Universal underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network

Abstract

The present invention discloses a universal underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network, which belongs to the technical field of underwater acoustic communication signal demodulation. It can solve the technical problems of poor versatility, low accuracy and high complexity in the existing underwater acoustic signal demodulation technology. Firstly, Matlab and BELLHOP are used to generate the underwater acoustic modulation signal under the simulated underwater acoustic channel, and then the signal features are enriched by using Hilbert transform to construct a four-channel sample; then an improved lightweight channel convolutional network is designed to combine deep separable convolution, GELU activation function, channel attention mechanism and shuffling method, which effectively extracts the key features of the signal in the time domain and reduces the computational complexity. In addition, the application of layer normalization ensures the stability of the network and improves the training effect; finally, the method realizes the accurate demodulation of multiple categories of underwater acoustic signals such as PSK, ASK, FSK, QAM, etc. through the Softmax function.

Description

Technical Field

[0001] The invention relates to the technical field of underwater acoustic communication signal demodulation, and in particular to a universal underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network. Background Art

[0002] Underwater communication refers to a communication method that uses water as a transmission medium and transmits information through means such as sound waves and light waves. It is mainly used in underwater detection, marine resource development, marine safety and other fields. With the continuous advancement of science and technology, human beings are increasingly developing and utilizing marine resources, and the importance of underwater communication technology is becoming increasingly prominent. Since underwater acoustic communication using sound waves as a transmission method has the characteristics of relatively small attenuation and strong penetration ability, it is currently the only way to achieve long-distance underwater communication.

[0003] However, underwater acoustic communication faces many unique and severe challenges. First, the high absorption and scattering characteristics of seawater lead to high loss and multipath effects in the signal transmission process, making the signal susceptible to interference, attenuation and distortion during transmission, making it difficult for the communication receiver to accurately receive data. Therefore, it is crucial to study stable and efficient demodulation methods.

[0004] As a method of converting the amplitude spectrum of the original signal into the phase spectrum, the Hilbert transform produces a complex function as a result, which contains the amplitude and phase information of the original signal. Through the Hilbert transform, various important information can be extracted from the original signal, such as the instantaneous frequency, phase and amplitude changes of the signal.

[0005] Traditional demodulation methods such as coherent demodulation usually require accurate phase and frequency information of the signal, which is difficult to adapt to complex and changeable channel characteristics, and has poor demodulation performance for signals with low signal-to-noise ratios. Especially under non-cooperative communication conditions, it is difficult to quickly obtain information such as the frequency and phase of the signal. If the transmitter uses a variety of mixed modulation methods, the design of the receiver is very difficult.

[0006] Therefore, there is an urgent need for a method with strong versatility, low complexity and high accuracy to achieve universal and efficient demodulation of underwater acoustic signals. Summary of the invention

[0007] The purpose of the present invention is to provide a universal underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network with strong versatility, low complexity and high accuracy.

[0008] To achieve the above object, the present invention provides a general underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network, comprising the following steps:

[0009] S1. Use Matlab and BELLHOP to generate the underwater acoustic modulation signal under the simulated underwater acoustic channel, and divide the signal into signal samples according to the symbol, and use the symbol as the label of the signal sample; use the normalization method and Hilbert transform to preprocess the data set, where the continuous Hilbert transform converts the original signal into four parts and splices them to form a four-channel sample data set to prepare for input into the subsequent convolutional neural network;

[0010] S2. Design an improved convolutional neural network model. The network backbone consists of four stages. Each stage contains a convolution block and a downsampling layer. A convolution layer and a normalization layer are added before the first stage. A fully connected layer and a Softmax activation function are added after the fourth stage to output the symbol category of the corresponding signal.

[0011] S3. Train the model based on the data set generated in step S1. After sufficient training, the model has the ability to learn to extract the time domain features of signal samples and classify them; use the trained model to classify unknown signal samples, thereby achieving signal demodulation.

[0012] Preferably, step S1 includes step S11 of generating a simulated underwater acoustic signal, and the specific steps are as follows:

[0013] First, determine the number of signal symbols, frequency, phase, modulation mode and other parameters, and use Matlab to generate symbols of dimension (1, 1), where the number of symbol categories corresponds to the modulation order; use PSK, FSK, ASK and QAM modulation methods to baseband modulate the signal, and then use the root raised cosine roll-off filter to perform pulse shaping, and finally multiply the modulated signal with the carrier and perform digital-to-analog conversion to form a frequency band signal for transmission; the frequency band signal transmitted is expressed as follows:

[0014] s t (t)=A(t)·cos(2πf(t)·t+φ(t));

[0015] Among them, s t (t) represents the frequency band signal, t represents time, A(t), f(t), φ(t) represent the amplitude, frequency and phase of the signal, these three parameters can be constants or functions of time, which depends on the modulation method adopted; when the transmitted symbol changes, the corresponding function value also changes;

[0016] The frequency band is transmitted in the channel and superimposed with various noises, and is finally received by the receiving end. The formula of the signal is expressed as follows:

[0017] s r (t) = H(t)·s t (t)+n(t);

[0018] Among them, s r (t) represents the signal received by the receiver, H(t) represents the channel equation, and n(t) represents additive white Gaussian noise;

[0019] Finally, the signal received by the receiving end is taken as a data set, segmented according to the symbol, and the time domain signal data corresponding to each symbol is taken as a sample, and the symbol is used as a label.

