, as the underwater environment is complex and changeable, uncertain noise interference, multipath interference and Doppler effect interference exist, the transmission quality of the traditional acoustic signals under water is influenced by different degrees, mainly expressed as low signal-to-noise ratio and short acting distance, and thus the hidden danger of misidentification among different beacon signals is promoted. The other is a capture algorithm corresponding to the transmitted signal, which generally includes a wavelet de-noising algorithm, a matched filtering algorithm, a fractional interval equalization function algorithm, and the like. However, in the actual complex and variable application environment, the algorithms have limitations, algorithm parameters cannot be adjusted in real time to meet the variable denoising requirement, and the purpose of remote identification of multiple beacon signals is difficult to achieve.

In radar and ground positioning systems, a spread spectrum coding technology is often used, and the technology has the advantages of strong anti-interference capability, long measurement distance, high ranging accuracy and the like due to the long periodicity and sharp autocorrelation characteristics, and has an information communication function while ranging. The convolutional neural network is one of the representative algorithms of deep learning, has stable effect on learning of pixels and audio, and has no additional characteristic engineering requirement on data.

Disclosure of Invention

In view of the above, the present invention provides a CNN-based multi-underwater beacon signal identification method. The method can improve the action distance and the recognition precision of the underwater beacon and effectively solve the problem of insufficient transmission distance of the long-baseline beacon system.

The technical scheme of the invention is as follows:

a CNN-based multi-underwater beacon signal identification method, the method comprising the steps of:

1) the modulation of beacon emission signal adopts n-order spread spectrum coding technique to make each underwater beacon have only one corresponding coding sequence, and makes phase-frequency homoshift keying mixed modulation on said coding sequence, and different code elements have different phases and frequencies so as to form the emission signal of beacon, so that every beacon possesses unique phase and frequency characteristics, and is favourable for identification of convolutional neural network, and every beacon is formed from 2ⁿ1 symbol component of 1 or-1, respectively 2^n-11 code element and 2^n-11-1 symbols, where 1 denotes m₁F with phase of 0 DEG₁Filling Hz sine waves; -1 represents by m₂F with 180 DEG phase₂Hz sine wave filling. The modulated beacon analog signal is as follows:

when in useThen, the formula is simplified as follows (2):

where Y represents the beacon-modulated analog signal, T represents time, X [ i ]]Is shown in (2)ⁿ-1 [ i ] th of symbols]Sequence values, 1 or-1;

2) collecting various beacon signals;

3) performing data enhancement on the sampled beacon signal sequence set;

the sampling signal sequence for data set expansion comprises an actual sampling signal sequence and a theoretical sampling signal sequence, and the actual sampling signal sequence is subjected to amplitude value conversion; the theoretical sampling signal sequence is processed as follows: setting different amplitudes for the beacon signal itself; adding white noise with different amplitudes; randomly carrying out distortion processing on partial signals;

4) converting one-dimensional signal information into two-dimensional picture information through short-time Fourier transform, wherein the process is as follows: taking a beacon signal period t₁The sampling sequence of ms time is the conversion target of one frame, and t is₂The length of the signal sampling sequence of ms time is the window length, and t is used₃The length of the signal sampling sequence of ms time is the translation step length to perform frame processing on the conversion target of one frame, and the number M of split sub-frames_fAs shown in formula (3):

wherein | represents integer division, |. represents non-integer division;

then, performing discrete fourier transform on each subframe sequence x (n) as shown in formula (4):

wherein X (k) is the sequence x (n) after DFT, f_sRepresenting the sampling rate (KHz), t₂Represents time (ms), f_s t₂Number of sampling points representing the length of a window, the frequency resolution being 1/t₂；

Thereby obtaining 0-f according to the formula (5)_sThe amplitude of the frequency component is shifted from the cosine in accordance with equation (6).

