Intelligent detection method for time domain transient characteristics of underwater target based on convolutional neural network

Abstract

The invention provides an underwater target time domain transient feature intelligent detection method based on a convolutional neural network. The method comprises the following steps: firstly, constructing a one-dimensional time domain waveform sample set with transient label information; secondly, designing a multi-scale convolution operator on the basis of a deep convolutional neural network method in combination with underwater acoustic target transient characteristic knowledge, and constructing a deep neural network detection model on the basis of the multi-scale convolution operator; then designing a transient waveform value reinforcement learning loss function and completing deep neural network detection model training based on the sample set; and finally, processing unknown target data based on the deep neural network detection model, and outputting a transient behavior detection result. According to the method, a deep neural network is constructed, a weak feature reinforcement learning method is designed, the comprehensiveness and the refinement degree of feature information mining and utilization are improved, and the method is an innovative method for application of an artificial intelligence algorithm in the field of underwater acoustic signal processing.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater target feature extraction and artificial intelligence, and mainly relates to an intelligent detection method for underwater target time-domain transient features based on a convolutional neural network. Background Technique

[0002] Passive sonar target recognition is an important research direction in the field of underwater acoustics and one of the difficulties faced in sonar information processing. The core of passive sonar target recognition lies in feature representation, construction, and extraction. Affected by the coupling of multiple factors such as the complex target sound generation mechanism, spatio-temporal variation of features, background noise, and complex sound propagation, it is difficult to model and construct passive target features. Currently, the available features mainly include typical steady-state features such as line spectra and propeller parameters, and the quantity is very scarce, with a large amount of useful information being discarded. In addition to steady-state features, ship targets will also generate certain transient features under conditions such as working condition conversion and drastic environmental changes, which can provide support for target identification. However, the transient feature generation mechanism is complex and diverse, and the duration is short. It is difficult to accurately model and extract using traditional information processing and probability statistics methods alone.

[0003] Currently, deep learning methods are a hot topic in the research and application of the artificial intelligence industry. They can achieve complex non-linear transformations through hierarchical neural network structures and directly mine useful information from high-dimensional complex data. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the existing technology and provide an intelligent detection method for underwater target time-domain transient features based on a convolutional neural network, which directly performs neural network processing on the original waveform to achieve transient feature identification and utilization, and provides auxiliary information support for target recognition.

[0005] The purpose of the present invention is completed through the following technical solutions. An intelligent detection method for underwater target time-domain transient features based on a convolutional neural network includes the following steps:

[0006] Step 1: Construction of a standardized sample set: According to the underwater target radiated noise signal library, obtain the target time-domain waveform sample set S with transient label information wave ;

[0007] Step 2: Construction of a transient feature detection model: Based on the deep convolutional neural network method, combined with the transient characteristics of underwater acoustic targets, design a multi-scale convolutional operator, and build a transient feature detection model on this basis;

[0008] Step 3: Design a transient waveform value reinforcement learning loss function, and train the intelligent detection model for time-domain transient features based on S wave ;

[0009] Step 4. Perform transient feature detection on the unknown underwater acoustic target data: Process the unknown underwater acoustic target data based on the transient feature detection model and output the transient behavior detection result.

[0010] The beneficial effects of the present invention are as follows: Compared with the existing passive target line spectrum extraction method, the method proposed by the present invention has a stronger non-linear relationship mapping ability, higher comprehensiveness and refinement in information utilization, and better comprehensive effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0012] Figure 1 Shown is the principle block diagram of the present invention.

[0013] Figure 2 Shown is the signal processing flow chart.

[0014] Figure 3 Shown are two basic modules used in the construction of the time-domain transient feature detection model in this patent.

[0015] Figure 4 Shown is a schematic diagram of the construction method of the loss function used in the training of the time-domain transient feature detection model.

[0016] Figure 5 Shown is a schematic diagram of the result of performing transient feature intelligent detection on the time-domain waveform diagram of the actual test target data based on the time-domain transient feature intelligent detection model.

