A Neural Network-Based Deinterleaving Method for Satellite-Borne AIS Signals

Abstract

The present invention belongs to the field of computer science artificial intelligence algorithms and relates to a signal deinterleaving method for a satellite-based ship automatic navigation system. The present invention first separates interleaved signals using a multi-channel AIS signal separation neural network. After separation, the separated signal is correlated with an AIS pilot reference signal and the maximum output is taken to determine the incident angle phase of the separated signal, and the steering vector of the separated AIS signal is calculated. Finally, the separated AIS signal is beamformed using minimum variance distortionless response beamforming technology to enhance the signal. The present invention combines array signal processing with convolutional neural network technology, improving separation performance while resisting mainlobe interference. This solves the problem that traditional blind source separation algorithms suffer performance degradation when the intensity of mainlobe interference is excessive.

Description

Technical Field

[0001] The invention belongs to the field of computer science artificial intelligence algorithms and relates to a signal deinterleaving method for a satellite-borne ship automatic navigation system. Background Art

[0002] The Automatic Identification System (AIS) is a new type of maritime safety and navigation aid system. AIS's functions include identifying ships, assisting in tracking targets, and exchanging ship information. Although AIS is powerful, the number, types, and size of ships on the oceans are constantly increasing, while the available bandwidth of the AIS channel is only 25 kHz. The available frequency band resources are very limited, resulting in increasingly serious time slot conflicts in AIS reception signals. Existing blind separation algorithms for resolving time slot conflicts are mostly derived in the absence of noise and interference, which does not meet the requirements of satellite-borne AIS satellite communications, which have low signal-to-noise ratios and the presence of mainlobe interference. Therefore, improving the ability of satellite-borne AIS systems to receive mixed signals with the same frequency is urgent. Summary of the Invention

[0003] This paper aims to address the issue of insufficient mainlobe interference resistance in satellite-borne AIS systems, particularly in the face of received signal time slot collisions. A deinterleaving method for satellite-borne AIS signals is proposed. This deinterleaving algorithm, based on a deep neural network, combines the powerful feature extraction capabilities of neural networks with the robust interference resistance of array signal processing beamforming to produce separated time-domain signals.

[0004] The technical solution of the present invention is:

[0005] The deinterleaving method for satellite-borne AIS signals is characterized by being based on a deep neural network and combined with array beamforming to demodulate AIS signals that are resistant to mainlobe interference and time slot collision. Figure 1 The specific steps are as follows:

[0006] Step 1: Create training data:

[0007] Building training data involves the following steps:

[0008] (1) Generate an AIS message sequence. Randomly generate a 208-bit AIS message stream of approximately 26.67ms, which contains 168 bits of data, a 24-bit training sequence, and an 8-bit start flag. The data transmission bit rate is 9600 bit / s, the modulation method is GMSK, the modulation index is 0.5, the Gaussian filter time-bandwidth product BT = 0.4, the oversampling factor is 8, the bit data is encoded using NRZI, and the modulation model is as shown in the attached figure. Figure 2As shown in Figure 1, an AIS signal s1 is generated. An interference signal s2 from another communication system is generated. The interference signal modulation method is still GMSK modulation. The real AIS signal is saved as the label for supervised training.

[0009] (2) Assuming that the number of receiving array elements is 4, the incident angles θ of s1 and s2 are randomly generated. i ,i=1,2, satisfying |θ1-θ2|<θ M ,θ M is the main lobe width of the array. Generate S=[s1,s2] T The steering vector matrix A = [a(θ1), a(θ2)], and finally the mixed signal is generated using the following formula.

[0010] X=AS#

[0011] (3) Pass the array received signal through an additive white Gaussian noise channel with a signal-to-noise ratio of -5 to 9 dB. Repeat the above process to generate 10,000 mixed signals for each signal-to-noise ratio.

[0012] Step 2: Build a multi-channel AIS signal separation network:

[0013] Construct a multi-channel AIS signal separation network (Multi-channel AIS Signal Separation), as shown in the attached Figure 3 As shown in the figure, the array receives AIS signals and separates the desired AIS incident signal from the interference signal through a multi-channel signal separation network. The observed mixed signal is processed by a convolutional neural network to find independent signal components. In the presence of mainlobe interference, the signal separation network can separate the interference signal from the desired signal, thereby reducing the impact of interference.

