Unmanned ship intelligence-oriented AIS weak signal blind separation robustness method and system

Abstract

The invention discloses an AIS weak signal blind separation robustness method for an unmanned ship, and the method comprises the following steps: 1) constructing a random signal matrix set, and generating an AIS data set simulating a real scene through random number seeds, signal source dynamic change and signal length adjustment; 2) designing a generator network fusing one-dimensional convolution and an LSTM time sequence module, extracting a space-time joint feature of the mixed signal and reconstructing a target signal; 3) establishing a multi-constraint optimization mechanism, and dynamically adjusting generator parameters by combining mean square error loss, interference suppression loss, statistical correlation loss and adversarial loss; 4) constructing a generative adversarial framework, and adopting a multi-layer perceptron discriminator to realize signal authenticity identification through cross entropy loss to form a parameter game mechanism; and 5) under a 5dB low signal-to-noise ratio condition, verifying the robust separation performance of the model in a complex multipath environment. According to the method, through adversarial training and multi-dimensional feature fusion, the blind separation precision of weak AIS signals under strong noise interference is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of maritime signal processing technology, and in particular to a robust blind separation method and system for AIS weak signals for intelligent unmanned boats. Background Art

[0002] Currently, unmanned vessels are developing towards intelligentization and playing an important role in fields such as ocean exploration and environmental monitoring. They rely on technologies such as multi-source sensor fusion, target detection, and path planning to achieve autonomous navigation and intelligent decision-making. Therefore, this places higher demands on the navigation efficiency, decision-making accuracy, and obstacle avoidance capabilities of unmanned vessels.

[0003] Commonly used sensing devices on ships include radar, ultrasonic sensors, optical cameras, and the Automatic Identification System (AIS). Ultrasonic sensors are susceptible to severe weather conditions, leading to target detection errors and high maintenance costs. Optical cameras are easily obscured in complex environments and experience performance degradation in extreme weather conditions, making these sensing devices prone to identification errors in extreme weather conditions. Radar and AIS, as important ship sensing systems, both rely on electromagnetic waves for radio communication and have the advantage of strong anti-interference capabilities. Compared to radar, AIS has lower operating power consumption, longer-range coverage, and provides more navigation information, making it indispensable for ensuring the navigation of unmanned ships.

[0004] However, because AIS uses self-organizing time-division multiple access (TDMA) communication, its signals are prone to time-slot collisions in areas with high traffic density, preventing the receiver from effectively resolving the target vessel's position, increasing navigation safety risks. Therefore, how to efficiently and accurately extract the original signal from the mixed AIS signals during autonomous navigation by unmanned vessels has become a pressing issue.

[0005] Traditional blind source separation (BSS) frameworks rely on the non-Gaussian properties of signals for signal separation. Their applicability presupposes that the source signal has a non-Gaussian distribution, while the mixed signal satisfies a Gaussian distribution. However, AIS signals undergo Gaussian filtering during the modulation process, which weakens the signal's higher-order statistical properties. This results in low accuracy and poor stability in traditional independent component analysis methods for AIS signal separation. Summary of the Invention

[0006] Aiming at the time slot collision problem of AIS signals, the present invention proposes a novel separation framework based on neural network to solve the above problem, aiming to provide a robust blind separation method and system for AIS weak signals for intelligent unmanned boats.

[0007] The present invention provides a robust blind separation method for AIS weak signals for intelligent unmanned boats, which has the following characteristics: Step 1, constructing a random signal matrix set: by introducing different random number seeds, signal source changes and signal lengths, a random AIS signal data set is generated to simulate the signal distribution in a real scene; Step 2, generator multi-dimensional feature extraction: based on the above random signal matrix set, a generator module including a one-dimensional convolution module and a temporal feature extraction module is used to extract the spatial features and time series dependencies of the mixed signal to generate a reconstructed signal; Step 3, adaptive constraint mechanism fusion : Based on the above feature extraction results, the generator module is jointly constrained by mean square error loss, interference suppression loss, statistical correlation loss and adversarial loss to optimize the generator training process; Step 4, construct a "zero-sum game" mechanism: At the same time, the adversarial training framework of the discriminator and the generator is adopted. The discriminator classifies the real signal and the generated signal through a multi-layer fully connected network, and updates the parameters through the cross-entropy loss function; Step 5, performance verification in low signal-to-noise ratio scenarios: Under 5dB signal-to-noise ratio conditions, the separation performance of the model in a complex channel environment is evaluated by the mean square error and signal-to-interference ratio indicators to verify the robustness of the method.

[0008] The robust blind separation method for AIS weak signals for intelligent unmanned watercraft provided by the present invention may also have the following features: wherein, in step 1, the process of constructing the random signal matrix set is as follows: step 1-1, generating a baseband signal, where the baseband signal is composed of a binary data sequence randomly generated by a pseudo-random number generator; step 1-2, encoding and modulating the data by Gaussian minimum shift keying modulation to generate an AIS signal; step 1-3, adding additive white noise with signal-to-noise ratio control to simulate a multi-channel mixed signal; and step 1-4, flattening the generated multi-channel mixed signal matrix into a vector and concatenating it with the random noise to serve as the input of the generator module.

[0009] In the robust blind separation method for AIS weak signals for intelligent unmanned boats provided by the present invention, it can also have the following characteristics: wherein, in step 2, the specific implementation process of the generator multi-dimensional feature extraction is: step 2-1, the combined signal is input into a four-layer one-dimensional convolution module, and the local time series features are extracted through the LeakyReLU activation function; step 2-2, the high-dimensional features output by the convolution are transposed and input into a two-layer long short-term memory network to capture the long-term time dependency of the signal and output the global time series features; step 2-3, the high-dimensional features output by the LSTM are mapped to a dimension matching the real source signal through a two-layer fully connected network, and the output range is constrained using the Tanh activation function; step 2-4, the generator generates a "false" signal based on the information extracted from the multi-dimensional features, and inputs the discriminator to complete the adversarial training.