[0017] Figure 6 Shown is a schematic diagram of the projection result of the transient and non-transient numerical features in the intermediate output when processing the time-domain waveform diagrams of multiple groups of actual test target data based on the time-domain transient feature intelligent detection model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the protection scope of the present invention.

[0019] Such as Figure 1-6As shown in the figure, the present invention proposes an intelligent detection method for the time-domain transient characteristics of underwater targets based on a convolutional neural network. First, a one-dimensional time-domain waveform sample set with transient label information is constructed; second, based on the deep convolutional neural network method and combined with the knowledge of the transient characteristics of underwater acoustic targets, a multi-scale convolutional operator is designed, and based on this, a deep neural network detection model is constructed; then, a transient waveform value reinforcement learning loss function is designed and the deep neural network detection model is trained based on the sample set; finally, the unknown target data is processed based on the deep neural network detection model, and the transient behavior detection result is output. As Figure 1 As shown in the implementation principle block diagram, first, the underwater acoustic target data is preprocessed to construct a time-frequency spectrum diagram sample, and then it is directly processed based on an end-to-end deep neural network model to output the time-domain transient feature detection result.

[0020] Figure 2 As shown in the signal processing flow chart, it includes four stages: constructing a standardized sample set, constructing a transient feature detection model, training the transient feature detection model, and detecting the transient features of unknown underwater acoustic target data. The specific implementation manner of the present invention is as follows:

[0021] (1) Construction of a standardized sample set

[0022] (1.1) Denote the labeled samples in the underwater target radiated noise signal library as x = {x1, x2,..., x n , (n ∈ N * )} (each row of data in the matrix corresponds to a target data, and the label information corresponding to each target data is the time when the transient feature appears in the data). Perform normalization preprocessing on x to obtain x std .

[0023] (1.2) Take out the data from x std one by one, and frame it according to a time window l frame of a certain length. Generally, it can be set that 0.5f s ≤ l frame ≤ f s . Taking x n,std as an example, it can be expressed as x n,std = {x n,std (t1), x n,std (t2),..., x n,std (t Nn ), N = floor(T n / l frame )}, where floor represents rounding down, and the same applies hereinafter. According to the data label, mark whether each frame of data has a transient feature. After processing all the data in x n , a labeled target time-domain waveform sample set can be obtained std .

[0024] (2) Construction of the transient feature detection model, the basic process is as follows.

[0025] (2.1) Construct two basic modules, and the specific construction methods are as follows.

[0026] (2.1.1) Construct basic module 1. To enhance the adaptability between the convolutional layer and the one-dimensional waveform, the number of output channels of all convolutional layers is a settable parameter x, and the size of some convolutional kernels is a variable parameter related to the input feature length. Add three parallel branches. Branch 1 sequentially includes a convolutional layer (D1×1, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (D2×1, 2x, [1, 1]), a convolutional layer (1×1, x, [1, 1]), and a GELU activation function, where the convolutional layer parameters (D1×1, x, [1, 1]) indicate that the convolutional kernel size is D1×1, the number of convolutional channels is x, and the horizontal and vertical strides of the convolutional kernel are both 1. D1 and x are both positive integers, and the same applies hereinafter; Branch 2 sequentially includes a convolutional layer (D3×1, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (D4×1, 2x, [1, 1]), a convolutional layer (1×1, x, [1, 1]), and a GELU activation function; Branch 3 is a direct connection layer; Add the convolutional features output by the three branches to obtain the final output result of this module. D1 to D4 generally take 5% to 10% of the input feature length.

[0027] (2.1.2) Construct basic module 2, and the number of output channels is a settable parameter x. Add two parallel branches. Branch 1 sequentially includes a convolutional layer (D5×1, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (D6×1, 2x, [1, 1]), a convolutional layer (1×1, x, [1, 1]), and a GELU activation function; Branch 2 is a direct connection layer; Add the convolutional features output by the two branches to obtain the final output result of this module.

[0028] Figure 3 The two basic modules used in the construction of the time-domain transient feature detection model in this patent are shown. They are mainly constructed based on multi-size convolutional operators, have the ability to mine multi-resolution features, and at the same time combine the direct connection structure to realize the construction of a large-depth network model.

