Passive radiation noise multipath channel estimation method based on complex residual neural network

Abstract

The invention provides a passive radiation noise multipath channel estimation method based on a complex residual neural network. According to the method, complex number time-frequency characteristics of ship radiation noise are fully utilized, in a non-cooperative scene without prior information, time delay and amplitude parameters of a multipath channel are accurately estimated through a deep neural network, and the problems that a traditional method is unstable in performance and low in estimation precision under the condition of a low signal-to-noise ratio are solved. According to a simulation experiment, the estimation precision, robustness and path resolution capability of the algorithm are systematically analyzed from multiple angles of different signal-to-noise ratios, path strength and the like, the result shows that the method has stable and reliable performance under the ship radiation noise background, and the effectiveness and popularization potential of the method are verified.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of underwater acoustic signal processing, in particular to a passive radiated noise multipath channel estimation method based on a complex-valued residual network. BACKGROUND

[0002] As a core part of underwater target detection and positioning technology, the underwater acoustic channel estimation is greatly difficult due to the complexity of the marine environment and the multipath effect of the signal. The signal propagation produces significant time delay spread and frequency selective fading, which greatly increases the difficulty of channel estimation. The traditional estimation method based on pilot signals depends on the periodicity or known modulation structure of the signal, and is suitable for cooperative active sonar systems, but in the face of the non-periodic, broadband and prior unknown characteristics of the ship radiated noise in the non-cooperative passive sonar system, there are fundamental limitations.

[0003] The existing non-cooperative channel estimation methods mainly include: subspace decomposition based on statistical characteristics, second-order statistics method; time reversal mirror technology based on propagation model and its variants; LMS adaptive filtering algorithm based on optimization theory; and multi-channel correlation analysis technology. These methods are often limited by the periodicity assumption of the signal, the dependence on environmental priors, and the performance degradation under low signal-to-noise ratio when dealing with ship radiated noise.

[0004] In recent years, new methods such as machine learning and deep learning have been gradually applied to underwater acoustic channel estimation. Related research has shown that deep neural networks perform well in pilot signal estimation, but research on multipath channel estimation of ship radiated noise is still insufficient, and there is a lack of deep models that can fully exploit the time-frequency complex characteristics of the signal and maintain the integrity of the phase. SUMMARY

[0005] The present application aims to solve the problems in the prior art and proposes a passive radiated noise multipath channel estimation method based on a complex-valued residual network (CRN-CE) for non-cooperative channel estimation of ship radiated noise. This method fully utilizes the complex convolution structure to extract the amplitude and phase information of the signal in the time-frequency domain, combines the residual learning framework and the double-branch regression structure, and realizes accurate prediction of the time delay and amplitude parameters of multiple paths.

[0006] The present application is realized by the following technical solutions, and the present application proposes a passive radiated noise multipath channel estimation method based on a complex-valued residual network, which comprises: Step 1: data acquisition and signal preprocessing; Step 2: Construction of complex residual neural network model; a dual-input structure is adopted to process the real and imaginary parts separately, and the real and imaginary parts are cross-calculated through grouped complex convolution, thereby preserving complete phase information; the overall architecture of the complex residual neural network model includes four core modules, namely the input layer, multi-scale feature extraction, dual prediction branch and output layer; Step 3: Model training and optimization; Step 4: Channel parameter estimation and performance evaluation; Step 5: Robustness testing and model application.

[0007] Furthermore, in step one, the BELLHOP sound field model is used to simulate and generate multipath channel data in a typical deep-sea munk environment to obtain the sound source frequency, depth, and receiver array position parameters.

[0008] Furthermore, in step two, input feature selection is performed, choosing the STFT data of ship radiated noise as the input feature of the network.