[0020] Preferably, step S1 further includes step S12 of preprocessing the data set using a normalization method, and the specific steps are as follows:

[0021] After the receiver receives the signal, it must first normalize the signal data to reduce the interference of noise sampling points with large amplitude offset on the data and enhance the stability of model training; normalization is performed with each carrier cycle as a window to compress the data while ensuring the signal time domain shape characteristics; the specific process is shown in the following formula:

[0022]

[0023] Among them, S norm represents the signal after normalization, S represents the signal before normalization, and S max and S min Respectively represent the maximum and minimum values ​​of the signal within the carrier period;

[0024] Preferably, step S1 further includes S13 preprocessing the data set using Hilbert transform, and the specific steps are as follows: using continuous Hilbert transform to process the signal to obtain multiple groups of mappings of the original samples, and the formula of the Hilbert transform is as follows:

[0025]

[0026] where s r (t) is the signal sample at time t, and H(s(t))(t) is its Hilbert transform. Through the Hilbert transform, a composite signal containing the amplitude and phase information of the original signal is obtained. The amplitude of the signal is obtained by taking the modulus of the signal after the Hilbert transform. The formula is as follows:

[0027] A h (t)=|H(s r (t))(t)|;

[0028] Among them A h (t) the amplitude of the transformed signal;

[0029] By calculating the phase angle represented by the composite signal, we can get a phase that is 90 degrees different from the original signal. The formula is as follows:

[0030] φ h (t) = arg(H(s r (t))(t));

[0031] Among them, φ h (t) is the phase of the transformed signal; combined with the frequency f(t) of the original signal, we finally get a signal s that is 90 degrees out of phase with the original signal. h (t), the formula is as follows:

[0032] s h (t) = A h (t)·cos(f(t)+φ h (t));

[0033] For the processed signal, continue to use Hilbert transform to calculate twice continuously, and extract the signal calculated each time, thus obtaining four signals s r (t), s h (t), s h '(t),s h ” (t), its h '(t),s h ”(t) The formula is as follows:

[0034] s h '(t)=A h '(t)·cos(f(t)+φ h '(t))

[0035] s h ”(t)=A h ” (t)·cos(f(t)+φ h ” (t));

[0036] where s h '(t),s h ” (t) are the signals obtained by the second and third Hilbert transform of the original signal, A h '(t),φ h '(t) are the amplitude and phase of the signal obtained by the second Hilbert transform, A h ” (t), φ h ” (t) are the amplitude and phase of the signal obtained by the third Hilbert transform;

[0037] The signals before and after the transformation are superimposed, and the original single-channel samples will become four-channel samples, ready to be input into the subsequent convolutional neural network.

[0038] Preferably, step S2 designs an improved convolutional neural network model, and the specific steps include the following: S21 establishes a shallow feature extraction module; S22 establishes a convolution block; S23 establishes an efficient channel attention module; S24 establishes a downsampling layer and a classifier;

[0039] Wherein, step S21 is specifically as follows:

[0040] The four-channel samples processed by S1 are sent to a shallow feature extraction structure consisting of a one-dimensional convolution layer with a convolution kernel size of 1×3 and a normalization layer to perform preliminary feature extraction and increase the number of channels; the number of input channels of the convolution layer is set to 4, and the number of output channels is set to 24, that is, after the convolution operation of the 4-channel signal samples carrying 4 kinds of phase information, 24 feature maps are finally output, each of which carries different information of the original 4-channel signal; the calculation formula of the one-dimensional convolution is as follows:

[0041]

[0042] Among them, C_I1 represents the input of the first convolutional layer, C_W 1,j and C_B 1,j Represent the j-th weight matrix and bias term in the first convolutional layer, C_O 1,j represents the output of the jth convolution kernel in the first convolutional layer, and C_A represents the activation function. A padding of size 1 is added during the processing to ensure that the output feature map is equal to the input sample size. The actual output dimension is [24, 64].