[0014] The entire separation network consists of an encoder, separator, and decoder architecture. The separator is the core of the network and consists of stacked convolutional modules. The separator generates an unmixing mask, which is combined with the encoder's output features as input to the decoder. The encoder contains a convolutional layer that extracts features from the input mixed signal. The decoder combines the unmixing mask and the encoder's output features to generate the deinterleaved signal.

[0015] (1) Encoder

[0016] The encoder is a complex convolution layer in the neural network. The encoder's function is to use one-dimensional convolution operation to obtain the encoding features of each source signal. Let the input signal x∈C N×L After the one-dimensional convolution operation, the input signal is converted into s encoder ∈C N×W :

[0017] sencoder =xU H

[0018] where U∈C W×L Contains W vectors, each of length L.

[0019] (2) Separator

[0020] Multi-channel AIS separation module f separator The separator is an important component of the present invention. It inputs the encoded signal x and outputs the separation mask for each independent component. The multi-channel AIS signal separation network actually contains N single-channel AIS separation modules. The separator inputs the encoded feature x and outputs the demixing mask for each signal, as shown in the following formula:

[0021] s mask =f separator (s encoder )

[0022] s mask ∈C K×W is the time domain separation mask, K is the number of independent components, which can measure the amplitude of each independent signal component in the mixed signal feature. In order to achieve the separation of the source signal, the mask is applied to the coding feature to obtain the characteristics of each independent component:

[0023] d i =m i ⊙s encoder ,i=1,2,…K

[0024] ⊙ represents the Hadamard product, d i As the input of the decoder, m i is the mask of the i-th signal.

[0025] The separator distributes the dilated convolution in a structure called a dilated convolution unit. The dilated convolution unit is composed of a convolution module, which includes a one-dimensional convolution, a complex layer normalization and a residual structure. A dilated convolution unit contains multiple of the above structures and is used to replace the recurrent neural network. In the dilated convolution, the core idea is to introduce a hole (Dilation) in the convolution kernel to change the size of the receptive field. The convolution kernel weights of the dilated convolution are not directly adjacent, but the coverage of the convolution kernel is expanded by inserting zero values at fixed intervals in the convolution kernel. By inserting zero values, AIS signals with longer time spans can participate in the convolution operation. The range of the receptive field is controlled by the parameter dilation rate. The introduction of holes increases the receptive field of the network without changing the parameter amount, thereby enhancing the nonlinear characteristics of the network. The present invention uses a different dilation coefficient for each dilated convolution unit to ensure that different convolution layers have different receptive fields. Appendix Figure 4 is a schematic diagram of the dilated convolution unit, Figure 5Schematic diagram of the convolution module.

[0026] The separator calculates the separation mask for the N signals, and then the decoder reconstructs the signal. To match the number of receiver array elements, the separator includes N parallel dilated convolution units to separate the signal for each channel. The steps of the separator are as follows:

[0027] 1: First, perform complex layer normalization on the features of the N array elements receiving signals of the encoder to ensure that the mean of each channel is close to 0 and the variance is close to 1. The features of the real and imaginary parts of the satellite-borne AIS receiving signal are consistent, and conventional normalization cannot be used. Otherwise, the feature distribution of the real and imaginary parts cannot be guaranteed to be consistent. The present invention uses complex layer normalization to ensure that the real and imaginary components are normal standard complex distributions with equal variances. The role of complex layer normalization is to pull the distribution of the input value of any neuron in each layer of the network back to the standard normal distribution with a mean of 0 and a variance of 1, so that the activation input value falls in the area where the activation function is more sensitive to the input, effectively improving the stability of the network training and enhancing the generalization ability of the network. The schematic diagram of complex layer normalization is shown in the attached figure. Figure 5 As shown. Complex layer normalization normalizes the tensor in the channel dimension and time dimension:

[0028]

[0029]

[0030]

[0031]

[0032] where Z∈C Q×T The input features for each layer are X and Y, which are the real and imaginary parts of the features, respectively. α and β are trainable parameters representing the scaling and translation of the features. They are then subjected to a complex one-dimensional convolution operation, which aims to adapt the input dimensions of the stacked convolutional units.