P(k)＝real(X(k))²+imag(X(k))² (5)

Calculating to obtain the component amplitude and offset phase of a subframe and drawing t according to the component amplitude and offset phase₃Time-frequency diagram and time phase diagram of ms time, M_fThe time-frequency diagram and the phase diagram of each sub-frame are respectively spliced into a time length t_3*M_fA time-frequency diagram and a phase diagram of ms (one frame), and finally, two characteristic diagrams generated by the signal sequence of the frame are fused into a phase-frequency characteristic diagram with pixels of 100 × 3;

the window length of the short-time Fourier transform adopts the design of an overlapped sliding window, and information in a longer time is mapped in a shorter time interval in the form of frequency component intensity. Using for a longer time (t)₂ms) as a window length to ensure frequency resolution accuracy; while using a shorter time (t)₃ms) as the step length of the sliding window to ensure the time resolution precision, so that the information difference between adjacent sub-frames is only reflected in the t of the head of the last sub-frame₃ms interval and t at the end of the next subframe₃The ms interval, this method is to improve the sensitivity to the signal change to give consideration to frequency resolution and time resolution, is suitable for the recognition of the beacon under the multi-water;

5) performing data enhancement on the picture data set again;

the data enhancement method for the data set of the phase-frequency characteristic diagram has the following three modes: firstly, setting different brightness, saturation and contrast for a phase frequency diagram by taking a beacon signal frequency band as a reference; secondly, performing Gaussian filtering processing and fuzzy processing of different degrees on the picture; thirdly, the two treatments are carried out simultaneously.

6) Training the data set by using a lightweight CNN network to obtain a network model for realizing real-time identification of various beacon signals;

the convolutional neural network maps a spectrogram onto a beacon sequence number, and the network structure of the convolutional neural network is composed of a plurality of nonlinear structures, which are 1 input layer, 4 convolutional layers, 3 pooling layers, 1 fully-connected layer and 1 softmax layer, wherein the first three convolutional layers have pooling layers, and from the input layer to the output layer, the convolutional layers with 64 3 × 3 convolutional kernels, the maximum pooling layer of 2 × 2, the convolutional layers with 32 3 × 3 convolutional kernels, the maximum pooling layer of 2 × 2, the convolutional layers with 12 3 × 3 convolutional kernels, the maximum pooling layer of 2 × 2, the convolutional layers with 8 × 3 convolutional kernels, the fully-connected layer with 128 neurons, and the last softmax layer are taken as output layers.

The invention has the beneficial effects that: the range and the recognition accuracy of the beacon are effectively improved, and the lightweight model design is also suitable for real-time recognition of the signal.

Drawings

Fig. 1 is a flow chart of a CNN-based multi-underwater beacon signal identification method.

Fig. 2 is a table of the 7 th spreading code sequence for beacon # 1.

Fig. 3 is a schematic diagram of a beacon part transmission signal based on 7-order spread spectrum coding and phase-frequency co-shift keying.

Fig. 4 is a schematic diagram of the short-time fourier transform principle.

Fig. 5 is a time phase frequency diagram of a beacon signal for CNN network model training.

Fig. 6 is a schematic diagram of a convolutional neural network structure.

Detailed Description

To describe the present invention in more detail, further detailed description is given below with reference to the accompanying drawings.

Referring to fig. 1, a CNN-based multi-underwater beacon signal identification method includes the following steps:

1) modulating beacon transmitting signals, namely, adopting a 7-order spread spectrum coding technology to enable each underwater beacon to have only one corresponding coding sequence, and performing phase-frequency homoshift keying mixed modulation on the coding sequences, wherein different code elements have different phases and frequencies to form the transmitting signals of the beacons, and each beacon has unique phase and frequency characteristics to facilitate the identification of a convolutional neural network; -1 represents the filling with 10 20KHz sine waves 180 ° in phase, at this timeThe formula is simplified as formula (7):

wherein Y represents the analog signal after beacon modulation, T represents time, X [ i ] represents the [ i ] th sequence value in 127 code elements, 1 or-1; the period of a beacon signal can be calculated to be 63.5ms from the formula, and the partial transmission signal of the beacon 1 is shown in fig. 3 and is seen to be composed of sine waves with different phases and frequencies;

2) collecting various beacon signals;