[0010] The robust method for blind separation of weak AIS signals for intelligent unmanned boats provided by the present invention may also have the following features: in step 3, a multi-dimensional constraint mechanism is set in the generator part, and the mean square error loss, interference suppression loss, statistical correlation loss and adversarial loss are introduced to jointly constrain the generator module, and the total loss is back-propagated to update the training parameters of the generator. The introduction of multi-dimensional constraints can optimize the generator performance from multiple angles to achieve high-quality blind source separation of AIS signals.

[0011] The robust blind separation method for weak AIS signals for intelligent unmanned watercraft provided by the present invention may also have the following features: In step 4, the "zero-sum game" structure is primarily based on the game mechanism between the discriminator and the generator. After the true and false signals in the discriminator are input into the discriminator, they are gradually reduced in dimension through a three-layer fully connected network. The first fully connected network maps the input signal to a 256-dimensional feature space, which is then compressed to 128 dimensions through the second fully connected network. The third fully connected layer outputs a probability, indicating the confidence that the signal is a true signal. Since the AIS signal is GMSK modulated, the present invention uses the AIS modulated GMSK signal as the true signal, and uses it together with the false signal generated by the generator as the input to the discriminator.

[0012] The robust blind separation method for weak AIS signals for intelligent unmanned boats provided by the present invention may also have the following features: wherein, the specific process of step 1-1 is as follows: after the baseband signal undergoes HDLC bit stuffing, synchronization sequence and buffer addition, and NRZI encoding, it is passed through a Gaussian low-pass filter to impart Gaussian characteristics. The time domain impulse response of the Gaussian low-pass filter is:

[0013]

[0014] Among them, σ represents the standard deviation of the Gaussian function, which determines the extent of the impulse response;

[0015] The specific process of steps 1-2 is: the signal after passing through the Gaussian low-pass filter is modulated using MSK to generate the AIS signal. The MSK signal can be expressed as:

[0016] S(t)=cos(2πf c t+φ(t)) (2)

[0017] where f c is the carrier frequency, φ(t) is the instantaneous phase of the signal;

[0018] The specific process of steps 1-3 is as follows: the carrier frequency band of the AIS signal modulation is 156.025 MHz to 162.025 MHz, and the present invention sets the carrier frequency to 156.775 MHz. Therefore, the GMSK signal can be expressed as:

[0019]

[0020] Among them, the modulation index h is used to characterize the intensity of signal modulation, b(t) represents the baseband signal, and the pulse shaping function h g (t) is used to filter the baseband signal to limit the signal bandwidth and control its signal characteristics.

[0021] The specific process of steps 1-4 is: the mixed signal matrix S (S∈R (N*N) ) and d Z dimensional random noise Z(Z∈R (dz) ) splicing, flatten S into a vector vec(S)∈R (N*N) Then concatenate it with random noise Z to get the combined input:

[0022]

[0023] The robustness method for blind separation of weak signals of AIS for intelligent unmanned boats provided by the present invention may also have the following features: wherein the specific process of step 2-1 is: expanding the combined signal I into a three-dimensional length quantity I 3D Input spatial feature extraction module, which consists of four layers of 1D convolutional neural network. The convolution calculation formula of the first layer is:

[0024] X (l) =LeakyReLU(Conv1D (l) (X (l-1) )) (5)

[0025] The output size of the convolution operation is Among them C l is the number of channels, T (l) is the time series length, at this time The first layer of convolutional neural network sets the number of input channels to 1, the number of output channels to 16, and the convolution kernel size to 1*3 to extract basic low-level local temporal features. The next three layers use the number of input channels to 16, 32, and 64, and the number of output channels to 32, 64, and 128, respectively. The convolution kernel size is 1*3. The last three convolutional layers will gradually improve the feature abstraction level and capture higher-level signal patterns. LeakyReLU is used as the activation function after each convolution layer. The calculation process of multi-dimensional feature extraction is as follows:

[0026]

[0027] X (l) =LeakyReLU(Conv1D (l) (X (l-1) )) (7)

[0028]

[0029] h t ,c t =LSTM(X LSTM-in ,h t-1 ,c t-1 ) (9)

[0030]

[0031] X output =Tanh(ω2·LeakyReLU(ω1h final +b1)+b2) (11)

[0032] Among them, S represents the combined signal matrix, Z represents random noise, I represents the mixed signal splicing noise matrix, and the output size of the convolution operation is where h t ∈R 256 is the hidden state to capture the temporal features, c t is the cell state at the current moment, h t-1 is the hidden state at the previous moment, c t-1 is the cell state at the previous moment, ω1 and b1 are the weight matrix and bias term of the first fully connected layer, ω2 and b2 are the weight matrix and bias term of the second fully connected layer, X output is the output of the temporal feature extraction module, and Tanh and LeakyReLU are activation functions. The specific process of step 2-2 is: after feature extraction, in order to adapt to the temporal feature extraction module, X CNN Transpose into dimension N×T×128, that is, X (LSTM-in) =X CNN T ∈R N×T×128 , X (LSTM-in) The input hidden unit is a time series feature extraction module with 256 hidden units. The time series feature extraction module is used to accept the high-dimensional features extracted by the convolution layer and use its memory mechanism to capture long-term dependencies. The two-layer stacked long short-term memory network provides a relatively powerful time modeling capability. The time series data after processing by the long short-term memory network is:

[0033] h t ,c t =LSTM(X LSTM-in ,h t-1 ,ct-1 ) (12)