[0009] Furthermore, in step two, a complex convolutional layer structure is designed. By using channel grouping constraints and orthogonal parameterization, the computational complexity is reduced while ensuring the phase-amplitude correlation of complex features. The specific implementation is as follows: Let the input tensor be ,in For an even number of channels, a complex representation is created through explicit channel partitioning:

[0010] Number of output channels Similarly, the number is forced to be even to ensure that the feature space satisfies... To ensure the integrity of complex dimensions and prevent dimension mismatch; grouped convolution is used to achieve independent processing of real and imaginary components:

[0011] in Independent convolutional kernels, This indicates an effective convolution operation; to maintain rotation invariance in the complex space, a bi-orthogonal constraint is applied to the convolution kernel:

[0012] The initialization process employs a modified Gramm-Schmidt orthogonal model, where the gain factor... Used to compensate for activation decay caused by grouping operations; bias terms are uniformly set to zero vectors to maintain the zero-central symmetry of complex features: .

[0013] Furthermore, the multi-scale feature extraction includes a residual learning module. The complex domain residual learning module employs a pre-activation structure to construct complex domain pre-activated residuals, and its calculation process is as follows:

[0014] Stable propagation of complex features is achieved using a pre-activation structure. Specifically, the pre-activation structure adopts a sequential connection from instance normalization to GELU activation and then to convolution. Instance normalization standardizes the real and imaginary parts separately, and the complex GELU activation function ensures that the phase changes continuously during the nonlinear transformation. Second, identity mapping is performed. When the input and output dimensions are inconsistent, 1×1 convolution is used to adjust the dimensions. Finally, for the characteristics of complex features, instance normalization is chosen instead of batch normalization.

[0015] Furthermore, a multi-path regression head based on a dual-branch structure was designed in the complex residual neural network model, consisting of a delay prediction branch and an amplitude prediction branch. The delay prediction branch adopts a structure of fully connected layers plus a Softplus activation function, while the amplitude prediction branch adopts a simple fully connected structure, maintaining the algebraic properties of the amplitude through L2 regularization constraints. The two branches share a feature extraction backbone network composed of complex convolutional layers, and finally output an estimation matrix, where the first and second columns correspond to the delay and amplitude estimates of multiple distinguishable paths, respectively.

[0016] Furthermore, in step two, the loss function is designed, and the design process for the loss function is as follows: First, assume the model outputs the following prediction set:

[0017] The real tag set is:

[0018] in , , ; Calculate the Euclidean distance between the estimated value and the true value to obtain the distance matrix. Specifically:

[0019] The globally optimal matching relationship is found based on the distance matrix and using the Hungarian algorithm. This ensures that each estimated path corresponds one-to-one with the actual path, minimizing the matching cost.

[0020] Define permutation space For all from Select from the estimates The permutations and combinations of , where Represent the true target Estimated index of allocation; After matching is completed, the latency and amplitude are calculated using MSE and Smooth L1 loss respectively, and learnable weights are introduced. and To balance the effects of the two, so that .

[0021] Furthermore, in step three, the AdamW optimizer is used to set the initial learning rate and weight decay for network parameter training; Dropout and instance normalization techniques are employed to prevent overfitting and improve training stability; the learning rate is dynamically adjusted using a cosine annealing scheduling strategy, combined with gradient accumulation to achieve large-scale training; and the model is repeatedly trained on a large-scale simulation dataset until it converges to obtain the best estimation performance.

[0022] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the passive radiated noise multipath channel estimation method based on complex residual neural networks.

[0023] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the passive radiated noise multipath channel estimation method based on a complex residual neural network.

[0024] The beneficial effects of this invention are: Compared with existing technologies, this invention innovatively utilizes a complex residual neural network to implicitly learn the delay and amplitude parameters of multipath channels directly from ship radiated noise. This eliminates reliance on prior information, fully preserves the phase and amplitude information of the complex time-frequency characteristics, and significantly improves the accuracy and robustness of channel estimation. Verification results based on Bellhop simulation data demonstrate that the proposed method performs superiorly in low signal-to-noise ratio environments, significantly outperforming traditional subspace and LMS algorithms. It provides an efficient, accurate, and easily implemented new technical solution for multipath channel estimation of ship radiated noise in non-cooperative passive sonar systems, and has broad application prospects. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0026] Figure 1 This is a profile of seawater sound velocity in a Munk environment.

