A Classification Protection Method for Identifying Anti-Bait Signals of Ship Radiated Noise

Abstract

The present invention provides a method for identifying anti-bait signals of ship radiated noise with classified protection. Two generators in the time domain and time-frequency domain of the signal are respectively designed according to the data structure characteristics of the signal in the time domain and time-frequency domain. Among the three discriminators, two can realize the scoring of the generated data performance of the two generators, and the other can realize the fusion of the signal characteristics in the time domain and time-frequency domain and give a score. An adversarial training strategy is adopted, and the training set only contains ship radiated noise signals. In the test stage, the score of the current sample is directly obtained from the model output. Whether the current sample is a bait signal can be judged according to the size of the score, so as to realize the anti-bait interference of ship radiated noise. The present invention realizes anti-bait signals through the classification idea; currently, the achievements of anti-interference for ship radiated noise signals in the field of underwater acoustics are basically blank, and deep learning methods are used to solve the problem of anti-bait signals of ship radiated noise in underwater acoustic countermeasures.

Description

Technical Field

[0001] The present invention relates to the field of underwater acoustic signal processing, and specifically to a method for identifying anti-bait signals. Background Art

[0002] The ship radiated noise signal is one of the important ways to perceive ships. In the scenario of underwater acoustic signal perception and countermeasure. In order to confuse the enemy's perception system, releasing bait signals is a typical interference means. The bait signal imitates the radiated noise signal of the target ship to cover the target ship. Since the bait signal is highly similar to the target ship signal, it is difficult to distinguish the real target ship radiated noise from the bait signal. It is very difficult to have labeled data for target recognition in the confrontation situation, which greatly increases the recognition difficulty. If it is possible to realize the recognition of ship radiated noise and bait signals under single-class target data, currently, there is no dedicated anti-interference research result for ship radiated noise in the field of underwater acoustic signal processing. Facing the future complex marine war environment, it is urgent to study anti-interference methods for ship radiated noise. Summary of the Invention

[0003] In order to overcome the deficiencies of the prior art, the present invention provides a method for identifying anti-bait signals of ship radiated noise with classification protection. The present invention includes an end-to-end method for anti-bait signal interference of ship radiated noise without supervision. Two generators in the time domain and time-frequency domain of the signal are designed respectively according to the data structure characteristics of the signal in the time domain and time-frequency domain. The method includes three discriminators, two of which can realize the scores of the generated data performance of the two generators, and the other can realize the feature fusion of the time domain and time-frequency domain signals and give scores. Using the adversarial training strategy, the training set only contains ship radiated noise signals. In the test phase, the score of the current sample is directly obtained from the model output. According to the size of the score, it can be judged whether the current sample is a bait signal, so as to realize the anti-bait interference of ship radiated noise.

[0004] The specific steps of the technical solution adopted by the present invention to solve its technical problems are as follows:

[0005] Step 1: Frame the ship radiated noise signal to obtain ship radiated noise signal frames, and frame the bait signal to obtain bait signal frames. Label the ship radiated noise signal frames and bait signal frames respectively, that is, assign the label of the ship radiated noise signal frame as 1, and assign the label of the bait signal frame as 0, to form a ship radiated noise anti-bait signal database;

[0006] Step 2: Perform time-domain conversion and frequency-domain conversion on the model input signals respectively; the time-domain conversion is to segment the input signal frames into signal segments with overlap again, and perform concatenation operations on the signal segments in the channel dimension to form batch parallel computing during the training process. The frequency-domain conversion is to obtain the time-frequency matrix of ship radiated noise through short-time Fourier transform;

[0007] Step 3: Establish a deep learning model. Two generators respectively process the two types of input data obtained after the time-domain conversion and frequency-domain conversion in Step 2. The generator for processing time-domain information adopts a gated convolutional neural network to form an autoencoder structure, and the generator for processing frequency-domain information adopts a convolutional neural network to form an autoencoder structure. Three discriminators all adopt convolutional neural network and feedforward neural network structures. Two of the three discriminators respectively score the two, and the third discriminator realizes fusing time-domain and frequency-domain information to score the generator. The model structure is as Figure 1 shown;

[0008] Step 4: Divide the data in the ship radiated noise anti-bait signal database into a training set and a test set. The training set is used for training the network model, and the test set is used for testing the performance of the network model;