[0034] where h t ∈R 256 is the hidden state to capture the temporal features, c t is the cell state at the current moment, h t-1 is the hidden state at the previous moment, c t-1 is the cell state at the previous moment. The output of the LSTM layer is the hidden state h at the last time step final =h T ∈R N ×256 Represents the global characteristics of the sequence as a whole, where h T is the hidden state at the last time step. The specific process of steps 2-3 is as follows: the output of the time series feature extraction network passes through two layers of fully connected networks. The first layer represents the high-dimensional feature map, and the second layer outputs signal data that matches the dimension of the real source signal. Each layer of the fully connected network uses Tanh as an activation function to limit the output to the range of [-1, 1] to simulate the characteristics of the real signal. The Tanh function formula is:

[0035]

[0036] The Tanh activation function can maintain the normalization of the generated signal amplitude, making it easier to compare with the actual target signal and calculate the loss. final After the full connection layer mapping is:

[0037] X output =Tanh(ω2·LeakyReLU(ω1h final +b1)+b2) (14)

[0038] Where ω1 and b1 are the weight matrix and bias term of the first fully connected layer, respectively, and ω2 and b2 are the weight matrix and bias term of the second fully connected layer, respectively.

[0039] The robustness method for blind separation of weak AIS signals for intelligent unmanned boats provided by the present invention may also have the following features: wherein the specific process of step 3 is as follows: step 3-1, introducing mean square error loss on the basis of adversarial loss:

[0040]

[0041]

[0042] Among them, L adv is the generator adversarial loss, G(z) represents the sample generated by the generator G after receiving z, D(G(z)) is the judgment result of the discriminator D on G(z), and L MSE is the mean square error loss, yi is a GMSK signal; in step 3-2, the interference suppression loss is introduced as:

[0043]

[0044] Among them, L SIR For interference suppression loss, ||·||2 represents the Euclidean norm, and ò is set to a small constant to prevent division by zero errors; Step 3-3, introduce statistical correlation loss as follows:

[0045]

[0046] Among them, L cMI is the statistical correlation loss, L G is the total loss of the generator network, z is the noise distribution, y i is the true value signal, p(y i ,G(z)) is y i The joint probability density function with G(z), p(y i ) and p(G(z)) are their respective marginal probability density functions. The numerical scales of different loss terms in the loss function may be different. In steps 3-4, the optimal weight parameters are set for each loss term to meet the multi-objective task requirements and ensure the stability of training. The total loss of the generator network is calculated as follows:

[0047] L G =λ1·L MSE +λ2·L SIR +λ3·L cMI +λ4·L adv (19)

[0048] In step 3-5, the difference between the "false" signal generated by the generator and the true value is measured by the generator loss, forcing the generator to generate a "false" signal that the discriminator cannot "distinguish"; in step 3-6, the gradient descent algorithm is used to optimize the generator loss, and steps 3-5 and 3-6 are repeated until the discriminator cannot judge the authenticity of the generated signal.

[0049] In the robust method for blind separation of weak signals of AIS for intelligent unmanned boats provided by the present invention, it can also have the following features: wherein, the specific process of step 4 is: step 4-1, after the signal is input to the discriminator, it passes through a three-layer fully connected network to gradually reduce the dimension, and the first layer of the fully connected network maps the input to a 256-dimensional space; step 4-2, the second layer further maps the input to 128 dimensions; step 4-3, the third layer outputs a single probability value to represent the confidence that the signal is true, and at the same time uses the LeakyReLU function to provide nonlinear representation capabilities, and uses the Sigmoid activation function to output a probability value of [0,1] for binary classification. In the above process, the goal of the discriminator is to distinguish between real samples and generated samples as accurately as possible, so the cross entropy loss function is used to calculate the loss L of the real samples respectively. real and the loss L of the generated samples fake , and then add them together to get the total discriminator loss L D :

[0050] L D =L real +L fake (20)

[0051]

[0052] Among them, L real is the loss of the discriminator for real samples, L fake is the loss of the discriminator on the generated samples.

[0053] The present invention also provides an intelligent AIS weak signal blind separation system for unmanned boats, which has the following characteristics: a random signal matrix set construction module: used to generate a random AIS signal data set; a generator multi-dimensional feature extraction module: used to extract the spatial characteristics and time series dependencies of the mixed signal to generate a reconstructed signal; an adaptive constraint mechanism fusion module: used to optimize the training process of the generator; a "zero-sum game" mechanism construction module: used to constrain the output range; a performance verification module: used to evaluate the separation performance of the model in a complex channel environment and verify the robustness of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Flowchart of the robust blind separation method for AIS weak signals for intelligent unmanned boats in this embodiment;

[0055] Figure 2 AIS signal modulation generation flow chart in this embodiment;

[0056] Figure 3 Schematic diagram of multi-dimensional feature extraction in this embodiment;

[0057] Figure 4Schematic diagram of the fusion of the adaptive constraint mechanism in this embodiment;

[0058] Figure 5 Schematic diagram of the “zero-sum game” mechanism in this embodiment;

[0059] Figure 6 : is a comparison chart of the experimental mean square error results in this embodiment;

[0060] Figure 7 2 is a comparison chart of the experimental signal-to-interference ratio results in this embodiment. DETAILED DESCRIPTION

[0061] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0062] This example aims to provide a blind source separation method based on a time series generative adversarial network (TSeq-GAN). This method addresses the loss of high-order statistical properties of AIS signals during the modulation process, resulting from Gaussian filtering. First, this example establishes and uses a set of AIS random mixed signal matrices to replace the ordered training set. Second, in the network generator, the multidimensional feature capture and high-precision signal reconstruction of random signal data are achieved by integrating spatial and time series feature extraction networks with a generative-discriminative mechanism. Finally, this example optimizes the network loss function based on a global multi-objective optimization approach to fully balance the separation requirements of minimizing error and maximizing signal generation quality.