[0033] 2: Divide the features output in step 1 into N parts, and the N sub-features correspond one-to-one to the array elements.

[0034] 3: N sub-features are passed through stacked dilated convolution units. The structure of the dilated convolution unit is as shown in the attached figure. Figure 4 The present invention uses 4 stacked dilated convolution units to form a separator, each of which contains 6 Figure 5The structure shown in the figure has convolution expansion rates of 1, 2, 4, 8, 16, and 32. The output of the previous convolution unit is used as the input of the next convolution unit. Using repeated connections to form a deep separator network can not only enhance the ability to extract deep features of the source signal, but also effectively fuse the extracted multi-layer features. In order to avoid the gradient vanishing problem caused by too deep a network layer, each convolution module includes a channel attention mechanism. The input of each convolution module is a tensor of N×B×W. The data flow is divided into a main path and a residual path. The main path passes through the residual path. Figure 5 The output dimension after the operation shown is still N×B×W, and the residual path passes through the attached Figure 6 The output dimension after the channel attention mechanism shown is N×B×W, and then the output of the main path and the residual path are added element by element as the output of the convolution module.

[0035] The residual path includes a channel attention mechanism, which redistributes weights according to the importance of the features of interest, thereby highlighting certain important features. The present invention adds an attention mechanism to the blind separation convolutional network in the channel dimension of each convolution block to implement an attention mechanism for each channel feature and automatically learn the weight of the channel feature. The process of the channel attention mechanism is shown in the attached figure. Figure 6 shown.

[0036] The N output features are concatenated and then subjected to a complex one-dimensional convolution operation and a ReLU nonlinear function to map the time domain separation mask of the N-way mixed signal.

[0037] d=s mask ⊙s encoder #

[0038] (3) Decoder

[0039] The decoder's role is to restore the source signal waveform. The decoder uses a complex one-dimensional convolution operation to represent the reconstruction of the source signal:

[0040]

[0041] where V∈C L×W Contains L vectors, each of length W.

[0042] (4) Steering vector estimation module

[0043] Since the AIS received signal has mainlobe interference, conventional beamforming technology cannot suppress the mainlobe interference. After separating the desired signal from the interference, beamforming is performed on the desired signal. The present invention utilizes the characteristics of the AIS signal containing training sequences and start flags to construct array steering vectors based on matched filtering technology. Assuming that the AIS receiver is a linear receiving array and that the satellite transmits a signal S(t), the received signal X(t) can be described as:

[0044] X(t)=AS(t)

[0045] where A∈C N×K is the steering vector matrix, the number of array elements is N, and the number of received signals is K. i , assuming that the training sequence part of the satellite transmission signal is s(t), then the array element p i The response to receiving s(t) is:

[0046]

[0047] where h i (t) is the array element p i The impulse response of “*” indicates the convolution operation, τ i Indicates that the signal reaches array element p i The delay, e i (t) represents additive noise. If the incident signal is a narrowband signal, then It can be expressed as:

[0048]

[0049] where f c is the carrier frequency, H i (2πf) is the array element p i The transfer function of .

[0050] The pilot signal s(t) is generated by the local receiver and is combined with Do related operations:

[0051]

[0052]

[0053] Let τ = 0, then:

[0054]

[0055] Where is the energy of the P signal s(t), In turn against Yuan p i ,i=1,2,…,N

[0056] Do the above operation to get the vector

[0057]

[0058] Assume h i (t)=δ(t),n i (t) is negligible, then

[0059]

[0060] matrix Contains the time delay information of the incident signal, and then derives the array steering vector

[0061]

[0062] Wherein θ is the incident angle. As discussed above, the present invention utilizes the prior information of the training sequence and obtains the steering vector of the incident signal through correlation operation without estimating the incident angle.