[0009] Step 5: Adopt an adversarial training strategy to train the deep learning model; use the backpropagation algorithm to update the model parameters;

[0010] Step 6: Perform performance testing on the trained deep learning model through the test set generated in Step 4. Take p s in Step 5 as the score of the model for the current test sample. Through the performance of the model on the test set, obtain the receiver operating characteristic curve ROC and the area under the receiver operating characteristic curve AUC;

[0011] Step 7: According to the test results in Step 6, that is, the AUC result, obtain the threshold according to the optimal ROC curve with the largest ROC value. When the data to be tested is input into the deep learning model to obtain the final score p s After that, if the final score p s output is greater than or equal to the threshold, the data to be tested is judged as a bait signal. If the model output is less than the threshold, it is a normal ship radiated noise signal, so as to realize the identification of bait signals and ship radiated noise signals.

[0012] In the above Step 4, 70% of the data in the ship radiated noise anti-bait signal database is used as the training set, and the remaining 30% is used as the test set.

[0013] In the above Step 5, when updating the model parameters through the backpropagation algorithm, the model parameter update method adopts AdamW, the learning rate is 0.00001, the momentum parameter is 0.9, and the batch size is 32.

[0014] In step 5, the training objective function of the deep learning model is as follows:

[0015] p t = D t (s t )

[0016] p f = D F (s f )

[0017] p s = D S (D T (s t ), D F (s f ))

[0018] GLoss cl = sign(p t + p f + p s )

[0019]

[0020] GLoss = GLoss score + GLoss content

[0021] DLoss cl = H(sign(p t )) + H(sign(p f )) + H(sign(p s ))

[0022]

[0023] where p t , p f , p s are the output scores of three discriminators, p t is the generator score for processing time-domain information, p f is the generator score for processing time-frequency domain information, p s is the score for fusing time-domain and time-frequency domain information, D t , D F , D S are three discriminators respectively, GLOSS cl represents the loss function, E(.) is to calculate the expected value, where s t , s f are the time-domain form and frequency-domain form of the ship radiated noise signal respectively, They are the time-domain form and the frequency-domain form after the reconstruction of the generator respectively.

[0024] The beneficial effects of the present invention are as follows: an unsupervised learning method is adopted and end-to-end classification of decoy signals and ship radiated noise signals can be achieved. Anti-decoy signals are realized through the classification idea; currently, in the field of underwater acoustics, the achievements of anti-interference for ship radiated noise signals are basically blank. The present invention mainly uses deep learning methods to solve the problem of anti-decoy signals of ship radiated noise in underwater acoustic countermeasures. The actual collected ship radiated noise signals and decoy signals are used for testing, and the AUC of the test result is 0.84. Description of the Drawings

[0025] Figure 1 It is a schematic diagram of the model structure of the present invention.

[0026] Figure 2 It is a bar chart of the test scores of the actual data of the present invention.

[0027] Figure 3 It is the TSNE visualization of the data in the actual data test set of the present invention. Detailed Embodiments

[0028] The present invention will be further described below in conjunction with the drawings and embodiments.

[0029] Step 1: Frame the ship radiated noise signals and decoy signals. The long signals are cut into signal frames with a length of 10 s, and the ship radiated noise signal frames are labeled as 1 and the decoy signal frames are labeled as 0. All the signal frames are written into an h5 file to form a ship radiated noise anti-decoy signal database.

[0030] Step 2: Perform time-domain transformation and time-frequency domain transformation on all the signal frames in Step 1. The time-domain transformation method is to cut the signal frames again and connect the segments in the 1 dimension. We use a 25% overlap with a length of 512 for cutting. The time-frequency domain transformation method is the short-time Fourier transform.

[0031] Step 3: Build a deep learning model. The specific structure of the deep learning is as Figure 1 shown. There are two generators, both of which are 10-layer autoencoder structures with cross-layer connections. Three generators are convolutional neural networks with cross-layer connections.

[0032] Step 4: Use 70% of the two types of data in the ship radiated noise anti-decoy signal database in Step 1 as the training set, and the remaining 30% as the test set.

[0033] Step 5: Train the deep learning model by training it on the training set using the backpropagation algorithm, and test the model performance using the test set data. The model parameter update method uses AdamW, with a learning rate of 0.00001, a momentum parameter of 0.9, and a batch size of 32.

