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

Abstract

The invention provides an intelligent detection method for frequency domain transient characteristics of a target in water based on a convolutional neural network. The method comprises the following steps: firstly, constructing a two-dimensional time-frequency spectrogram 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 frequency domain feature 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 frequency domain transient behavior detection result.

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 frequency-domain transient features based on a convolutional neural network. Background Art

[0002] Passive sonar target recognition uses the target radiated noise received by the sonar and other sensor information to discriminate the target type, and its focus lies in the construction and extraction of separable features. For a long time, the research on passive target feature extraction technology has mainly focused on steady-state features, and typical steady-state features such as line spectra and propeller parameters have been applied in actual scenarios. However, the underwater acoustic environment is extremely complex. Affected by factors such as multi-target strong interference, time-varying and space-varying features, background and platform noise, it is difficult to extract the steady-state features of weak targets at long distances. In addition to steady-state features in ship radiated noise, there are also transient features caused by changes in working conditions, behavior postures, etc. Compared with steady-state features, transient features have a short duration, and it is difficult to stably capture and utilize them using traditional information processing methods. Generally speaking, the transient features of ship radiated noise will cause frequency changes, and thus can be characterized in the frequency domain.

[0003] Deep learning is a type of artificial intelligence method that has been rapidly developed and applied in recent years. It can achieve non-linear fitting from high-dimensional input to target output by constructing complex network structures. With the support of data-driven, it can autonomously mine and utilize useful feature information. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies existing in the prior art and provide an intelligent detection method for underwater target frequency-domain transient features based on a convolutional neural network.

[0005] The purpose of the present invention is achieved through the following technical solutions. An intelligent detection method for underwater target frequency-domain transient features based on a convolutional neural network, which uses the deep learning method for target transient feature extraction. To utilize the transient characteristics represented in the frequency domain, facing the time-frequency spectrogram, by designing and constructing a deep learning model and its processing system adapted to the target characteristics, it is expected to achieve autonomous extraction and utilization of transient features, providing a basic support for passive target recognition; including the following steps:

[0006] Step 1: Construction of a standardized sample set: According to the underwater target radiated noise signal library, obtain a two-dimensional time-frequency spectrogram sample set S with transient label information STFT ;

[0007] Step 2: Construction of a frequency-domain transient feature intelligent 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 deep neural network detection model on this basis;

[0008] Step 3: Design a transient frequency domain feature enhanced learning loss function and train the intelligent detection model for frequency domain transient features based on S STFT ;

[0009] Step 4: Conduct transient feature detection on unknown underwater acoustic target data: Process the unknown underwater acoustic target data based on the intelligent detection model for frequency domain transient features and output the detection results of frequency domain transient behaviors.

[0010] The beneficial effects of the present invention are as follows: Compared with the existing passive target line spectrum extraction methods, the method proposed by the present invention has stronger non-linear relationship mapping ability, higher comprehensiveness and refinement in information utilization, and higher level in capturing weak transient features. Applying this method to actual passive target transient feature detection has achieved good results. 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 skilled in the art or ordinary technicians, other drawings can be obtained based on these drawings without creative efforts.

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

[0013] Figure 2 The signal processing flow chart is shown.

[0014] Figure 3 Two basic modules used in constructing the frequency domain transient feature detection model in the present invention are shown.

[0015] Figure 4 The construction method of the loss function used in training the intelligent detection model for frequency domain transient features is shown.

[0016] Figure 5 The results of intelligent detection of frequency domain transient features for the spectrogram of actual test target data based on the intelligent detection model for frequency domain transient features are shown.

[0017] Figure 6 The projection results of transient and non-transient numerical features in the intermediate output when processing the spectrograms of multiple groups of actual test target data based on the intelligent detection model for frequency domain transient features are shown. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all 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 shall fall within the protection scope of the present invention.

[0019] As Figure 1-6 shown, the present invention proposes an intelligent detection method for transient frequency domain features of underwater targets based on a convolutional neural network. Aiming at the characteristics of the time-frequency spectrogram of underwater targets, a basic convolutional operator is designed and a transient feature detection network model is constructed. Further, a loss function for weighted reinforcement learning of weak frequency domain transient fluctuation features is proposed to capture and detect target transient feature information from the spectrogram, providing feature support for passive target recognition.