[0043] The layer normalization technology is used to process the feature map, and the output of each neuron in the convolution layer is normalized to achieve sample scale processing. The specific formula is as follows:

[0044]

[0045] Where L_O i Represents the normalized output of the layer, L_I i represents the normalized input of the layer, μ and σ represent the mean and standard deviation of the output of the neuron in the entire mini-batch, respectively, and ε is a small constant used to prevent the denominator from being zero.

[0046] Preferably, step S22 of establishing the convolution block is specifically as follows:

[0047] All feature maps are evenly divided into two groups in the channel dimension, one of which is an identity mapping, directly mapping the input to the output; then the other group of channels is processed, first using a one-dimensional convolutional layer of size 1x1 to perform linear combinations between different channels, preliminarily learn feature representation, and further use layer normalization to adjust data distribution; then use a one-dimensional depthwise separable convolution of size 1x3 to extract features, each filter of the depthwise separable convolution only operates on one feature map, and the activation function is set to GELU. The specific formula is as follows:

[0048] G o =G I φ(x);

[0049] Among them, G o , G I are the output and input of GELU respectively, φ(x) represents the cumulative distribution function of the standard normal distribution; after using depth-wise separable convolution to extract channel-level data, a one-dimensional convolution of size 1x1 is used again to aggregate the data of each channel.

[0050] Preferably, step S23 establishes an efficient channel attention module, specifically:

[0051] An efficient channel attention module is added to the convolution block. Specifically, it consists of an adaptive average pooling layer, a one-dimensional convolution layer with dynamically adjustable convolution kernel size, and a Sigmoid activation function. The module first uses the global adaptive average pooling technology to convert the feature map of each channel into a scalar. For the input data with C feature maps, all elements on each channel of the feature map are averaged to obtain a vector with a length of C, the number of channels. The vector is processed by one-dimensional convolution to obtain a vector with a length of C, which represents the attention weight of each channel. The size K of the convolution kernel is determined by the following method. The specific formula is:

[0052]

[0053] Among them, c is the number of input channels, b is a constant, usually set to 1, and γ is a scaling factor, usually set to 2; the weight of the convolution output is calculated using Sigmoid to obtain the final weight matrix, which is multiplied by the original feature map.

[0054] Preferably, step S24 establishes a downsampling layer and a classifier, specifically:

[0055] A downsampling layer is added between different convolution blocks, which is processed by a convolution kernel of size 3 and stride 2, and the number of output channels is doubled, and then normalized; the convolution block does not change the size and number of channels of the sample, while the downsampling operation will halve the size of the sample and double the number of channels;

[0056] The final output feature map will be input to the classifier, which contains a normalization layer, a fully connected layer, and a Softmax function to output the category probability corresponding to the modulation order; after the normalization layer integrates the data, the fully connected layer provides a smooth mapping to prepare for the final classification; the fully connected layer is represented as follows:

[0057]

[0058] Among them, L_I represents the input of the fully connected layer, L_O represents the output of the fully connected layer, L_W1 represents the weight matrix between the two hidden layers, L_W2 represents the weight matrix between the second hidden layer and the output layer, L_B1 and L_B2 represent bias terms, and L_A1 and L_A2 represent activation functions;

[0059] The final output layer uses Softmax to predict the category, and its formula is as follows:

[0060]

[0061] where z j Represents the input of Softmax, and n represents the number of neurons; Sofmax will output multiple probabilities and use the maximum probability of the output as the predicted category.

[0062] Preferably, step S3 is specifically as follows:

[0063] The mini-batch gradient descent method is used to train the neural network. The training set is divided into several groups of data. The neural network learns information from one group of data in each iteration. The multi-classification cross entropy function is used as the loss function of the network to predict the error between the predicted value and the true value of the neural network during the training process. The calculation formula of the multi-classification cross entropy function is expressed as follows:

[0064]

[0065] in, represents the loss function, represents the total predicted value, X represents the total true value, N represents the total number of output values, X i represents the i-th output value, represents the i-th predicted value;

[0066] After saving the trained model, input the unlabeled modulated signal, and the model will automatically output the predicted symbols to achieve signal demodulation.

[0067] Beneficial effects of the present invention:

[0068] The present invention discloses a universal underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network. Firstly, Hilbert transform is used to enhance the time domain characteristics of the signal. Then, the proposed improved convolutional neural network is used to extract and classify the features. The method can universally demodulate a variety of modulated signals.

[0069] The present invention specially designs each convolution block, uses depthwise separable convolution to efficiently extract the features of each channel, and utilizes efficient channel attention module, layer normalization technology and channel shuffling technology to aggregate each channel to resist the influence of noise, ensuring that the model can efficiently extract the most representative features.

[0070] In order to ensure the classification ability while reducing the complexity, the present invention specially designs the overall structure of the model: adopts a single-branch structure to avoid the wider model bringing more training and inference time; uses 1D convolution instead of 2D convolution; outputs fewer channels in the shallow layer of the model and more channels in the deep layer; and simplifies the design of the downsampling layer.

