A method for evaluating target detection distance of underwater acoustic sonar under large water surface ship interference based on deep learning

Abstract

The present application belongs to the technical field of signal processing, and particularly relates to a large water surface ship interference underwater acoustic target detection distance evaluation method based on deep learning. The present application is based on deep convolution and multilayer perception machine method, takes standardized samples as input, faces multi-dimensional feature extraction, fusion and regression fitting learning, designs deep learning structure and parameter scheme, constructs an underwater acoustic target detection distance evaluation deep learning model, and completes model training based on a sample set. Finally, the underwater acoustic target detection distance under unknown large water surface ship interference is evaluated based on the deep learning model, and the corresponding result is output.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to a method for evaluating the detection range of sonar targets under interference from large surface ships based on deep learning. Background Technology

[0002] With the rapid development of marine resource development, underwater detection technology, and military offensive and defensive needs, sonar systems, as the core means of underwater target detection, have seen their performance evaluation and optimization become an important research direction in the field of underwater acoustics. Sonar target detection range is a core indicator for measuring the effectiveness of a sonar system, directly related to the reliability of underwater target perception and identification, and has a significant impact on the operational effectiveness of sonar.

[0003] Sonar detection range is affected by a variety of factors, including marine environmental parameters, target acoustic characteristics, and sonar platform hardware parameters. Traditional sonar detection range assessment is mainly based on theoretical calculations using sonar equations, without considering the impact of complex scenarios such as interference from large surface ships at sea. This makes it difficult for theoretical models to adapt to actual conditions, resulting in detection range prediction accuracy that fails to meet usage requirements.

[0004] In recent years, with the rapid development of technologies such as big data and artificial intelligence, performance evaluation technology based on deep learning has gradually emerged. By mining the complex nonlinear relationship between historical data and actual performance and using it for real-time data processing, the accuracy of performance evaluation in complex environments can be effectively improved.

[0005] Therefore, this application uses deep learning to evaluate the sonar target detection range under interference conditions of large surface ships. Based on establishing the relationship between parameters such as target, environment, and platform on the detection range, it further considers the differences in characteristics between the target and interference from large surface ships by mining deep learning. By making full use of multidimensional heterogeneous information, the accuracy of sonar detection range evaluation under complex and strong interference conditions is improved. Summary of the Invention

[0006] To improve the accuracy of sonar detection range assessment, this invention provides a deep learning-based method for assessing the sonar target detection range under interference from large surface ships.

[0007] The specific technical solution of this invention is as follows:

[0008] A deep learning-based method for evaluating the sonar target detection range under interference from large surface ships, the evaluation method comprising the following steps;

[0009] S1. Construction of standardized sample set: Based on the actual target detection sea trial data with interference from large surface ships, a training sample set is constructed. Each sample includes environmental parameters, interference parameters, target parameters, receiver directivity index, sonar detection range label, and LOFAR feature spectrum of target and interference.

[0010] S2. Deep Learning Model Construction: For environmental parameters, interference parameters, target parameters, and receiver directivity index, multilayer perceptron sub-modules are constructed based on multilayer perceptrons respectively. For target and interference LOFAR feature spectra, deep convolutional sub-modules are constructed based on deep convolutional network methods respectively. Based on the abstract features obtained from multiple multilayer perceptron sub-modules and deep convolutional sub-modules, heterogeneous data encoding sub-modules and deep fusion sub-modules are constructed to achieve alignment and deep fusion of multidimensional abstract features. Finally, a regression output sub-module is constructed.

[0011] S3. Iteratively train the constructed deep learning model: Initialize and fine-tune the deep learning model based on the actual sample set;

[0012] S4. Evaluation of sonar target detection range under interference from large surface ships: Based on environmental parameters, interference parameters, target parameters, receiver directivity index, and LOFAR feature spectra of the target and interference, standardized samples are constructed and input into a deep learning model to output the detection range evaluation results.

[0013] Furthermore, the construction of the deep convolutional submodule requires the construction of basic convolution operators, and the construction of deep convolutional submodules based on the basic convolution operators.

[0014] Furthermore, the construction of the basic convolution operator includes constructing basic module 1, basic module 2, basic module 3 and basic module 4. The construction of the basic module is achieved by establishing multiple branches and adding the convolutional features output by the multiple branches to obtain the output result of the module.

[0015] Furthermore, basic module 1 and basic module 2 are used for two-dimensional image feature mining, and basic module 3 and basic module 4 are used for one-dimensional sequence feature mining.

[0016] Furthermore, the construction of the deep convolutional submodule includes constructing deep convolutional submodule 1 and deep convolutional submodule 2.