[0020] As Figure 1 shown in the implementation principle block diagram, first, the underwater acoustic target data is preprocessed to construct a time-frequency spectrogram sample. Secondly, it is directly processed based on an end-to-end deep neural network model to output the detection result of the frequency domain transient feature.

[0021] As Figure 2 shown in the signal processing flow chart, it includes four stages: construction of a standardized sample set, construction of a transient feature detection model, training of the transient feature detection model, and detection of transient features of unknown underwater acoustic target data. The entire signal processing process is as follows.

[0022] (1) Construction of a standardized sample set

[0023] (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 transient features appear in the data). Perform normalization preprocessing on x to obtain x std .

[0024] (1.2) Take out the data one by one from x std and frame it according to a time window l frame of a certain length. Generally, 0.5f s ≤ l frame ≤ f s . Taking x n,std as an example, it can be expressed as T n where floor represents rounding down, and the same applies hereinafter. According to the data label, mark whether there are transient features in each frame of data. For x n,stdPerform wavelet transform time-frequency analysis on each sample in it to obtain the time-frequency spectrogram sample corresponding to each time-domain waveform sample. Taking as an example, its time-frequency diagram can be expressed as After processing all the data in x std , a labeled target time-frequency spectrogram sample set can be obtained

[0025] (2) Construction of the intelligent detection model for frequency-domain transient characteristics, and the basic process is as follows.

[0026] (2.1) Construct 2 basic modules, and the specific construction methods are as follows.

[0027] (2.1.1) Construct basic module 1, and the number of output channels of each convolutional layer is a parameter x that can be set. Add 3 parallel branches. Branch 1 sequentially includes a convolutional layer (7×7, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (5×5, 2x, [1, 1]), a convolutional layer (3×3, x, [1, 1]), and a GELU activation function; Branch 2 sequentially includes a convolutional layer (5×5, 4x, 1), a LayerNorm layer, a convolutional layer (3×3, x, [1, 1]), and a GELU activation function; Branch 3 is a direct connection layer; Add the convolutional features output by the 3 branches to obtain the final output result of this module.

[0028] (2.1.2) Construct basic module 2, and the number of output channels is a parameter x that can be set. Add 2 parallel branches. Branch 1 sequentially includes a convolutional layer (3×1, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (1×3, 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 2 branches to obtain the final output result of this module.

[0029] Figure 3 The 2 basic modules used in the construction of the frequency-domain transient characteristic detection model shown 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.

[0030] (2.2) Construct the intelligent detection model for frequency-domain transient characteristics, and the specific process is as follows.

[0031] Add a convolutional layer (10×10, 32, [1, 1]) and a convolutional layer (5×5, 64, [2, 2]) in sequence; add 5 basic modules 1, and set the output channel number x of each module to 128; add a convolutional layer (5×5, 128, [2, 2]); add 5 basic modules 1, and set the output channel number x of each module to 256; add a convolutional layer (5×5, 256, [2, 2]); add 5 basic modules 2, and set the output channel number x of each module to 512; add a convolutional layer (5×5, 256, [2, 2]); add 5 basic modules 2, and set the output channel number x of each module to 1024; add a global average pooling layer, a fully connected layer (1024, 512), a fully connected layer (512, 128), a fully connected layer (128, 2), a ReLU activation function, and a softmax function, where the fully connected layer parameter (1024, 512) indicates that the number of input and output nodes is 1024 and 512 respectively.

[0032] (3) Based on S STFT Train the intelligent detection model for frequency-domain transient features as follows.