3) performing data enhancement on the sampled beacon signal sequence set;

the sample signal sequence for data set expansion includes an actual sample signal sequence and a theoretical sample signal sequence. The two data set expansion methods are different, the former already contains various different environmental noises, so only the amplitude is transformed; for the latter, the following treatment is performed: setting different amplitudes for the beacon signal itself; adding white noise with different amplitudes; randomly carrying out distortion processing on partial signals;

4) by short timeThe fourier transform converts one-dimensional signal information into two-dimensional picture information, and the process is shown in fig. 4: taking a sampling sequence of 64ms time in a beacon signal period as a conversion target of one frame, taking the length of a signal sampling sequence of 0.5ms time as a window length, taking the length of the signal sampling sequence of 0.1ms time as a translation step length to perform framing processing on the conversion target of one frame, wherein the number M of split sub-frames_fAs shown in formula (8):

where | represents an integer division and | represents a non-integer division. Obtaining M according to calculation_fIs 636.

Then, performing discrete fourier transform on each subframe sequence x (n) as shown in formula (9):

wherein X (k) is the sequence x (n) after DFT, f_sRepresenting the sampling rate (KHz), t₂Represents time (ms), f_s t₂Number of samples representing the length of a window, where f_sTake 100KHz, t₂The time is 0.5ms, the length of one window obtained by calculation is 50 sampling points, and the frequency resolution is 2 KHz;

thereby obtaining the component amplitude of 0-100KHz frequency according to the formula (10), and obtaining the offset phase relative to cosine according to the formula (11);

P(k)＝real(X(k))²+imag(X(k))² (10)

and calculating to obtain the component amplitude and the offset phase of one subframe, drawing a time-frequency diagram and a time-phase diagram of 0.1ms time according to the component amplitude and the offset phase, and respectively splicing the time-frequency diagrams and the phase diagrams of 636 subframes into a time-frequency diagram and a phase diagram with the time length of 63.6ms (one frame). And finally, fusing the two feature maps generated by the signal sequence into a phase-frequency feature map with pixels of 100 × 3, as shown in fig. 5.

The window length of the short-time Fourier transform adopts the design of an overlapping sliding window, information of a long time is mapped in a short time interval in the form of frequency component intensity, and the sequence length of the long time (0.5ms) is used as the window length to ensure the frequency resolution precision; meanwhile, the sequence length of shorter time (0.1ms) is used as the step length of a sliding window to ensure the time resolution precision, so that the information difference between adjacent subframes is only embodied in the 0.1ms interval at the head of the last subframe and the 0.1ms interval at the tail of the next subframe, the method improves the sensitivity to signal change to take frequency resolution and time resolution into account, and is suitable for the identification of multi-underwater beacons;

5) performing data enhancement on the picture data set again;

the data enhancement method for the data set of the phase-frequency characteristic diagram has the following three modes: firstly, setting different brightness, saturation and contrast for a phase frequency diagram by taking a beacon signal frequency band as a reference; secondly, performing Gaussian filtering processing and fuzzy processing of different degrees on the picture; thirdly, the two treatments are carried out simultaneously.

6) Training the data set by using a lightweight CNN network to obtain a network model for realizing real-time identification of various beacon signals;

fig. 6 is a schematic diagram of a convolutional neural network structure, which maps a spectrogram onto a beacon sequence number, and the network structure of the convolutional neural network is composed of a number of nonlinear structures, which are 1 input layer, 4 convolutional layers, 3 pooling layers, 1 fully-connected layer, and 1 softmax layer, where the first three convolutional layers have pooling layers, and specifically, a picture input layer, a convolutional layer having 64 3 × 3 convolutional cores, a maximum pooling layer of 2 × 2, a convolutional layer having 32 3 × 3 convolutional cores, a maximum pooling layer of 2 × 2, a convolutional layer having 12 3 × 3 convolutional cores, a maximum pooling layer of 2 × 2, a convolutional layer having 8 3 × 3 convolutional cores, a fully-connected layer having 128 neurons, and a last softmax layer are sequentially arranged from an input layer to an output layer.