[0017] Furthermore, the construction of the multilayer sensing submodule includes constructing multilayer sensing submodule 1, multilayer sensing submodule 2, multilayer sensing submodule 3 and multilayer sensing submodule 4.

[0018] Furthermore, the environmental parameters include the sound speed profile, sea depth, and receiving platform depth obtained on-site; the interference parameters include interference depth, interference source level, interference distance, and interference azimuth; and the target parameters include target depth, target source level, target distance, and target azimuth.

[0019] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0020] This invention, based on establishing the influence relationship between parameters such as target, environment, and platform on detection range, further considers the differences in characteristics between the target and large underwater acoustic ship interference based on deep learning. It makes full use of the differences in characteristics between the target and interference, as well as other data such as sound source level and propagation loss, and adopts a data-driven deep learning method to establish a fitting relationship between multi-dimensional heterogeneous information and detection range, thereby improving the accuracy of sonar detection range assessment in complex scenarios such as multiple targets and strong interference. Attached Figure Description

[0021] Figure 1 This is a flowchart of the sonar target detection range evaluation method of the present invention;

[0022] Figure 2 This is a schematic diagram illustrating the technical implementation principle of the sonar target detection range evaluation method of the present invention.

[0023] Figure 3 This is a flowchart illustrating the construction process of basic module 1 and basic module 2 of the present invention.

[0024] Figure 4 This is a flowchart illustrating the construction process of basic module 3 and basic module 4 of the present invention.

[0025] Figure 5 This is a schematic diagram of the multilayer perceptron submodule constructed based on the multilayer perceptron of the present invention. Detailed Implementation

[0026] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this invention.

[0027] Combination Figures 1-5 As shown, a method for evaluating the detection range of sonar targets under interference from large surface ships based on deep learning is described. The evaluation method includes the following steps.

[0028] S1. Construction of Standardized Sample Set: Based on the actual target detection sea trial data with interference from large surface ships, a training sample set is constructed. Each sample includes environmental parameters, interference parameters, target parameters, receiver directivity index, sonar detection range label, and LOFAR feature spectrum of the target and interference. The LOFAR feature spectrum is generated by framing, LOFAR analysis, and normalization based on the time-domain waveforms of the target and interference.

[0029] S2. Deep Learning Model Construction: For environmental parameters, interference parameters, target parameters, and receiver directivity index, multilayer perceptron sub-modules are constructed based on multilayer perceptrons respectively. For target and interference LOFAR feature spectra, deep convolutional sub-modules are constructed based on deep convolutional network methods respectively. Based on the abstract features obtained from multiple multilayer perceptron sub-modules and deep convolutional sub-modules, heterogeneous data encoding sub-modules and deep fusion sub-modules are constructed to achieve alignment and deep fusion of multidimensional abstract features. Finally, a regression output sub-module (i.e., a regression output layer) is constructed to output the target detection distance evaluation results in the subsequent output.

[0030] S3. Iteratively train the constructed deep learning model: Initialize and fine-tune the deep learning model based on the actual sample set;

[0031] S4. Evaluation of sonar target detection range under interference from large surface ships: Based on environmental parameters, interference parameters, target parameters, receiver directivity index, and LOFAR feature spectra of the target and interference, standardized samples are constructed and input into a deep learning model to output the detection range evaluation results.

[0032] Specifically, the construction of the deep convolution submodule requires the construction of basic convolution operators, and the construction of deep convolution submodules based on the basic convolution operators.

[0033] Specifically, the construction of the basic convolution operator includes constructing basic module 1, basic module 2, basic module 3 and basic module 4. The construction of the basic module is achieved by establishing multiple branches and adding the convolutional features output by the multiple branches to obtain the output result of the module.

[0034] The specific process of constructing the basic convolution operator is as follows:

[0035] 1.1 Construct basic module 1, with the number of output channels adjustable by parameter x, and add 3 parallel branches. Branch 1 includes a convolutional layer (7×3, 4x, [1,1]), a LayerNorm layer, a convolutional layer (3×1, x, [1,1]), and a GELU activation function. The convolutional layer parameter (7×3, x, [1,1]) indicates that the kernel size is 7×3, the number of convolutional channels is x, and the horizontal and vertical strides of the kernel are both 1, and so on. Branch 2 includes a convolutional layer (3×1, 4x, 1), a LayerNorm layer, a convolutional layer (1×1, x, [1,1]), and a GELU activation function. Branch 3 is a direct connection layer. The convolutional features output from the 3 branches are summed to obtain the final output of this module.