[0063] (5) MVDR beamformer

[0064] After obtaining the array's steering vector, this algorithm uses MVDR beamforming to generate a strong gain for the desired signal. Let the input signal be x(t), and first calculate the covariance matrix R of the input signal:

[0065] R=E[x(t)x H (t)]

[0066] Where E[·] is the mathematical expectation, in practice we can use the sample covariance matrix Instead of R

[0067]

[0068] Then calculate the optimal beamforming weight vector w opt :

[0069]

[0070] Finally, the separated expected signal is obtained:

[0071]

[0072] Step 3: Use the training data constructed by S1 to train the multi-channel AIS signal separation network constructed by S2 to obtain the trained multi-channel AIS signal separation network:

[0073] The loss function of multi-channel AIS signal separation network training consists of multiple parts. The loss value includes the sum of the loss of the separated signals of N array elements. Define the array element p i The received signal is x i (t), then the array element p i The corresponding separation network labels are:

[0074] x target,i (t) = λ i x i (t)#

[0075] where λ i ∈C 1×1is a constant and the array element p i It is related to the location of the array element. i The time delay of receiving the target signal is τ i , the steering vector of the array is as follows:

[0076]

[0077] The loss value of the entire network is:

[0078] loss=loss1+loss2+…+loss N

[0079] Among them, loss i ,i=1,2,…,N is the loss of each element single channel separation network, loss i is defined as follows:

[0080]

[0081] e=xx aim

[0082]

[0083] Step 4: Generate a mixed signal from the acquired AIS signal and interference signal according to the method in S1 and input it into the trained multi-channel AIS signal separation network to obtain a separated signal.

[0084] The beneficial effects of the present invention are:

[0085] This paper proposes a deinterleaving method for satellite-borne AIS signals, addressing the insufficient anti-interference capabilities of traditional deinterleaving algorithms. Based on a neural network algorithm and beamforming, this method online separates AIS received signals from interference backgrounds, addressing the requirements for low signal-to-noise ratio (SNR) and interference-prone satellite-borne AIS signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is a schematic diagram of the overall architecture of the AIS signal deinterleaving method of the present invention;

[0087] Figure 2 This is the flow chart of satellite-borne AIS signal modulation;

[0088] Figure 3 This is a schematic diagram of a multi-channel signal separation network;

[0089] Figure 4 Schematic diagram of a dilated convolution unit;

[0090] Figure 5 Schematic diagram of the convolution module;

[0091] Figure 6 Schematic diagram of the attention mechanism;

[0092] Figure 7 The following is the bit error rate curve of the AIS separated signal under different signal-to-noise ratios when the main lobe interference is 10dB;

[0093] Figure 8 The following is the bit error rate curve of the AIS separated signal under different signal-to-noise ratios when the main lobe interference is 20dB;

[0094] Figure 9 Figure 2 is the bit error rate curve of the AIS separated signal under different interference-to-signal ratios. DETAILED DESCRIPTION

[0095] The present invention is described in detail below with reference to the accompanying drawings and examples.

[0096] like Figure 1 As shown in the figure, the process of the present invention is as follows: first, the received AIS signal sequence is processed as input data through maximum and minimum normalization; the target signal and the interference signal are separated by a multi-channel AIS signal separation neural network; the data is input into the network, and the weights and biases of the network neurons are continuously iterated to finally learn the signal separation strategy.

[0097] Example:

[0098] Step 1: Construct training data

[0099] First, generate the target desired signal and interference signal. The desired signal is the AIS baseband receive signal, with a data transmission bit rate of 9600 bit / s, a fixed 24-bit training sequence, and an 8-bit start flag. The modulation scheme is GMSK, with a modulation index of 0.5, a Gaussian filter time-bandwidth product (BT) of 0.4, an oversampling factor of 8, and NRZI encoding of the bit data. The interference signal is a signal using another modulation scheme, GMSK in this example. The number of receiving elements is 4. An AIS signal S1 and an interference signal S2, each approximately 26.67 milliseconds long, are randomly generated, along with a 4×2 matrix A. The mixed signal is generated using the following formula:

[0100] X=A[S1,S2] T

[0101] Step 2: Build a multi-channel AIS signal separation network

[0102] The network structure is as follows Figure 2As shown in the figure, the system consists of three structures: an encoder, a decoder, and a separator. The encoder obtains the encoded features of each source signal through complex one-dimensional convolution operations. The separator obtains the separation mask of the source signal through stacked dilated convolution units. The separator incorporates a channel attention mechanism. Channel attention captures local cross-channel information through a local cross-channel interaction strategy, allowing the separator to learn important information in different channels of the feature while discarding unimportant information. The decoder maps the encoded features and separation mask into an estimate of the desired signal through a one-dimensional linear connection layer. The decoder outputs the separated target AIS signal and interference signal.