[0029] (2.2) Construct the intelligent detection model for time-domain transient features, and the specific process is as follows.

[0030] Add a convolutional layer (128×1, 32, [4, 1]) and a convolutional layer (128×1, 64, [4, 1]) in sequence; add 6 basic modules 1, with D1, D2, D3, and D4 set to 64, 64, 32, and 32 respectively, and the number of output channels set to 128; add a convolutional layer (32×1, 128, [4, 1]); add 6 basic modules 1, with D1, D2, D3, and D4 set to 32, 32, 16, and 16 respectively, and the number of output channels set to 256; add a convolutional layer (16×1, 256, [4, 1]); add 6 basic modules 2, with D5 and D6 set to 8 and 8 respectively, and the number of output channels set to 512; add a convolutional layer (8×1, 512, [4, 1]); add 6 basic modules 2, with D5 and D6 set to 4 and 4 respectively, and the number of output channels set to 1024; add a global average pooling layer, a fully connected layer (1024, 256), a fully connected layer (256, 2), a ReLU activation function, and a softmax function, where the parameters of the fully connected layer (1024, 256) indicate that the number of input and output nodes is 1024 and 256 respectively.

[0031] (3) Based on S wave Train the time-domain transient feature intelligent detection model as follows.

[0032] (3.1) Construct the training loss function J of the time-domain transient feature intelligent detection model wave , generally, a basic function can be constructed based on the cross-entropy calculation method, and then a regularization term can be constructed by combining the transient fluctuation feature values of each sample to enhance the learning and mining ability of weak transient features. The loss function where N bz is the sample block size during model training, L i is the label of the i-th sample, P wave,i is the transient feature identification result of the i-th sample, E wave,i is the transient fluctuation feature value of the i-th sample, μ wave is the weighting coefficient, and its value range is 0 to 1. Generally, it can be set in segments according to the size of E wave,i and usually has an inverse relationship with the size of E wave,i , that is, the larger E wave,i , the smaller μ wave . E wave,i can be expressed as where S i represents the i-th sample, including N s data points, is the mean of S i , l is a positive integer, and generally its value range is 3 to 6. When a - l < 0, k = 0, and when a + l > N S , k = N sWhen the sample has no transient features, L i = 0, and at this time E wave,i = 0; when the sample has transient features, that is, L i = 1, the transient fluctuation eigenvalue E wave,i is related to the average fluctuation of the amplitude of the sample.

[0033] Figure 4 The method for constructing the loss function used in the training of the time-domain transient feature detection model is shown. The loss function mainly includes a basic term and a regularization term. The basic term is constructed based on the cross-entropy calculation method and is used for learning the basic transient feature detection ability of the model; the regularization term is mainly the transient fluctuation value of the sample constructed in this patent, which can enhance the learnability of weak transient features and enable the model to gradually master the weak change law of transient features.

[0034] (3.2) Train the time-domain transient feature intelligent detection model, and set training parameters such as the optimizer, learning rate, and data block size N bz during iterative training. The optimizer is set to Adam and the learning rate is set to 0.001. Randomly select N wave samples from the training dataset S bz with replacement, perform forward inference calculation based on the time-domain transient feature intelligent detection model to obtain the transient feature detection result. Combine the sample label, and obtain the loss value based on the loss function constructed in (3.1). Then use the set Adam optimizer to optimize the structural parameters of the time-domain waveform transient numerical feature extraction network model.

[0035] (3.3) Based on (3.2), perform multiple rounds of training on the time-domain transient feature intelligent detection model until the loss function converges.

[0036] (4) Perform transient feature detection on the unknown underwater acoustic target data, and the basic process is as follows. Use the method in step (1) to preprocess the unknown underwater acoustic target data to be recognized to generate a one-dimensional time-domain waveform sample sequence; use the time-domain transient feature intelligent detection model constructed in step (2) and trained in step (3) to perform forward inference on each one-dimensional time-domain waveform sample one by one to obtain the transient feature detection results of the underwater acoustic target data at different time periods.