[0036] 1.2 Construct basic module 2, where the number of output channels can be set with parameter x, and add two parallel branches. Branch 1 consists of a convolutional layer (5×1, 2x, [1,1]), a LayerNorm layer, a convolutional layer (1×5, x, [1,1]), and a GELU activation function; Branch 2 consists of a convolutional layer (3×1, 2x, [1,1]), a LayerNorm layer, a convolutional layer (1×3, x, [1,1]), and a GELU activation function. The convolutional features output from the two branches are added together to obtain the final output of this module.

[0037] 1.3 Construct basic module 3, with the number of output channels adjustable by parameter x, and add 3 parallel branches. Branch 1 consists of a convolutional layer (16×1, 4x, [1,1]), a LayerNorm layer, a convolutional layer (8×1, x, [1,1]), and a GELU activation function; Branch 2 consists of a convolutional layer (7×1, 4x, 1), a LayerNorm layer, a convolutional layer (3×1, x, [1,1]), and a GELU activation function; Branch 3 is a direct-connected layer; the convolutional features output from the 3 branches are summed to obtain the final output of this module.

[0038] 1.4 Construct basic module 4, where the number of output channels can be set with parameter x, and add two parallel branches. Branch 1 consists of a convolutional layer (7×1, 2x, [1,1]), a LayerNorm layer, a convolutional layer (3×1, x, [1,1]), and a GELU activation function; Branch 2 consists of a convolutional layer (5×1, 2x, [1,1]), a LayerNorm layer, a convolutional layer (3×1, x, [1,1]), and a GELU activation function; the convolutional features output from the two branches are added together to obtain the final output of this module.

[0039] Specifically, the basic module 1 and basic module 2 are used for two-dimensional image feature mining, and the basic module 3 and basic module 4 are used for one-dimensional sequence feature mining.

[0040] Specifically, the construction of the deep convolutional submodule includes constructing deep convolutional submodule 1 and deep convolutional submodule 2.

[0041] The specific process of constructing the deep convolutional submodule is as follows:

[0042] 2.1 Construct a deep convolutional submodule 1, add a convolutional layer (13×7, 32, [1, 1]); add 5 basic modules 1, each with an output channel number x set to 128; add 1 basic module 2, with an output channel number x set to 128; add 5 basic modules 1, each with an output channel number x set to 256; add a global average pooling layer and a fully connected layer (256, 128), where the fully connected layer parameters (1024, 512) represent the number of input and output nodes as 1024 and 512, respectively.

[0043] 2.2 Constructing deep convolutional submodule 2, the specific process is the same as that of constructing deep convolutional submodule 1 in 2.1.

[0044] Specifically, the construction of the multilayer sensing submodule includes constructing multilayer sensing submodule 1, multilayer sensing submodule 2, multilayer sensing submodule 3 and multilayer sensing submodule 4.

[0045] The specific process of constructing the multilayer sensing machine submodule is as follows:

[0046] 3.1 Construct a multilayer perceptron submodule 1, add a one-dimensional sequence extension and merging layer to realize the splicing and addition of several one-dimensional sequences; add a fully connected layer (128, 256); add a fully connected layer (256, 128); add a fully connected layer (128, 64).

[0047] 3.2 Construct multilayer perceptron submodule 2 and add a fully connected layer (4, 16).

[0048] 3.3 Constructing the multilayer perceptron submodule 3, the specific process is the same as that of constructing the deep convolutional submodule 2 in 3.2.

[0049] 3.4 Construct the multilayer perceptron submodule 4 and add a fully connected layer (1, 16).

[0050] The specific process of constructing the heterogeneous data encoding submodule: add a fully connected layer (368, 512).

[0051] The specific process for building the deep fusion submodule is as follows:

[0052] Add matrix dimension transformation layer ,in and The numbers in parentheses represent the matrix dimensions before and after the transformation, respectively, and the numbers in parentheses represent the dimensions of each dimension of the matrix. A convolutional layer (24×1, 512, [1, 1]) is added. Five basic modules 3 are added, with each module having an output channel count x of 1024. One basic module 2 is added, with an output channel count x of 1024. Five basic modules 1 are added, with each module having an output channel count x of 256. One basic module 1 is added, with an output channel count x of 256. A global average pooling layer is added.

[0053] The specific process of constructing the regression output submodule is as follows: add a fully connected layer (256, 64); add a fully connected layer (64, 1).

[0054] Specifically, the environmental parameters include the sound speed profile, sea depth, and receiving platform depth obtained on-site; the interference parameters include interference depth, interference source level, interference distance, and interference azimuth; and the target parameters include target depth, target source level, target distance, and target azimuth.