[0071] By comparing with the existing methods under Gaussian channels and BELLHOP simulated underwater acoustic channels, it is shown that the present invention has higher demodulation accuracy and lower complexity, can meet the requirements of real-time underwater acoustic communication, and has good application prospects in future communication systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is an overall model structure diagram of an embodiment of the present invention;

[0073] Figure 2 is an application diagram of an underwater acoustic communication system according to an embodiment of the present invention;

[0074] Figure 3 It is a convolution block structure and down-sampling structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0075] In order to make the technical means, creative features and effects achieved by the present invention easy to understand, the technical solutions in the embodiments of the present invention will be further clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0076] Example 1

[0077] Signals are easily affected by interference, attenuation and distortion during transmission, which makes it difficult for communication receivers to accurately receive data. The receiving end needs to recover and reconstruct the received signal and reduce the bit error rate under complex conditions. Existing receiving methods are difficult to achieve low bit error rate signal demodulation. Therefore, in order to accurately and efficiently demodulate underwater acoustic signals, a universal underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network is used to enhance the accuracy and versatility of underwater acoustic signal demodulation.

[0078] The general underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network of the present invention has an overall model structure diagram as shown in Figure 1 As shown, it specifically includes: using Matlab and BELLHOP to generate an underwater acoustic modulation signal under a simulated underwater acoustic channel, dividing the signal into samples according to symbols, and using continuous Hilbert transform to convert the original signal into four parts and splice them to form a four-channel sample.

[0079] An improved lightweight channel convolutional network is designed to process signal samples. The backbone of the network consists of four stages, each of which contains a convolution block and a downsampling layer. The convolution block uses depthwise separable convolution to improve the feature extraction capability for a single channel and reduce the complexity. The GELU function is used as the activation function to enhance the expressiveness of the network, and an efficient channel attention mechanism and channel shuffling method are introduced to enhance the feature interaction between channels.

[0080] In addition, layer normalization technology is used to normalize the output to ensure the stability of the network and better training results. At the end of the model, the Softmax function is used to output the probability of the corresponding category. The present invention uses the idea of ​​classification to reconstruct the original symbol corresponding to the signal. The specially designed convolution block can efficiently extract the universal frequency and phase characteristics of the signal in the time domain under low complexity conditions, and can realize end-to-end demodulation of multiple categories of underwater acoustic signals such as PSK, ASK, FSK, QAM, etc., solving the problems of poor versatility, low accuracy and high complexity of traditional methods.

[0081] Example 2

[0082] The present invention provides a universal underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network, such as Figure 1 As shown, the following steps are included:

[0083] S1. Use Matlab and BELLHOP to generate the underwater acoustic modulation signal under the simulated underwater acoustic channel, and divide the signal into signal samples according to the symbol, and use the symbol as the label of the signal sample; use the normalization method and Hilbert transform to preprocess the data set, where the continuous Hilbert transform converts the original signal into four parts and splices them to form a four-channel sample data set to prepare for input into the subsequent convolutional neural network. The specific steps are as follows:

[0084] S11 generates a simulated underwater acoustic signal, and the specific steps are as follows:

[0085] First, determine the number of signal symbols, frequency, phase, modulation mode and other parameters, and use MATLAB to generate symbols of dimension (1, 1), where the number of symbol categories corresponds to the modulation order; use PSK, FSK, ASK and QAM modulation methods to baseband modulate the signal, and then use the root raised cosine roll-off filter to perform pulse shaping, and finally multiply the modulated signal with the carrier and perform digital-to-analog conversion to form a frequency band signal for transmission; the frequency band signal transmitted is expressed as follows:

[0086] s t (t)=A(t)·cos(2πf(t)·t+φ(t));

[0087] Among them, s t (t) represents the frequency band signal, t represents time, A(t), f(t), φ(t) represent the amplitude, frequency and phase of the signal, these three parameters can be constants or functions of time, which depends on the modulation method adopted; when the transmitted symbol changes, the corresponding function value also changes;

[0088] The frequency band is transmitted in the channel and superimposed with various noises, and is finally received by the receiving end. The formula of the signal is expressed as follows:

[0089] s r (t) = H(t)·s t (t)+n(t);

[0090] Among them, s r (t) represents the signal received by the receiver, H(t) represents the channel equation, and n(t) represents additive white Gaussian noise;

[0091] Finally, the signal received by the receiving end is taken as a data set, segmented according to the symbol, and the time domain signal data corresponding to each symbol is taken as a sample, and the symbol is used as a label.