[0063] Figure 1 This is a flow chart of the robustness method for blind separation of AIS weak signals for intelligent unmanned boats in this embodiment.

[0064] like Figure 1 As shown, the specific implementation method of the AIS weak signal blind separation robustness method for the intelligent unmanned boat in this embodiment is as follows:

[0065] Step S1, constructing a random signal matrix set: The AIS signal dataset used in this embodiment is generated by a random process, aiming to replicate the inherent randomness of real-world AIS transmissions. To overcome the potential dataset bias introduced by a sequential training set that fails to simulate the data distribution in real-world scenarios, leading to inaccurate or unstable model performance under specific conditions, this embodiment generates a randomized AIS signal dataset by introducing different random number seeds, signal source variations, and signal lengths. This simulates the signal distribution in real-world scenarios and maintains the wide variation of AIS signals in real-world environments.

[0066] Figure 2 The following is a flow chart of AIS signal modulation generation in this embodiment.

[0067] The specific process of step S1 is:

[0068] Step S1-1, as Figure 2 As shown in Figure 1, the generated baseband signal consists of a binary data sequence randomly generated by a pseudo-random number generator. After HDLC bit stuffing, synchronization sequence and buffer addition, and NRZI encoding, the baseband signal is passed through a Gaussian low-pass filter to give it Gaussian characteristics. The time domain impulse response of the Gaussian low-pass filter is:

[0069]

[0070] Among them, σ represents the standard deviation of the Gaussian function, which determines the extent of the impulse response;

[0071] Step S1-2: Encode and modulate the data by Gaussian minimum shift keying modulation to generate an AIS signal.

[0072] The AIS signal can be generated by using MSK modulation on the signal after passing through the Gaussian low-pass filter. The MSK signal can be expressed as:

[0073] S(t)=cos(2πf c t+φ(t)) (2)

[0074] where f c is the carrier frequency, φ(t) is the instantaneous phase of the signal;

[0075] In step S1-3, in order to more accurately simulate the transmission environment, additive white noise with a noise level controlled by the signal-to-noise ratio is introduced into the data set. The mixing of signals between multiple channels is simulated by a random noise matrix to ensure the authenticity and diversity of the data set.

[0076] The carrier frequency range of the AIS signal modulation is 156.025 MHz to 162.025 MHz. In this embodiment, the carrier frequency is set to 156.775 MHz. Therefore, the GMSK signal can be expressed as:

[0077]

[0078] Among them, the modulation index h is used to characterize the intensity of signal modulation, b(t) represents the baseband signal, and the pulse shaping function h g (t) is used to filter the baseband signal to limit the signal bandwidth and control its signal characteristics.

[0079] Step S1-4: Flatten the generated multi-channel mixed signal matrix into a vector and concatenate it with random noise, which serves as the input to the generator module. Therefore, the data set used in this embodiment exhibits inherent randomness at each stage of signal generation and mixing, and the signal data in each simulation process is randomly generated, fully meeting the task requirements of blind source separation. The specific process is as follows:

[0080] The mixed signal matrix S(S∈R (N*N) ) and d Z dimensional random noise Splicing, flatten S into a vector vec(S)∈R (N*N) After concatenating with random noise Z, the combined input is:

[0081]

[0082] Figure 3 The following is a flow chart of AIS signal modulation generation in this embodiment.

[0083] Step S2, generator multi-dimensional feature extraction: Based on the above random signal matrix set, such as Figure 3 As shown, a generator module consisting of four one-dimensional convolutional modules and two temporal feature extraction modules is used to extract the spatial features and temporal dependencies of the mixed signal to generate a reconstructed signal. LeakyReLU is used as the activation function after each convolution layer to enhance the network's ability to handle complex, variable, and noisy signals. This also fully utilizes the local feature extraction capabilities of convolutional neural networks, the time series modeling capabilities of long short-term memory networks, and the adversarial training capabilities of generative adversarial networks, especially when dealing with highly nonlinear and multimodal signals.

[0084] At the same time, the local time series features of the mixed signal splicing noise matrix are gradually extracted and the nonlinear expression capability is enhanced. Next, the features are converted into dimensions suitable for the time series feature extraction layer. The two stacked time series feature extraction layers perform time series modeling to capture long-term dependencies. After two layers of linear connection layers, the output signal data that matches the real source signal is mapped. Finally, a "false" signal is generated based on the data features extracted by the generator. The specific process is as follows:

[0085] In step S2-1, the combined signal is input into a four-layer one-dimensional convolution module, and the local temporal features are extracted through the LeakyReLU activation function.

[0086] Expand the combined signal I into a three-dimensional length quantity I 3D Input spatial feature extraction module, which consists of four layers of 1D convolutional neural network. The convolution calculation formula of the first layer is:

[0087] X (l)=LeakyReLU(Conv1D (l) (X (l-1) )) (5)

[0088] The output size of the convolution operation is Among them C l is the number of channels, T (l) is the time series length, at this time The first layer of convolutional neural network sets the number of input channels to 1, the number of output channels to 16, and the convolution kernel size to 1*3 to extract basic low-level local temporal features. The next three layers use the number of input channels to 16, 32, and 64, and the number of output channels to 32, 64, and 128, respectively. The convolution kernel size is 1*3. The last three convolutional layers will gradually improve the feature abstraction level and capture higher-level signal patterns. LeakyReLU is used as the activation function after each convolution layer. The calculation process of multi-dimensional feature extraction is as follows:

[0089]

[0090] X (l) =LeakyReLU(Conv1D (l) (X (l-1) )) (7)

[0091]

[0092] h t ,c t =LSTM(X LSTM-in ,h t-1 ,c t-1 ) (9)

[0093]

[0094] X output =Tanh(ω2·LeakyReLU(ω1h final +b1)+b2) (11)

[0095] Among them, S represents the combined signal matrix, Z represents random noise, I represents the mixed signal splicing noise matrix, and the output size of the convolution operation is where h t ∈R 256 is the hidden state to capture the temporal features, c t is the cell state at the current moment, h t-1 is the hidden state at the previous moment, c t-1 is the cell state at the previous moment, ω1 and b1 are the weight matrix and bias term of the first fully connected layer, ω2 and b2 are the weight matrix and bias term of the second fully connected layer, Xoutput is the output of the time series feature extraction module, and Tanh and LeakyReLU are activation functions.