[0103] Step 3: MVDR beamforming

[0104] First, the beamforming steering vector is calculated using the separated target AIS signal and the local reference signal. Then the covariance matrix of the source mixed signal is calculated. Finally, the MVDR algorithm is called to enhance the estimated desired signal.

[0105] The GPU used in this experiment is RTX3080Ti, and the network is trained for 50 rounds. The network prediction results are shown in the attached Figure 6 , Attachment Figure 7 and attached Figure 8 As shown, the decoding error rate of the separated signal is lower than that of other blind separation algorithms at 0-9 dB, and the present invention can meet the requirements of practical applications.

Claims

1. A method for deinterleaving satellite-borne AIS signals based on a neural network, characterized in that: The following steps are involved: S1. Build training data: Generate an AIS signal s1: Randomly generate a 208-bit, 26.67ms AIS message stream, which contains 168 bits of data, a 24-bit training sequence, and an 8-bit start flag. Define the data transmission bit rate as 9600 bit / s, the modulation method as GMSK, the modulation index as 0.5, the Gaussian filter time-bandwidth product BT = 0.4, the oversampling factor as 8, and the bit data as NRZI encoding. Generate an interference signal s2 from another communication system, the interference signal modulation method is still GMSK modulation, and the real AIS signal is saved as the label for supervised training. Define the number of receiving array elements as 4, and randomly generate the incident angles θ of s1 and s2 i ,i=1,2, satisfying |θ1-θ2|<θ M ,θ M is the main lobe width of the array; generate S = [s1, s2] T The steering vector matrix A=[a(θ1),a(θ2)], and finally the mixed signal is generated using the following formula: X=AS Pass the array received signal through an additive white Gaussian noise channel with a signal-to-noise ratio of -5 to 9 decibels; repeat the process to generate 10,000 mixed signals for each signal-to-noise ratio; S2. Build a multi-channel AIS signal separation network: Construct a multi-channel AIS signal separation network, including an encoder, separator, and decoder. The separator generates a demixing mask, which is used as input to the decoder along with the features of the encoder output. The encoder extracts features from the input mixed signal, while the decoder combines the demixing mask and the features of the encoder output to generate a deinterleaved signal. The encoder uses a one-dimensional convolution operation to obtain the encoding features of each source signal; the input signal is defined as x∈C N×L , then after one-dimensional convolution operation, the input signal is converted to s encoder ∈C N×W : s encoder =xU H where U∈C W×L Contains W vectors, each of length L; The separator inputs the encoded signal x and outputs a separation mask for each independent component. The multi-channel AIS signal separation network includes N single-channel AIS separation modules. The separator inputs the encoded feature x and outputs a demixing mask for each signal: s mask =f seperator (s encoder ) s mask ∈C K×W is the time domain separation mask, K is the number of independent components, which is used to measure the amplitude of each independent signal component in the mixed signal feature, and f seperator As a separator, in order to separate the source signal, the mask is applied to the encoding feature to obtain the features of each independent component: d i =m i ⊙s enscoder ,i=1,2,…K ⊙ represents the Hadamard product, d i As the input of the decoder, m i is the mask of the i-th signal; The separator distributes the dilated convolutions in a structure defined as a dilated convolution unit. The dilated convolution unit consists of multiple convolution modules, which include one-dimensional convolution, complex layer normalization, and residual structure. By using different dilation coefficients for each dilated convolution unit, different convolution layers have different receptive fields. The separator calculates the separation mask for N signals, and the decoder reconstructs the signal. To match the number of receiver array elements, the separator includes N parallel dilated convolution units for separating the signals of each channel. The separator steps are as follows: 1) Perform complex layer normalization on the features of the signals received by the N elements of the encoder to ensure that the mean of each channel is close to 0 and the variance is close to 1. The complex layer normalization is used to pull the distribution of the input value of any neuron in each layer of the network back to a standard normal distribution with a mean of 0 and a variance of 1. The complex layer normalization normalizes the tensor in the channel dimension and the time dimension: where Z∈C Q×T Input