[0092] S12 uses the normalization method to preprocess the data set. The specific steps are as follows:

[0093] After the receiver receives the signal, it must first normalize the signal data to reduce the interference of noise sampling points with large amplitude offset on the data and enhance the stability of model training. Strong noise interference in the channel causes the amplitude of some sampling points to deviate far from the range of the transmitted signal. If normalization is performed directly on the entire data set, the signal loses its envelope characteristics, making it difficult for the model to extract information from it. Therefore, normalization is performed with each carrier cycle as a window to compress the data while ensuring the signal's time domain shape characteristics. The specific process is shown in the following formula:

[0094]

[0095] Among them, S norm represents the signal after normalization, S represents the signal before normalization, and S max and S min Respectively represent the maximum and minimum values ​​of the signal within the carrier period;

[0096] S13 uses Hilbert transform to preprocess the data set. The specific steps are as follows:

[0097] Continuous Hilbert transform is used to process the signal to obtain multiple sets of mappings of the original samples. The formula of the Hilbert transform is as follows:

[0098]

[0099] where s r (t) is the signal sample at time t, and H(s(t))(t) is its Hilbert transform. Through the Hilbert transform, a composite signal containing the amplitude and phase information of the original signal is obtained. The amplitude of the signal is obtained by taking the modulus of the signal after the Hilbert transform. The formula is as follows:

[0100] A h (t)=|H(s r (t))(t)|;

[0101] Among them A h (t) The amplitude of the signal obtained by transformation; Calculate the phase angle represented by the composite signal to obtain a phase that is 90 degrees different from the original signal. The formula is as follows:

[0102] φ h (t) = arg(H(s r (t))(t));

[0103] Among them, φ h (t) is the phase of the transformed signal; combined with the frequency f(t) of the original signal, we finally get a signal s that is 90 degrees out of phase with the original signal. h (t), the formula is as follows:

[0104] s h (t) = A h (t)·cos(f(t)+φ h (t));

[0105] For the processed signal, continue to use Hilbert transform to calculate twice continuously, and extract the signal calculated each time, thus obtaining four signals s r (t), sh (t), s h '(t),s h ” (t), its h '(t),s h ”(t) The formula is as follows:

[0106] s h '(t)=A h '(t)·cos(f(t)+φ h '(t))

[0107] s h ”(t)=A h ” (t)·cos(f(t)+φ h ” (t));

[0108] where s h '(t),s h ” (t) are the signals obtained by the second and third Hilbert transform of the original signal, A h '(t),φ h '(t) are the amplitude and phase of the signal obtained by the second Hilbert transform, A h ” (t), φ h ” (t) are the amplitude and phase of the signal obtained by the third Hilbert transform;

[0109] The signals before and after the transformation are superimposed, and the original single-channel samples will become four-channel samples, ready to be input into the subsequent convolutional neural network.

[0110] S2: Design an improved convolutional neural network model. The network backbone consists of four stages. Each stage contains a convolution block and a downsampling layer. A convolution layer and a normalization layer are added before the first stage. A fully connected layer and a Softmax activation function are added after the fourth stage to output the symbol category of the signal.

[0111] S21 establishes a shallow feature extraction module, and the specific steps are as follows:

[0112] The four-channel samples processed by S1 are sent to a shallow feature extraction structure consisting of a one-dimensional convolution layer with a convolution kernel size of 1×3 and a normalization layer to perform preliminary feature extraction and increase the number of channels; the number of input channels of the convolution layer is set to 4, and the number of output channels is set to 24, that is, after the convolution operation of the 4-channel signal samples carrying 4 kinds of phase information, 24 feature maps are finally output, each of which carries different information of the original 4-channel signal; the calculation formula of the one-dimensional convolution is as follows:

[0113]

[0114] Among them, C_I1 represents the input of the first convolutional layer, C_W 1,j and C_B 1,j Represent the j-th weight matrix and bias term in the first convolutional layer, C_O 1,j represents the output of the jth convolution kernel in the first convolutional layer, C_A represents the activation function; a padding of size 1 is added during the processing to ensure that the output feature map is equal to the input sample size, and the actual output dimension is [24, 64].

[0115] The layer normalization technology is used to process the feature map, and the output of each neuron in the convolution layer is normalized to achieve sample scale processing. The specific formula is as follows:

[0116]

[0117] Where L_O i Represents the normalized output of the layer, L_I i represents the normalized input of the layer, μ and σ represent the mean and standard deviation of the output of the neuron in the entire mini-batch, respectively, and ε is a small constant used to prevent the denominator from being zero.