[0096] In step S2-2, the high-dimensional features output by the convolution are transposed and input into a two-layer long short-term memory network to capture the long-term temporal dependency of the signal and output the global temporal features.

[0097] After feature extraction, in order to adapt to the temporal feature extraction module, X CNN Transpose into dimension N×T×128, that is, X (LSTM-in) =X CNN T ∈R N×T×128 , X (LSTM-in) The input hidden unit is a time series feature extraction module with 256 hidden units. The time series feature extraction module is used to accept the high-dimensional features extracted by the convolution layer and use its memory mechanism to capture long-term dependencies. The two-layer stacked long short-term memory network provides a relatively powerful time modeling capability. The time series data after processing by the long short-term memory network is:

[0098] h t ,c t =LSTM(X LSTM-in ,h t-1 ,c t-1 ) (12)

[0099] where h t ∈R 256 is the hidden state to capture the temporal features, c t is the cell state at the current moment, h t-1 is the hidden state at the previous moment, c t-1 is the cell state at the previous moment. The output of the LSTM layer is the hidden state h at the last time step final =h T ∈R N ×256 Represents the global characteristics of the sequence as a whole, where h T is the hidden state of the last time step;

[0100] In step S2-3, the high-dimensional features output by LSTM are mapped to the dimension matching the real source signal through a two-layer fully connected network, and the Tanh activation function is used to constrain the output range.

[0101] The output of the time series feature extraction network passes through two layers of fully connected networks. The first layer represents the high-dimensional feature map, and the second layer outputs signal data that matches the dimension of the real source signal. Each layer of the fully connected network uses Tanh as an activation function to limit the output to the range of [-1, 1] to simulate the characteristics of the real signal. The Tanh function formula is:

[0102]

[0103] The Tanh activation function can maintain the normalization of the generated signal amplitude, making it easier to compare with the actual target signal and calculate the loss. final After the full connection layer mapping is:

[0104] X output =Tanh(ω2·LeakyReLU(ω1h final +b1)+b2) (14)

[0105] Where ω1 and b1 are the weight matrix and bias term of the first fully connected layer, respectively, and ω2 and b2 are the weight matrix and bias term of the second fully connected layer, respectively.

[0106] In step S2-4, the generator generates a "false" signal based on the information extracted from the multi-dimensional features and inputs it into the discriminator to complete the adversarial training.

[0107] Figure 4 Schematic diagram of the fusion of the adaptive constraint mechanism in this embodiment;

[0108] Step S3, adaptive constraint mechanism fusion: This embodiment is to meet the multi-source perception needs of USV in scenarios with high traffic density and poor communication conditions, and avoid the network's excessive pursuit of a single goal, such as Figure 4 As shown, this embodiment adds a multi-dimensional constraint mechanism to the generator part. Based on the above feature extraction results, the generator module is jointly constrained by mean square error loss, interference suppression loss, statistical correlation loss and adversarial loss to optimize the training process of the generator.

[0109] In this embodiment, a multi-dimensional constraint mechanism is set up in the generator part. The mean square error loss, interference suppression loss, statistical correlation loss, and adversarial loss are introduced to jointly constrain the generator module. The total loss is back-propagated to update the generator training parameters. The introduction of multi-dimensional constraints can optimize the generator performance from multiple angles to achieve high-quality blind source separation of AIS signals. The specific process is as follows:

[0110] Step S3-1, introduce mean square error loss based on adversarial loss:

[0111]

[0112] Among them, L adv is the generator adversarial loss, G(z) represents the sample generated by the generator G after receiving z, D(G(z)) is the judgment result of the discriminator D on G(z), and L MSE is the mean square error loss, y i It is a GMSK signal.

[0113] Step S3-2, introduce interference suppression loss as:

[0114]

[0115] Among them, L SIR For interference rejection loss, ||·||2 represents the Euclidean norm, and ò is set to a small constant to prevent division by zero errors;

[0116] Step S3-3, introduce statistical correlation loss as:

[0117]

[0118] Among them, L cMI is the statistical correlation loss, L G is the total loss of the generator network, z is the noise distribution, y i is the true value signal, p(y i ,G(z)) is y i The joint probability density function with G(z), p(y i ) and p(G(z)) are their respective marginal probability density functions, and the numerical scales of different loss terms in the loss function may be different.

[0119] In step S3-4, the optimal weight parameters are set for each loss item to meet the multi-objective task requirements and ensure the stability of training. The total loss of the generator network is calculated as follows:

[0120] L G =λ1·L MSE +λ2·L SIR +λ3·L cMI +λ4·L adv (19)

[0121] In step S3-5, the difference between the "false" signal generated by the generator and the true value is measured by the generator loss, forcing the generator to generate a "false" signal that the discriminator cannot "distinguish".

[0122] In step S3-6, the gradient descent algorithm is used to optimize the generator loss, and steps S3-5 and S3-6 are repeated until the generated discriminator cannot determine the authenticity of the generated signal.

[0123] In this embodiment, it is found through experiments that the weight λ1 of the mean square error loss is 0.4, the weight λ2 of the interference suppression loss is 0.01, the weight λ3 of the statistical correlation loss is 0.5, and the weight λ4 of the adversarial loss is set to 0.05.