[0037] Figure 5 The result of performing transient feature intelligent detection on the time-domain waveform diagram of the actual test target data based on the time-domain transient feature intelligent detection model is shown. The red box indicates the detected transient feature. It can be seen that the model can detect transient features from a complex noise environment.

[0038] Figure 6The figure shows the projection results of transient and non-transient numerical features in the intermediate output when the intelligent detection model based on the time-domain transient features processes the time-domain waveform diagrams of multiple groups of actual test target data. It can be seen that the transient and non-transient numerical features have good separability, indicating that the model can correctly identify and detect the frequency-domain transient features.

[0039] The present invention constructs a deep neural network and designs a weak feature enhancement learning method, which helps to improve the comprehensiveness and refinement of feature information mining and utilization, and is an innovative method for the application of artificial intelligence algorithms in the field of underwater acoustic signal processing.

[0040] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claimed rights.

Claims

1. An intelligent detection method for time-domain transient features of underwater targets based on convolutional neural network, characterized in that: It includes the following steps: Step 1. Construction of standardized sample set: According to the underwater target radiated noise signal library, obtain the target time-domain waveform sample set S with transient label information wave ; Step 2. Construction of transient feature detection model: Based on the deep convolutional neural network method and combined with the transient characteristics of underwater acoustic targets, a multi-scale convolutional operator is designed, and based on this, a transient feature detection model is constructed; Step 3. Design a transient waveform value reinforcement learning loss function and train the intelligent detection model for time-domain transient features based on S wave ; Step 4. Transient feature detection of unknown underwater acoustic target data: The unknown underwater acoustic target data is processed based on the transient feature detection model, and the transient behavior detection result is output.

2. The transient numerical feature extraction method for underwater targets based on deep learning according to claim 1, wherein: The construction of the standardized sample set in Step 1 is specifically as follows: (1.1) Denote the labeled samples in the underwater target radiated noise signal library as \(x = \{x_1, x_2, \ldots, x_{n}\}\) (\(n\in N\)), where each row of data in the matrix corresponds to a target data, and the label information corresponding to each target data is the time when transient features appear in the data; perform normalization preprocessing on \(x\) to obtain \(\widetilde{x}\); n , (\(n\in N\) * )}, where each row of data in the matrix corresponds to a target data, and the label information corresponding to each target data is the time when transient features appear in the data; perform normalization preprocessing on \(x\) to obtain \(\widetilde{x}\); std ; (1.2) Fetch data item by item from x std and frame it according to a time window l of a certain length frame Label whether there is a transient feature for each frame of data according to the data label; Process all the data in x std to obtain the labeled target time-domain waveform sample set S wave .