[0118] S22 establishes a convolution block, and the specific steps are as follows:

[0119] All feature maps are evenly divided into two groups in the channel dimension, one of which is an identity mapping, directly mapping the input to the output; then the other group of channels is processed, first using a one-dimensional convolutional layer of size 1x1 to perform linear combinations between different channels, preliminarily learn feature representation, and further use layer normalization to adjust data distribution; then use a one-dimensional depthwise separable convolution of size 1x3 to extract features, each filter of the depthwise separable convolution only operates on one feature map, and the activation function is set to GELU. The specific formula is as follows:

[0120] G o =G I φ(x);

[0121] Among them, G o , G I are the output and input of GELU respectively, φ(x) represents the cumulative distribution function of the standard normal distribution; after using depth-wise separable convolution to extract channel-level data, a one-dimensional convolution of size 1x1 is used again to aggregate the data of each channel.

[0122] S23 establishes an efficient channel attention module. The specific steps are as follows:

[0123] An efficient channel attention module is added to the convolution block. Specifically, it consists of an adaptive average pooling layer, a one-dimensional convolution layer with dynamically adjustable convolution kernel size, and a Sigmoid activation function. The module first uses the global adaptive average pooling technology to convert the feature map of each channel into a scalar. For the input data with C feature maps, all elements on each channel of the feature map are averaged to obtain a vector with a length of C, the number of channels. The vector is processed by one-dimensional convolution to obtain a vector with a length of C, which represents the attention weight of each channel. The size K of the convolution kernel is determined by the following method. The specific formula is:

[0124]

[0125] Among them, c is the number of input channels, b is a constant, usually set to 1, and γ is a scaling factor, usually set to 2; the weight of the convolution output is calculated using Sigmoid to obtain the final weight matrix, which is multiplied by the original feature map.

[0126] S24 establishes a downsampling layer and a classifier, and the specific steps are as follows:

[0127] A downsampling layer is added between different convolution blocks, which is processed by a convolution kernel of size 3 and stride 2, and the number of output channels is doubled, and then normalized; the convolution block does not change the size and number of channels of the sample, while the downsampling operation will halve the size of the sample and double the number of channels;

[0128] The final output feature map will be input to the classifier, which contains a normalization layer, a fully connected layer, and a Softmax function to output the category probability corresponding to the modulation order; after the normalization layer integrates the data, the fully connected layer provides a smooth mapping to prepare for the final classification; the fully connected layer is represented as follows:

[0129]

[0130] Among them, L_I represents the input of the fully connected layer, L_O represents the output of the fully connected layer, L_W1 represents the weight matrix between the two hidden layers, L_W2 represents the weight matrix between the second hidden layer and the output layer, L_B1 and L_B2 represent bias terms, and L_A1 and L_A2 represent activation functions;

[0131] The final output layer uses Softmax to predict the category, and its formula is as follows:

[0132]

[0133] where z jRepresents the input of Softmax, and n represents the number of neurons; Sofmax will output multiple probabilities and use the maximum probability of the output as the predicted category.

[0134] S3. Train the model based on the data set generated in step S1. After sufficient training, the model has the ability to learn to extract the time domain features of signal samples and classify them, and use the trained model to classify unknown signal samples, thereby achieving signal demodulation.

[0135] The S3 step is as follows:

[0136] The mini-batch gradient descent method is used to train the neural network. The training set is divided into several groups of data. The neural network learns information from one group of data in each iteration. The multi-classification cross entropy function is used as the loss function of the network to predict the error between the predicted value and the true value of the neural network during the training process. The calculation formula of the multi-classification cross entropy function is expressed as follows:

[0137]

[0138] in, represents the loss function, represents the total predicted value, X represents the total true value, N represents the total number of output values, X i represents the i-th output value, Represents the i-th predicted value.

[0139] After saving the trained model, input the unlabeled modulated signal, and the model will automatically output the predicted symbols to achieve signal demodulation.

[0140] In summary, the present invention proposes a general underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network. The simulated underwater acoustic signal is generated by Matlab and BELLHOP, the signal time domain characteristics are enhanced by Hilbert transform, and an improved lightweight channel convolutional network is designed to process the signal. The network adopts deep separable convolution, GELU activation function, channel attention mechanism and channel shuffling technology, as well as layer normalization, to achieve efficient extraction of the frequency and phase characteristics of the signal in the time domain while maintaining low complexity. The method can demodulate multiple categories of underwater acoustic signals such as PSK, ASK, FSK, QAM, etc. end-to-end, significantly improves the demodulation accuracy, reduces the complexity, and has good anti-noise ability, meeting the needs of real-time underwater acoustic communication. Compared with the existing methods, the present invention shows significant advantages in both demodulation accuracy and complexity, and shows broad prospects for the application of future communication systems.

[0141] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for a person skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to replace some of the technical features therein by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions claimed to be protected by the present invention.