[0124] Step S4, constructing a "zero-sum game" mechanism: At the same time, an adversarial training framework between the discriminator and the generator is adopted. The discriminator performs binary classification on the real signal and the generated signal through a multi-layer fully connected network, and updates the parameters through the cross-entropy loss function.

[0125] Figure 5 Schematic diagram of the “zero-sum game” mechanism in this embodiment;

[0126] like Figure 5 As shown, the "zero-sum game" structure is mainly based on the game mechanism between the discriminator and the generator. After the true value signal and the false signal in the discriminator are input into the discriminator, the signal dimension is gradually reduced through a three-layer fully connected network. The first layer of the fully connected network maps the input signal to a 256-dimensional feature space, and then the dimension is compressed to 128 dimensions through the second layer of the fully connected network. The third layer of the fully connected layer outputs a probability, which indicates the confidence that the signal is a true signal. Since the AIS signal is GMSK modulated, the AIS modulated signal GMSK signal is used as the true signal in the present invention, and it is used as the input of the discriminator together with the false signal generated by the generator. The specific process is:

[0127] Step S4-1: After the signal is input into the discriminator, it passes through a three-layer fully connected network to gradually reduce the dimension. The first layer of the fully connected network maps the input to a 256-dimensional space.

[0128] Step S4-2, the second layer further maps the input to 128 dimensions;

[0129] In step S4-3, the third layer outputs a single probability value to indicate the confidence that the signal is true. At the same time, the LeakyReLU function is used to provide nonlinear representation capabilities, and the Sigmoid activation function is used to output a probability value of [0, 1] for binary classification.

[0130] In the above process, the goal of the discriminator is to distinguish the real samples from the generated samples as accurately as possible, so the cross entropy loss function is used to calculate the loss L of the real samples respectively. real and the loss L of the generated samples fake , and then add them together to get the total discriminator loss L D :

[0131] L D =L real +L fake (20)

[0132]

[0133] Among them, L real is the loss of the discriminator for real samples, L fake is the loss of the discriminator on the generated samples.

[0134] Step S5, performance verification in low signal-to-noise ratio scenario: Under a 5dB signal-to-noise ratio, the separation performance of the model in a complex channel environment is evaluated by the mean square error and signal-to-interference ratio indicators to verify the robustness of the method.

[0135] Figure 6 This is a comparison chart of the experimental mean square error results in this embodiment;

[0136] Figure 7 2 is a comparison chart of the experimental signal-to-interference ratio results in this embodiment.

[0137] like Figure 6 and Figure 7 As shown, this embodiment designs a low signal-to-noise ratio scenario to verify the model separation performance. The specific process is as follows:

[0138] 1. The model signal-to-noise ratio parameter is designed to be 5dB, and the number of source signal mixtures is adjusted to 4-10 for experiments. The mean square error and signal-to-interference ratio are used as evaluation indicators to measure the model performance, and the experimental results of the model proposed in this invention are recorded.

[0139] 2. For some existing blind source separation methods, experiments are conducted under the same experimental background, and the mean square error and signal-to-interference ratio after separation of each method under different numbers of source signal mixtures are recorded.

[0140] This embodiment also provides an AIS weak signal blind separation system for intelligent unmanned boats, including:

[0141] Random signal matrix set construction module: Generate a random AIS signal data set using the method in step S1 of this implementation.

[0142] Generator multi-dimensional feature extraction module: uses the method in step S2 of this implementation to extract the spatial features and time series dependencies of the mixed signal to generate a reconstructed signal.

[0143] Adaptive constraint mechanism fusion module: Optimize the generator training process using the method in step S3 of this implementation.

[0144] "Zero-sum game" mechanism building module: Use the method in step S3 of this implementation to constrain the output range.

[0145] Performance verification module: Use the method in step S4 of this implementation to evaluate the separation performance of the model in a complex channel environment and verify the robustness of the system.

[0146] Compared with the prior art, the robust method and system for blind separation of weak AIS signals for intelligent unmanned boats provided in this embodiment has the following beneficial effects:

[0147] 1. The present invention adopts a random AIS signal data set. By introducing different random number seeds, signal source changes, and signal lengths into the data set to replace the traditional network model ordered training set, the AIS signal training set is made variable. At the same time, dynamic test data spliced ​​by mixed signals and source signals is used in the signal test data set. This enables the AIS signal separation network to have stronger real-time processing capabilities and high fault tolerance when dealing with complex, unknown, and dynamic environments, significantly enhancing the versatility of the system.

[0148] 2. Based on the traditional generative-adversarial mechanism, the present invention introduces a temporal relationship modeling network and a spatial information processing structure to optimize the traditional fully connected network layer. This combined approach fully combines the advantages of spatial feature extraction with the powerful capabilities of time series modeling, breaking through the limitations of traditional signal processing methods in local feature extraction and long-term dependency modeling. It not only enhances the robustness and generalization ability of the present invention in processing nonlinear and time-varying signals, but also promotes the application potential of AIS signal processing technology in complex environments.

[0149] Based on the idea of ​​multi-objective optimization, this invention introduces mean square error loss constraints, interference suppression loss constraints and statistical correlation loss constraints to achieve separation performance optimization from different angles. Especially in dynamic targets and complex water environments, the model can quickly adapt to signal changes and effectively improve separation accuracy and real-time response capabilities.

[0150] The above embodiments are preferred examples of the present invention and are not intended to limit the scope of protection of the present invention.