[0033] (3.1) Construct the training loss function J of the intelligent detection model for frequency-domain transient features 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 STFT,i is the transient feature identification result of the i-th sample, E STFT,i is the transient frequency-domain fluctuation feature value of the i-th sample, μ STFT 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 STFT,i and is usually inversely proportional to the size of E STFT,i , that is, the larger E STFT,i , the smaller μ wave . E STFT,i can be calculated as follows. Assume that the spectrogram samples used for training include n t time points and n f frequency points, that is, a two-dimensional matrix of n t ×n f . Divide it equally along the time axis and frequency axis to obtain sub-spectrograms of the same size. l t and l f are the number of time points and frequency points of the sub-spectrogram respectively, and the general value range is 5 to 10. Then Among them, S i represents the i-th sample, including N s sub-spectrum diagrams, is the mean of S i 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 eigenvalue E wave,i is related to the average fluctuation of the amplitude of this sample.

[0034] Such as Figure 4 the construction method of the loss function used in the training of the intelligent detection model for frequency-domain transient features shown. This 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 the model to learn the basic transient feature detection ability; the regularization term is mainly the sample frequency-domain transient fluctuation value constructed in the present invention, which can enhance the learnability of weak frequency-domain transient features and enable the model to gradually master the weak change law of frequency-domain transient features.

[0035] (3.2) Train the intelligent detection model for frequency-domain transient features, and set training parameters such as the optimizer, learning rate, and data block size N bz during iterative training. Among them, the optimizer is set to Adam and the learning rate is set to 0.001. Randomly and with replacement, select N STFT samples from the training data set S bz , perform forward inference calculation based on the intelligent detection model for frequency-domain transient features 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), and then use the set Adam optimizer to optimize the structural parameters of the intelligent detection model for frequency-domain transient features.

[0036] (3.3) Based on the processing flow in (3.2), perform multiple rounds of training on the intelligent detection model for frequency-domain transient features until the loss function converges.

[0037] (4) Perform frequency-domain transient feature detection on 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 set of time-frequency spectrogram samples; use the time-domain transient feature intelligent detection model constructed in step (2) and trained in step (3) to perform forward inference on each time-frequency spectrogram sample one by one to obtain the frequency-domain transient feature detection results of this underwater acoustic target data at different time periods.

[0038] Such as Figure 5The results of the intelligent detection of the frequency-domain transient features of the actual test target data spectrogram by the intelligent detection model based on the frequency-domain transient features are shown. The red box indicates the detected transient features. It can be seen that the model can detect transient features from a complex noise environment.

[0039] As Figure 6 shown are the projection results of the transient and non-transient numerical features in the intermediate output when the intelligent detection model based on the frequency-domain transient features processes the spectrograms 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.

[0040] As mentioned above, the above are only specific embodiments 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 transient frequency domain features of underwater targets based on a convolutional neural network, characterized in that: It includes the following steps: Step 1. Construction of a standardized sample set: Based on the underwater target radiated noise signal library, obtain a two-dimensional time-frequency spectrogram sample set S with transient label information STFT ; Step 2. Construction of the intelligent detection model for frequency-domain transient features: 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 deep neural network detection model is constructed; Step 3. Design a transient frequency domain feature enhanced learning loss function and train the intelligent detection model for frequency domain transient features based on S STFT train the intelligent detection model for frequency domain transient features; Step 4. Detection of transient features of unknown underwater acoustic target data: Based on the intelligent detection model for frequency-domain transient features, the unknown underwater acoustic target data is processed, and the detection results of frequency-domain transient behaviors are output.