Claims

1. A general underwater acoustic signal demodulation method based on Hilbert transform and improved convolutional neural network, characterized in that: The following steps are involved: S1. Use Matlab and BELLHOP to generate the underwater acoustic modulation signal under the simulated underwater acoustic channel, and divide the signal into signal samples according to the symbol, and use the symbol as the label of the signal sample; use the normalization method and Hilbert transform to preprocess the data set, where the continuous Hilbert transform converts the original signal into four parts and splices them to form a four-channel sample data set to prepare for input into the subsequent convolutional neural network; S2. Design an improved convolutional neural network model. The network backbone consists of four stages. Each stage contains a convolution block and a downsampling layer. A convolution layer and a normalization layer are added before the first stage. A fully connected layer and a Softmax activation function are added after the fourth stage to output the symbol category of the corresponding signal. S3, training a model based on the data set generated in step S1. After sufficient training, the model has the ability to learn to extract time domain features of signal samples and classify them; using the trained model to classify unknown signal samples, thereby achieving signal demodulation; Step S1 includes step S11 of generating a simulated underwater acoustic signal, and the specific steps are as follows: First, determine the signal parameters, including the number of symbols, frequency, phase, and modulation mode, and use Matlab to generate symbols of dimension (1, 1), where the number of symbol categories corresponds to the modulation order; use PSK, FSK, ASK, and QAM modulation methods to baseband modulate the signal, and then use a root raised cosine roll-off filter to perform pulse shaping. Finally, multiply the modulated signal with the carrier and perform digital-to-analog conversion to form a frequency band signal for transmission; the frequency band signal transmitted is expressed as follows: s t (t)=A(t)·cos(2πf(t)·t+φ(t)); Among them, s t (t) represents the frequency band signal, t represents time, A(t), f(t), φ(t) represent the amplitude, frequency and phase of the signal, these three parameters can be constants or functions of time, which depends on the modulation method adopted; when the transmitted symbol changes, the corresponding function value also changes; The frequency band signal is transmitted in the channel and superimposed with various noises, and is finally received by the receiving end. The formula of the signal is expressed as follows: s r (t)=H(t)·s t (t)+n(t); Among them, s r (t) represents the signal received by the receiver, H(t) represents the channel equation, and n(t) represents additive white Gaussian noise; Finally, the signal received by the receiving end is used as a data set, segmented according to the symbol, and the time domain signal data corresponding to each symbol is used as a sample, and the symbol is used as a label; Step S1 also includes step S12 of preprocessing the data set using a normalization method, and the specific steps are as follows: After the receiver receives the signal, it must first normalize the signal data to reduce the interference of noise sampling points with large amplitude offset on the data and enhance the stability of model training; normalization is performed with each carrier cycle as a window to compress the data while ensuring the signal time domain shape characteristics. The specific process is shown in the following formula: Among them, S norm represents the signal after normalization, S represents the signal before normalization, and S max and S min Respectively represent the maximum and minimum values ​​of the signal within the carrier period; Step S1 also includes S13 using Hilbert transform to pre-process the data set, and the specific steps are as follows: Continuous Hilbert transform is used to process the signal to obtain multiple sets of mappings of the original samples. The formula of the Hilbert transform is as follows: where s r (t) is the signal sample at time t, and H(s(t))(t) is its Hilbert transform. Through the Hilbert transform, a composite signal containing the amplitude and phase information of the original signal is obtained. The amplitude of the signal is obtained by taking the modulus of the signal after the Hilbert transform. The formula is as follows: A h (t)=|H(s r (t))(t)|; Among them A h (t) The amplitude of the signal obtained by transformation; Calculate the phase angle represented by the composite signal to obtain a phase that is 90 degrees different from the original signal. The formula is as follows: φ h (t)=arg(H(s r (t))(t)); Among them, φ h (t) is the phase of the transformed signal; combined with the frequency f(t) of the original signal, we finally get a signal s that is 90 degrees out of phase with the original signal. h (t), the formula is as follows: s h (t)=A h (t)·cos(f(t)+φ h (t)); For the processed signal, continue to use Hilbert transform to calculate twice continuously, and extract the signal calculated each time, thus obtaining four signals s r (t), s h (t), s h '(t),s h ” (t), its h '(t),s h ”(t) The formula is as follows: s h '(t)=A h '(t)·cos(f(t)+φ h '(t)) s h ”(t)=A h ”(t)·cos(f(t)+φ h ”(t)); where s h '(t),s h ” (t) are the signals obtained by the second and third Hilbert transform of the original signal, A h '(t),φ h '(t) are the amplitude and phase of the signal obtained by the second Hilbert transform, A h ” (t), φ h ” (t) are the amplitude and phase of the signal obtained by the third Hilbert transform; The signals before and after the transformation are superimposed, and the original single-channel samples will become four-channel samples, ready to be input into the subsequent convolutional neural network; Step S2 designs an improved convolutional neural network model, and the specific steps include the following: S21 establishes a shallow feature extraction module; S22 establishes a convolution block; S23 establishes an efficient channel attention module; S24 establishes a downsampling layer and a classifier; Wherein, step