3. The method for extracting transient numerical features of underwater targets based on deep learning according to claim 2, characterized in that: The construction of the transient feature detection model in Step 2 is specifically as follows: (2.1) Construct 2 basic modules, and the specific construction method is as follows: (2.1.1) Construct basic module 1. To enhance the adaptability between the convolutional layer and the one-dimensional waveform, the number of output channels of all convolutional layers is the adjustable parameter x, and the size of some convolutional kernels is the variable parameter related to the input feature length; Add 3 parallel branches. Branch 1 sequentially includes a convolutional layer (D1×1, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (D2×1, 2x, [1, 1]), a convolutional layer (1×1, x, [1, 1]), and a GELU activation function. Here, the convolutional layer parameter (D1×1, x, [1, 1]) means that the convolutional kernel size is D1×1, the number of convolutional channels is x, the horizontal and vertical strides of the convolutional kernel are both 1, and D1 and x are both positive integers, the same below; Branch 2 sequentially includes a convolutional layer (D3×1, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (D4×1, 2x, [1, 1]), a convolutional layer (1×1, x, [1, 1]), and a GELU activation function; Branch 3 is a direct connection layer; The convolutional features output by the 3 branches are added together to obtain the final output result of this module. D1 to D4 take 5% to 10% of the input feature length; (2.1.2) Construct basic module 2. The number of output channels is the adjustable parameter x, and 2 parallel branches are added; Branch 1 sequentially includes a convolutional layer (D5×1, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (D6×1, 2x, [1, 1]), a convolutional layer (1×1, x, [1, 1]), and a GELU activation function; Branch 2 is a direct connection layer; The convolutional features output by the 2 branches are added together to obtain the final output result of this module; (2.2) Construct the intelligent detection model for time-domain transient features, and the specific process is as follows: Add a convolutional layer (128×1, 32, [4, 1]) and a convolutional layer (128×1, 64, [4, 1]) in sequence; add 6 basic modules 1, with D1, D2, D3, and D4 set to 64, 64, 32, and 32 respectively, and the number of output channels set to 128; add a convolutional layer (32×1, 128, [4, 1]); add 6 basic modules 1, with D1, D2, D3, and D4 set to 32, 32, 16, and 16 respectively, and the number of output channels set to 256; add a convolutional layer (16×1, 256, [4, 1]); add 6 basic modules 2, with D5 and D6 set to 8 and 8 respectively, and the number of output channels set to 512; add a convolutional layer (8×1, 512, [4, 1]); add 6 basic modules 2, with D5 and D6 set to 4 and 4 respectively, and the number of output channels set to 1024; add a global average pooling layer, a fully connected layer (1024, 256), a fully connected layer (256, 2), a ReLU activation function, and a softmax function, where the parameters of the fully connected layer (1024, 256) represent that the number of input and output nodes are 1024 and 256 respectively.

4. The method for extracting transient numerical features of underwater targets based on deep learning according to claim 3, wherein: In step 3, based on S wave Train the intelligent detection model for time-domain transient characteristics. The specific steps are as follows: (3.1) Construct the training loss function J of the time-domain transient feature intelligent detection model wave , construct the basic function based on the cross-entropy calculation method, and then combine the transient fluctuation feature values of each sample to construct the regularization term, which is used to enhance the learning and mining ability of weak transient features; the loss function where N bz is the sample block size during model training, L i is the label of the i-th sample, P wave,i is the transient feature identification result of the i-th sample, E wave,i is the transient fluctuation feature value of the i-th sample, μ wave is the weighting coefficient, and its value range is 0 to 1. It is set in segments according to the size of E wave,i , and is inversely proportional to the size of E wave,i , that is, the larger E wave,i , the smaller μ wave ; E wave,i is expressed as where, S i represents the i-th sample, including N s data points, is the mean value of S i , l is a positive integer, and its value range is 3 to 6. When a - l < 0, k = 0. When a + l > N S , k = N s ; when this sample has no transient features, L i = 0, and at this time E wave,i = 0; when this sample has transient features, that is, L i = 1, the transient fluctuation feature value E wave,i is related to the average fluctuation of the amplitude of this sample; (3.2) Train the intelligent detection model for time-domain transient features, and set the optimizer, learning rate, and data block size N during iterative training. bz Related training parameters, where the optimizer is set to Adam, the learning rate is set to 0.001, and N wave samples are randomly selected with replacement from the training dataset S bz samples are randomly selected with replacement from the training dataset S, and forward inference calculation is performed based on the intelligent detection model for time-domain transient features to obtain the transient feature detection results. Combining with the sample labels, the loss value is obtained based on the loss function constructed in (3.1), and then the structure parameters of the time-domain waveform transient numerical feature extraction network model are optimized using the set Adam optimizer. (3.3) Based on (3.2), conduct multiple rounds of training on the time-domain transient feature intelligent detection model until the loss function converges.

5. The method for extracting transient numerical features of underwater targets based on deep learning according to claim 4, wherein: The specific steps for detecting the transient features of the unknown underwater acoustic target data in step four are as follows: use the method in step one to preprocess the unknown underwater acoustic target data to be recognized to generate a one-dimensional time-domain waveform sample sequence; use the time-domain transient feature intelligent detection model constructed in step two and trained in step three to perform forward inference on each one-dimensional time-domain waveform sample one by one to obtain the transient feature detection results of the underwater acoustic target data at different time periods.

Images