[0034] Step 6: Perform a performance test on the trained deep learning model using the test set generated in Step 4. The ROC and AUC can be obtained from the model's performance on the test set.

[0035] Step 7: Obtain the threshold based on the optimal ROC curve. After splitting the signal to be measured and inputting it into the model, compare the output score of the model with the threshold to determine whether it is a decoy signal. The final actual data test results are as Figure 2 and Figure 3 shown.

Claims

1. A method for identifying anti-bait signals of ship radiated noise with classified protection, characterized in that It includes the following steps: Step 1: Frame the ship radiated noise signal to obtain the ship radiated noise signal frames, and frame the decoy signal to obtain the decoy signal frames. Label the ship radiated noise signal frames and the decoy signal frames respectively, that is, assign the label of 1 to the ship radiated noise signal frames and the label of 0 to the decoy signal frames to form a ship radiated noise anti-decoy signal database; Step 2: Perform time-domain conversion and frequency-domain conversion on the model input signal respectively; the time-domain conversion is to re-cut the input signal frames into signal segments with overlap, and perform a concatenation operation on the signal segments in the channel dimension to form batch parallel computing during the training process. The frequency-domain conversion is through the short-time Fourier transform to obtain the time-frequency matrix of the ship radiated noise; Step 3: Establish a deep learning model. Two generators respectively process the two types of input data obtained after the time-domain conversion and frequency-domain conversion in Step 2. The generator for processing time-domain information adopts a gated convolutional neural network to form an autoencoder structure, and the generator for processing frequency-domain information adopts a convolutional neural network to form an autoencoder structure. Three discriminators all adopt a convolutional neural network and a feedforward neural network structure. Two of the three discriminators score the two respectively, and the third discriminator realizes scoring the generator by fusing time-domain and frequency-domain information; Step 4: Divide the data in the ship radiated noise anti-decoy signal database into a training set and a test set. The training set is used for training the network model, and the test set is used for testing the performance of the network model; Step 5: Train the deep learning model using an adversarial training strategy; use the backpropagation algorithm to update the model parameters; Step 6: Perform performance testing on the trained deep learning model using the test set generated in Step 4, and take the p in Step 5 s as the score of the model for the current test sample. Based on the performance of the model on the test set, obtain the Receiver Operating Characteristic curve (ROC) and the Area Under the Curve (AUC) of the Receiver Operating Characteristic curve; Step 7: According to the test results in Step 6, that is, the AUC result, obtain the threshold based on the optimal ROC curve with the largest ROC value. When the data to be tested is input into the deep learning model to obtain the final score p s After that, if the final score p s output is greater than or equal to the threshold, the data to be tested is judged as a decoy signal. If the model output is less than the threshold, it is a normal ship radiated noise signal, so as to realize the identification of decoy signals and ship radiated noise signals.

2. The method for identifying anti-bait signals of ship radiated noise with classified protection according to claim 1, characterized in that: In Step 4, 70% of the data in the ship radiated noise anti-decoy signal database is used as the training set, and the remaining 30% is used as the test set.

3. The method for identifying anti-bait signals of ship radiated noise with classified protection according to claim 1, characterized in that: In Step 5, when updating the model parameters through the backpropagation algorithm, the model parameter update method adopts AdamW, the learning rate is 0.00001, the momentum parameter is 0.9, and the batch size is 32.

4. The method for identifying anti-bait signals of ship radiated noise with classified protection according to claim 1, characterized in that: In Step 5, the training objective function of the deep learning model is: p t = D T (s t ) p f = D F (s f ) p s = D S (D T (s t ), D F (s f )) GLoss cl = sign(p t + p f + p s ) GLoss = GLoss score + GLoss content Among them, p t , p f , p s are the output scores of three discriminators, p t is the generator score for processing time-domain information, p f is the generator score for processing time-frequency domain information, p s is the score for fusing time-domain and time-frequency domain information, D t , D F , D S are the three discriminators respectively, GLOSS cl represents the loss function, E(.) is to calculate the expected value, H(x) = max(0, 1 - x), where, s t , s f are the time-domain form and frequency-domain form of the ship radiated noise signal respectively, are the time-domain form and frequency-domain form after reconstruction by the generator respectively.

Images