2. The intelligent detection method for the frequency-domain transient features of underwater targets based on a convolutional neural network according to claim 1, characterized in that: In the construction of the standardized sample set in Step 1, the specific steps are as follows: (1) Construction of the standardized sample set (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 , (\(n\in N\) * )}, and each row of data in the matrix corresponds to a target data. 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 certain length time window l frame Then, according to the data label, mark whether each frame of data has transient features; For x n,std When performing wavelet transform time-frequency analysis on each sample in as an example, its time-frequency diagram is expressed as For x std After processing all the data in 3. The intelligent detection method for frequency-domain transient features of underwater targets based on a convolutional neural network according to claim 2, characterized in that: In the construction of the intelligent detection model for frequency-domain transient features in Step 2, the specific steps are as follows: (2.1) Construction of 2 basic modules, and the specific construction methods are as follows: (2.1.1) Construction of basic module 1, where the number of output channels of each convolutional layer is a settable parameter x; Add 3 parallel branches. Branch 1 sequentially includes a convolutional layer (7×7, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (5×5, 2x, [1, 1]), a convolutional layer (3×3, x, [1, 1]), and a GELU activation function; Branch 2 sequentially includes a convolutional layer (5×5, 4x, 1), a LayerNorm layer, a convolutional layer (3×3, 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; (2.1.2) Construction of basic module 2, and the number of output channels is a settable parameter x; Add 2 parallel branches. Branch 1 sequentially includes a convolutional layer (3×1, 4x, [1, 1]), a LayerNorm layer, a convolutional layer (1×3, 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) Construction of the intelligent detection model for frequency-domain transient features, and the specific process is as follows: Sequentially add a convolutional layer (10×10, 32, [1, 1]) and a convolutional layer (5×5, 64, [2, 2]); Add 5 basic modules 1, and the number of output channels x of each module is set to 128; Add a convolutional layer (5×5, 128, [2, 2]); Add 5 basic modules 1, and the number of output channels x of each module is set to 256; Add a convolutional layer (5×5, 256, [2, 2]); Add 5 basic modules 2, and the number of output channels x of each module is set to 512; Add a convolutional layer (5×5, 256, [2, 2]); Add 5 basic modules 2, and the number of output channels x of each module is set to 1024; Add a global average pooling layer, a fully connected layer (1024, 512), a fully connected layer (512, 128), a fully connected layer (128, 2), a ReLU activation function, and a softmax function, where the parameters of the fully connected layer (1024, 512) indicate that the number of input and output nodes is 1024 and 512 respectively.

4. The intelligent detection method for transient features in the frequency domain of underwater targets based on a convolutional neural network according to claim 3, characterized in that: In the third step, based on S STFT train the intelligent detection model for frequency-domain transient characteristics. The specific steps are as follows: (3.1) Construct the training loss function J of the frequency-domain transient feature intelligent detection model wave , construct a basic function based on the cross-entropy calculation method, and then combine the transient fluctuation feature values of each sample to construct a regularization term 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 STFT,i is the transient feature identification result of the i-th sample, E STFT,i is the transient frequency-domain fluctuation feature value of the i-th sample, μ STFT is the weighting coefficient, and its value range is 0 to 1. It is set in segments according to the size of E STFT,i , and is inversely proportional to the size of E STFT,i , that is, the larger E STFT,i , the smaller μ wave ; E STFT,i is calculated as follows: Set the time-frequency spectrogram samples for training to include n t time points and n f frequency points, that is, an n t × n f two-dimensional matrix. Divide it equally along the time axis and frequency axis to obtain sub-spectrograms of the same size. l t and l f are the number of time points and frequency points of the sub-spectrogram respectively, and the value range is 5 to 10. Then where S i represents the i-th sample, including N s sub-spectrograms, is the mean of S i ; 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 frequency-domain transient features, and set the optimizer, learning rate, and data block size N during iterative training. bz Relevant training parameters, where the optimizer is set to Adam and the learning rate is set to 0.001; randomly select N STFT samples from the training dataset S bz with replacement, perform forward inference calculation based on the intelligent detection model for frequency-domain transient features to obtain the transient feature detection results. Combine with the sample labels, obtain the loss value based on the loss function constructed in (3.1), and then use the set Adam optimizer to optimize the structural parameters of the intelligent detection model for frequency-domain transient features. (3.3) Based on the processing flow in (3.2), conduct multiple rounds of training on the frequency-domain transient feature intelligent detection model until the loss function converges.

5. The intelligent detection method for the frequency-domain transient features of underwater targets based on a convolutional neural network according to claim 4, characterized in that: The specific steps for detecting the frequency-domain transient features of 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 and generate a set of time-frequency spectrogram samples; Use the time-domain transient feature intelligent detection model constructed in step two and trained in step three to perform forward inference on each time-frequency spectrogram sample one by one to obtain the frequency-domain transient feature detection results of the underwater acoustic target data in different time periods.

Images