[0055] Specifically, in step S3, the constructed deep learning model is trained iteratively, and the specific process is as follows:

[0056] The deep learning model was initialized and trained based on the actual sample set, using the Adam optimizer, with a learning rate (lr) set to 0.001 and a batch size of 64 for each training session.

[0057] The deep learning model was fine-tuned based on the actual sample set, using the Adam optimizer, with a learning rate (lr) of 0.0001 and a batch size of 64 for each training session.

[0058] Specifically, the evaluation of sonar target detection range under interference from large surface ships in step S4 is as follows: Based on the sound speed profile, sea depth, receiving platform depth, interference depth, interference source level, interference distance, interference azimuth, target depth, target source level, target distance, target azimuth, receiving directivity index, and LOFAR feature spectrum of the target and interference obtained on site, standardized samples are constructed and input into the deep learning model to output the detection range evaluation results.

[0059] Compared with existing sonar target detection range estimation methods, this invention makes full use of the differences between the target and the interference, as well as other sound source levels, propagation losses, and other data, utilizing more useful information. Furthermore, it adopts a data-driven deep learning method to establish a fitting relationship between multidimensional heterogeneous information and detection range, which can improve the accuracy of sonar detection range assessment in complex scenarios such as multiple targets and strong interference.

[0060] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications and substitutions based on the technical solutions and inventive concepts provided by the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for evaluating the sonar target detection range under interference from large surface ships, based on deep learning, characterized in that... The evaluation method Includes the following steps; S1. Construction of standardized sample set: Based on the actual target detection sea trial data with interference from large surface ships, a training sample set is constructed. Each sample includes environmental parameters, interference parameters, target parameters, receiver directivity index, sonar detection range label, and LOFAR feature spectrum of target and interference. S2. Deep Learning Model Construction: For environmental parameters, interference parameters, target parameters, and receiver directivity index, multilayer perceptron sub-modules are constructed based on multilayer perceptrons respectively. For target and interference LOFAR feature spectra, deep convolutional sub-modules are constructed based on deep convolutional network methods respectively. Based on the abstract features obtained from multiple multilayer perceptron sub-modules and deep convolutional sub-modules, heterogeneous data encoding sub-modules and deep fusion sub-modules are constructed to achieve alignment and deep fusion of multidimensional abstract features. Finally, a regression output sub-module is constructed. S3. Iteratively train the constructed deep learning model: Initialize and fine-tune the deep learning model based on the actual sample set; S4. Evaluation of sonar target detection range under interference from large surface ships: Based on environmental parameters, interference parameters, target parameters, receiver directivity index, and LOFAR feature spectra of the target and interference, standardized samples are constructed and input into a deep learning model to output the detection range evaluation results.

2. The method for evaluating the sonar target detection range under interference from large surface ships based on deep learning as described in claim 1, characterized in that: The construction of the deep convolution submodule requires the construction of basic convolution operators, and the construction of deep convolution submodules based on the basic convolution operators.

3. The method for evaluating the sonar target detection range under interference from large surface ships based on deep learning as described in claim 2, characterized in that: The construction of the basic convolution operator includes constructing basic module 1, basic module 2, basic module 3 and basic module 4. The construction of the basic module is achieved by establishing multiple branches and adding the convolutional features output by the multiple branches to obtain the output result of the module.

4. The method for evaluating the sonar target detection range under interference from large surface ships based on deep learning as described in claim 3, characterized in that: The basic modules 1 and 2 are used for two-dimensional image feature mining, and the basic modules 3 and 4 are used for one-dimensional sequence feature mining.

5. The method for evaluating the sonar target detection range under interference from large surface ships based on deep learning as described in claim 1, characterized in that: The construction of the deep convolutional submodule includes constructing deep convolutional submodule 1 and deep convolutional submodule 2.

6. The method for evaluating the sonar target detection range under interference from large surface ships based on deep learning as described in claim 1, characterized in that: The construction of the multilayer sensing submodule includes constructing multilayer sensing submodule 1, multilayer sensing submodule 2, multilayer sensing submodule 3 and multilayer sensing submodule 4.

7. The method for evaluating the sonar target detection range under interference from large surface ships based on deep learning as described in claim 1, characterized in that: The environmental parameters include the sound velocity profile, sea depth, and receiving platform depth obtained on-site; the interference parameters include interference depth, interference source level, interference distance, and interference azimuth; and the target parameters include target depth, target source level, target distance, and target azimuth.

Images