features to each layer, where X and Y are the real and imaginary parts of the features respectively, and α and β are trainable parameters representing the scaling and translation of the features. Then, a complex one-dimensional convolution operation is performed to adapt to the input dimension of the stacked convolution unit. 2) Divide the output features into N parts, and the N sub-features correspond one-to-one to the array elements; 3) The N sub-features are passed through stacked dilated convolution units respectively, and a separator is constructed using 4 stacked dilated convolution units. Each convolution unit contains 6 convolution modules with convolution dilation rates of 1, 2, 4, 8, 16, and 32. The output of the previous convolution unit serves as the input of the next convolution unit. Each convolution module contains a channel attention mechanism. The input of each convolution module is an N×B×W tensor. The data flow is divided into a main path and a residual path. After the main path passes through the convolution module, the output dimension is still N×B×W. After the residual path passes through the channel attention mechanism, the output dimension is N×B×W. Then, the output of the main path and the residual path are element-by-element added as the output of the convolution module. The N output features are concatenated and then subjected to a complex one-dimensional convolution operation and a ReLU nonlinear function to map the time domain separation mask of the N-way mixed signal: d=s mask ⊙s encoder The decoder uses a complex one-dimensional convolution operation to represent the reconstruction of the source signal: where V∈C L×W Contains L vectors, each of which has a length of W; After separating the desired signal from the interference, beamforming is performed on the desired signal. The steps are as follows: Using the characteristics of the AIS signal, which contains a training sequence and a start flag, an array steering vector is constructed based on matched filtering technology. The AIS receiver is defined as a linear receiving array. The satellite transmits the signal S(t), and the received signal X(t) is described as: X(t)=AS(t) where A∈C N×K is the steering vector matrix, the number of array elements is N, and the number of received signals is K; for a single array element p i , define the training sequence part of the satellite transmission signal as s(t), array element p i The response to receiving s(t) is: where h i (t) is the array element p i The impulse response of "*" indicates the convolution operation, τ i Indicates that the signal reaches array element p i The delay, e i (t) represents additive noise; If the incident signal is a narrowband signal, then Expressed as: where f c is the carrier frequency, H i (2πf) is the array element p i The transfer function of The pilot signal s(t) is generated by the local receiver and is combined with Do related operations: Let τ = 0, then: Where is the energy of the P signal s(t), In turn against Yuan p i ,i=1,2,…,N, Get vector Let h i (t) = δ(t), n i (t) is negligible, then matrix Contains the time delay information of the incident signal and derives the array steering vector: Where θ is the angle of incidence; After obtaining the steering vector of the array, MVDR beamforming is used to form a strong gain for the desired signal. The input signal is defined as x(t) and the covariance matrix R of the input signal is calculated as: R=E[x(t)x H (t)] Where E[·] is the mathematical expectation, using the sample covariance matrix Instead of R: Calculate the optimal beamforming weight vector w opt : Finally, the separated expected signal is obtained: S3, uses the training data constructed in S1 to train the multi-channel AIS signal separation network constructed in S2, and obtains the trained multi-channel AIS signal separation network: The multi-channel AIS signal separation network training loss function includes multiple parts, and the loss value includes the sum of the losses of the separated signals of N array elements; define the array element p i The received signal is x i (t), then the array element p i The corresponding separation network labels are: x target,i (t)=λ i x i (t)# where λ i ∈C 1×1 is a constant and the array element p i The position of the array element p is related to the definition of i The time delay of receiving the target signal is τ i , the steering vector of the array is as follows: The loss value of the entire network is: loss=loss1+loss2+…+loss N Among them, loss i ,i=1,2,…,N is the loss of each element single channel separation network, loss i is defined as follows: e=x-x aim S4: Generate a mixed signal from the acquired AIS signal and the interference signal according to the method of S1 and input it into the trained multi-channel AIS signal separation network to obtain a separated signal.

Images