S21 is specifically as follows: The four-channel samples processed in step S1 are sent to a shallow feature extraction structure consisting of a one-dimensional convolution layer with a convolution kernel size of 1×3 and a normalization layer to perform preliminary feature extraction and increase the number of channels; the number of input channels of the convolution layer is set to 4, and the number of output channels is set to 24, that is, after the convolution operation of the 4-channel signal samples carrying 4 kinds of phase information, 24 feature maps are finally output, each of which carries different information of the original 4-channel signal; the calculation formula of the one-dimensional convolution is as follows: Among them, C_I1 represents the input of the first convolutional layer, C_W 1,j and C_B 1,j Represent the j-th weight matrix and bias term in the first convolutional layer, C_O 1,j represents the output of the jth convolution kernel in the first convolutional layer, and C_A represents the activation function. A padding of size 1 is added during the processing to ensure that the output feature map is equal to the input sample size. The actual output dimension is [24, 64]. The layer normalization technology is used to process the feature map, and the output of each neuron in the convolution layer is normalized to achieve sample scale processing. The specific formula is as follows: Where L_O i Represents the normalized output of the layer, L_I i Represents the normalized input of the layer, μ and σ represent the mean and standard deviation of the output of the neuron in the entire mini-batch, and ε is a small constant used to prevent the denominator from being 0; Step S22 establishes the convolution block specifically as follows: All feature maps are evenly divided into two groups in the channel dimension, one of which is an identity mapping, directly mapping the input to the output; then the other group of channels is processed, first using a one-dimensional convolutional layer of size 1x1 to perform linear combinations between different channels, preliminarily learn feature representation, and further use layer normalization to adjust data distribution; then use a one-dimensional depthwise separable convolution of size 1x3 to extract features, each filter of the depthwise separable convolution only operates on one feature map, and the activation function is set to GELU. The specific formula is as follows: G o =G I ·φ(x); Among them, G o , G I are the output and input of GELU respectively, φ(x) represents the cumulative distribution function of the standard normal distribution; after using depth-wise separable convolution to extract channel-level data, a one-dimensional convolution of size 1x1 is used again to aggregate the data of each channel; Step S23 establishes an efficient channel attention module, specifically: An efficient channel attention module is added to the convolution block. Specifically, it consists of an adaptive average pooling layer, a one-dimensional convolution layer with dynamically adjustable convolution kernel size, and a Sigmoid activation function. The module first uses the global adaptive average pooling technology to convert the feature map of each channel into a scalar. For the input data with C feature maps, all elements on each channel of the feature map are averaged to obtain a vector with a length of C, the number of channels. The vector is processed by one-dimensional convolution to obtain a vector with a length of C, which represents the attention weight of each channel. The size K of the convolution kernel is determined by the following method. The specific formula is: Among them, c is the number of input channels, b is a constant, usually set to 1, and γ is a scaling factor, usually set to 2; the weight of the convolution output is calculated using Sigmoid to obtain the final weight matrix, which is multiplied with the original feature map; Step S24 establishes a downsampling layer and a classifier, specifically: A downsampling layer is added between different convolution blocks, which is processed by a convolution kernel of size 3 and stride 2, and the number of output channels is doubled, and then normalized; the final output feature map will be input into the classifier, which contains a normalization layer, a fully connected layer and a Softmax function to output the category probability corresponding to the modulation order; after the normalization layer integrates the data, the fully connected layer provides a smooth mapping to prepare for the final classification; the representation of the fully connected layer is as follows: Among them, L_I represents the input of the fully connected layer, L_O represents the output of the fully connected layer, L_W1 represents the weight matrix between the two hidden layers, L_W2 represents the weight matrix between the second hidden layer and the output layer, L_B1 and L_B2 represent bias terms, and L_A1 and L_A2 represent activation functions; The final output layer uses Softmax to predict the category, and its formula is as follows: where z j Represents the input of Softmax, and n represents the number of neurons; Sofmax will output multiple probabilities, and the maximum probability of the output is used as the predicted category; Step S3 is specifically as follows: The mini-batch gradient descent method is used to train the neural network. The training set is divided into several groups of data. The neural network learns information from one group of data in each iteration. The multi-classification cross entropy function is used as the loss function of the network to predict the error between the predicted value and the true value of the neural network during the training process. The calculation formula of the multi-classification cross entropy function is expressed as follows: in, represents the loss function, represents the total predicted value, X represents the total true value, N represents the total number of output values, X i represents the i-th output value, represents the i-th predicted value; After saving the trained model, input the unlabeled modulated signal, and the model will automatically output the predicted symbols to achieve signal demodulation.

Images