Claims

1. A robust blind separation method for weak AIS signals for intelligent unmanned boats, characterized by: The following steps are involved: Step 1: Construct a random signal matrix set: By introducing different random number seeds, signal source changes, and signal lengths, a random AIS signal dataset is generated to simulate the signal distribution in real scenarios. Step 2, generator multi-dimensional feature extraction: Based on the above random signal matrix set, a generator module including a one-dimensional convolution module and a temporal feature extraction module is used to extract the spatial features and time series dependencies of the mixed signal to generate a reconstructed signal; Step 3: Adaptive constraint mechanism fusion: Based on the above feature extraction results, the generator module is jointly constrained by mean square error loss, interference suppression loss, statistical correlation loss and adversarial loss to optimize the generator training process; Step 4: Build a "zero-sum game" mechanism: At the same time, adopt an adversarial training framework between the discriminator and the generator. The discriminator uses a multi-layer fully connected network to perform binary classification between the real signal and the generated signal, and updates its parameters using the cross-entropy loss function. Step 5: Performance verification in low SNR scenarios: Under a 5dB SNR, the separation performance of the model in a complex channel environment is evaluated using the mean square error and signal-to-interference ratio indicators to verify the robustness of the method.

2. The robust blind separation method for AIS weak signals for intelligent unmanned boats according to claim 1 is characterized by: in, In step 1, the construction process of the random signal matrix set is: Step 1-1, generating a baseband signal, wherein the baseband signal is composed of a binary data sequence randomly generated by a pseudo-random number generator; Step 1-2: Encode and modulate the data using Gaussian minimum shift keying modulation to generate an AIS signal; Steps 1-3, add additive white noise with signal-to-noise ratio control to simulate multi-channel mixed signals; Step 1-4: Flatten the generated multi-channel mixed signal matrix into a vector and concatenate it with random noise to serve as the input of the generator module.

3. The robust blind separation method for AIS weak signals for intelligent unmanned boats according to claim 1, Its characteristics are: in, In step 2, the specific implementation process of the generator multidimensional feature extraction is: Step 2-1: Input the combined signal into a four-layer one-dimensional convolution module and extract local temporal features through the LeakyReLU activation function; Step 2-2: Transpose the high-dimensional features output by the convolution and input them into a two-layer long short-term memory network to capture the long-term temporal dependency of the signal and output the global temporal features. In steps 2-3, the high-dimensional features output by the LSTM are mapped to a dimension matching the true source signal through a two-layer fully connected network, and the Tanh activation function is used to constrain the output range; In steps 2-4, the generator generates a "false" signal based on the information extracted from the multi-dimensional features and inputs it into the discriminator to complete the adversarial training.

4. The robust blind separation method for AIS weak signals for intelligent unmanned boats according to claim 1 is characterized by: in, In step 3, a multidimensional constraint mechanism is set up in the generator part. The mean square error loss, interference suppression loss, statistical correlation loss and adversarial loss are introduced to jointly constrain the generator module. The total loss is backpropagated to update the generator training parameters. The introduction of multidimensional constraints can optimize the generator performance from multiple angles to achieve high-quality blind source separation of AIS signals.

5. The robust method for blind separation of weak AIS signals for intelligent unmanned boats according to claim 1, characterized in that: in, In step 4, the "zero-sum game" architecture is primarily based on the game mechanism between the discriminator and the generator. The true and false signals in the discriminator are fed into the discriminator, where they are gradually reduced in dimensionality through a three-layer fully connected network. The first fully connected layer maps the input signal to a 256-dimensional feature space, which is then compressed to 128 dimensions by the second fully connected layer. The third fully connected layer outputs a probability, representing the confidence that the signal is true. Since the AIS signal uses GMSK modulation, this GMSK signal is used as the true signal in this invention, along with the false signal generated by the generator, as input to the discriminator.

6. The robust blind separation method for AIS weak signals for intelligent unmanned boats according to claim 2 is characterized by: in, The specific process of step 1-1 is: After HDLC bit stuffing, synchronization sequence and buffer addition, and NRZI encoding, the baseband signal is passed through a Gaussian low-pass filter to give it Gaussian characteristics. The time domain impulse response of the Gaussian low-pass filter is: Among them, σ represents the standard deviation of the Gaussian function, which determines the extent of the impulse response; The specific process of steps 1-2 is as follows: The AIS signal can be generated by using MSK modulation on the signal after passing through the Gaussian low-pass filter. The MSK signal can be expressed as: S(t)=cos(2πf c t+φ(t)) (2) where f c is the carrier frequency, φ(t) is the instantaneous phase of the signal; The specific process of steps 1-3 is as follows: The carrier frequency range of the AIS signal modulation is 156.025 MHz to 162.025 MHz, and the carrier frequency is set to 156.775 MHz. Therefore, the GMSK signal can be expressed as: Among them, the modulation index h is used to characterize the intensity of signal modulation, b(t) represents the baseband signal, and the pulse shaping function h g (t) Used to filter the baseband signal to limit the signal bandwidth and control its signal characteristics; The specific process of steps 1-4 is as follows: The mixed signal matrix S(S∈R (N*N) ) and d Z dimensional random noise Z Splicing, flatten S into a vector vec(S)∈R (N*N) Then concatenate it with random noise Z to get the combined input:

7. The robust blind separation method for AIS weak signals for intelligent unmanned boats according to claim 3 is characterized by: in, The specific process of step 2-1 is: Expand the combined signal I into a three-dimensional length quantity I 3D Input spatial feature extraction module, which consists of four layers of 1D convolutional neural network. The convolution calculation formula of the first layer is: X (l) =LeakyReLU(Conv1D (l) (X (l-1) )) (5) The output size of the convolution operation is Among them C l is the number of channels, T (l) is the time series length, at this time The first layer of convolutional neural network sets the number of input channels to 1, the number of output channels to 16, and the convolution kernel size to 1*3 to extract basic low-level local temporal features. The next three layers use the number of input channels to 16, 32, and 64, and the number of output channels to 32, 64, and 128, respectively. The convolution kernel size is 1*3. The last three convolutional layers will gradually improve the feature abstraction level and capture higher-level signal patterns. LeakyReLU is used as the activation function after each convolution layer. The calculation process of multi-dimensional feature extraction is as follows: X (l) =LeakyReLU(Conv1D (l) (X (l-1) )) (7) h t ,c t =LSTM(X LSTM-in ,h t-1 ,c t-1 ) (9) X output =Tanh(ω2·LeakyReLU(ω1h final +b1)+b2) (11) Among them, S represents the combined signal matrix, Z represents random noise, I represents the mixed signal splicing noise matrix, and the output size of the convolution operation is where h t ∈R 256 is the hidden state to capture the temporal features, c t is the cell state at the current moment, h t-1 is the hidden state at the previous moment, c t-1 is the cell state at the previous moment, ω1 and b1 are the weight matrix and bias term of the first fully connected layer, ω2 and b2 are the weight matrix and bias term of the second fully connected layer, X output is the output of the time series feature extraction module, Tanh and LeakyReLU are activation functions; The specific process of step 2-2 is: After feature extraction, in order to adapt to the temporal feature extraction module, X CNN Transpose into dimension N×T×128, that is, X (LSTM-in) =X CNN T ∈R N×T×128 , X (LSTM-in) The input hidden unit is a time series feature extraction module with 256 hidden units. The time series feature extraction module is used to accept the high-dimensional features extracted by the convolution layer and use its memory mechanism to capture long-term dependencies. The two-layer stacked long short-term memory network provides a relatively powerful time modeling capability. The time series data after processing by the long short-term memory network is: h t ,c t =LSTM(X LSTM-in ,h t-1 ,c t-1 ) (12) where h t ∈R 256 is the hidden state to capture the temporal features, c t is the cell state at the current moment, h t-1 is the hidden state at the previous moment, c t-1 is the cell state at the previous moment. The output of the LSTM layer is the hidden state h at the last time step final =h T ∈R N×256 Represents the global characteristics of the sequence as a whole, where h T is the hidden state of the last time step; The specific process of steps 2-3 is as follows: The output of the time series feature extraction network passes through two layers of fully connected networks. The first layer represents the high-dimensional feature map, and the second layer outputs signal data that matches the dimension of the real source signal. Each layer of the fully connected network uses Tanh as an activation function to limit the output to the range of [-1, 1] to simulate the characteristics of the real signal. The Tanh function formula is: The Tanh activation function can maintain the normalization of the generated signal amplitude, making it easier to compare with the actual target signal and calculate the loss. final After the full connection layer mapping is: X output =Tanh(ω2·LeakyReLU(ω1h final +b1)+b2) (14) Where ω1 and b1 are the weight matrix and bias term of the first fully connected layer, respectively, and ω2 and b2 are the weight matrix and bias term of the second fully connected layer, respectively.

8. The robust blind separation method for AIS weak signals for intelligent unmanned boats according to claim 4 is characterized by: in, The specific process of step 3 is: Step 3-1, introduce mean square error loss based on adversarial loss: Among them, L adv is the generator adversarial loss, G(z) represents the sample generated by the generator G after receiving z, D(G(z)) is the judgment result of the discriminator D on G(z), and L MSE is the mean square error loss, y i is a GMSK signal; In step 3-2, the interference suppression loss is introduced as: Among them, L SIR is the interference suppression loss, ||·||2 represents the Euclidean norm, Set to a small constant to prevent division by zero errors; Step 3-3, introduce statistical correlation loss as follows: Among them, L cMI is the statistical correlation loss, L G is the total loss of the generator network, z is the noise distribution, y i is the true value signal, p(y i ,G(z)) is y i The joint probability density function with G(z), p(y i ) and p(G(z)) are their respective marginal probability density functions, and the numerical scales of different loss terms in the loss function may be different; In steps 3-4, the optimal weight parameters are set for each loss item to meet the multi-objective task requirements and ensure the stability of training. The total loss of the generator network is calculated as follows: L G =λ1·L MSE +λ2·L SIR +λ3·L cMI +λ4·L adv (19) In steps 3-5, the difference between the "false" signal generated by the generator and the true value is measured by the generator loss, forcing the generator to generate "false" signals that the discriminator cannot "distinguish"; In steps 3-6, the gradient descent algorithm is used to optimize the generator loss. Steps 3-5 and 3-6 are repeated until the discriminator cannot determine the authenticity of the generated signal.

9. The robust method for blind separation of weak AIS signals for intelligent unmanned boats according to claim 5, characterized in that: in, The specific process of step 4 is: Step 4-1: After the signal is input into the discriminator, it passes through a three-layer fully connected network to gradually reduce the dimension. The first layer of the fully connected network maps the input to a 256-dimensional space. Step 4-2, the second layer further maps the input to 128 dimensions; In step 4-3, the third layer outputs a single probability value to indicate the confidence that the signal is true. At the same time, the LeakyReLU function is used to provide nonlinear representation capabilities, and the Sigmoid activation function is used to output a probability value of [0,1] for binary classification. In the above process, the goal of the discriminator is to distinguish the real samples from the generated samples as accurately as possible, so the cross entropy loss function is used to calculate the loss L of the real samples respectively. real and the loss L of the generated samples fake , and then add them together to get the total discriminator loss L D : L D =L real +L fake (20) Among them, L real is the loss of the discriminator for real samples, L fake is the loss of the discriminator on the generated samples.

10. An intelligent blind separation system for AIS weak signals of unmanned boats, characterized by: include: Random signal matrix set building module: used to generate random AIS signal data sets; Generator multi-dimensional feature extraction module: used to extract the spatial features and time series dependencies of mixed signals and generate reconstructed signals; Adaptive constraint mechanism fusion module: used to optimize the generator training process; "Zero-sum game" mechanism building block: used to constrain the output range; Performance verification module: used to evaluate the separation performance of the model in complex channel environments and verify the robustness of the system.

Images