[0027] Figure 2 This is a diagram of the complex residual neural network structure.

[0028] Figure 3 This is a comparative analysis chart of radiated noise from simulated ships.

[0029] Figure 4 This is a flowchart of a complex residual neural network.

[0030] Figure 5 This is a graph showing the change in the loss function.

[0031] Figure 6 The figure shows the subspace channel estimation results based on second-order statistics.

[0032] Figure 7 This is a diagram showing the channel estimation results based on the LMS algorithm.

[0033] Figure 8 The diagram shows the channel estimation results of the CRN-CE method of this invention.

[0034] Figure 9 This is a graph showing the variation of MAE with different signal-to-noise ratios. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Combination Figures 1-9 This invention proposes a passive radiated noise multipath channel estimation method based on a complex residual neural network, the method comprising: Step 1: Data Acquisition and Signal Preprocessing; In step one, using the BELLHOP sound field model, multipath channel data is generated in a typical deep-sea munk environment to obtain the sound source frequency, depth, and receiver array location parameters. The simulated seawater sound velocity profile is shown below. Figure 1 As shown.

[0037] Extract the top 6 strongest paths (in terms of amplitude) for each receiving point. (Sorting), constructing sparse impulse responses

[0038] The time delay is normalized. Sampling rate .

[0039] The sound source signal uses simulated ship radiated noise. Its generation process includes deterministic and random components, aiming to simulate key characteristics of real ship noise. The simulation signal duration is 5 seconds. The simulation signal is synthesized from the following elements: 1) Line spectrum components, including the axial frequency fundamental wave ( ) and its harmonics up to the 4th order, as well as characteristic high-frequency components (188 / 325 / 492Hz), where the phase follows Uniform distribution. 2) Continuous spectrum components are shaped by a 1024th-order FIR filter to achieve... Power characteristics. 3) Modulation effect: The modulation characteristics of a 4-bladed propeller are simulated using a Gaussian pulse sequence, with pulse width... Amplitude segmentation introduction Random perturbations.

[0040] Received signal Through source signal With channel impulse response The discrete convolution is obtained by adding additive white Gaussian noise to the received signal containing multipath effects. .

[0041] Time-frequency analysis was performed using short-time Fourier transform, and the final constructed eigenvectors were as follows:

[0042] Tag vector It includes 6 normalized amplitudes and 6 relative delays, with amplitudes in linear units and delays in seconds. The amplitude is amplified by 1e5 times to match the delay order of magnitude.

[0043] Ten sets of data are generated at each receiving point, for a total of 25,000 sets (50×50×10). The dataset is divided into a training set (70%), a validation set (10%), and a test set (20%) (fixed division, no overlap). Time-frequency masking and amplitude perturbation (±3dB) are added for data augmentation.

[0044] Step Two: Construction of the Complex Residual Neural Network Model; Addressing the issue of phase information loss when traditional real-valued convolutional networks process complex STFT features, this invention employs a dual-input structure to process the real and imaginary parts separately, and achieves cross-calculation of the real and imaginary parts through grouped complex convolutions, thereby preserving complete phase information; the overall architecture of the complex residual neural network model comprises four core modules: an input layer, multi-scale feature extraction, dual prediction branches, and an output layer; the overall architecture is as follows... Figure 2 As shown.

[0045] In step two, input feature selection is performed. In ship radiated noise signal processing, the choice of feature extraction method directly affects the accuracy and robustness of channel estimation. Traditional time-domain analysis methods, such as energy detection, struggle to characterize the dynamic characteristics of non-stationary signals, while frequency-domain methods, such as Fourier transform, cannot retain time evolution information. In contrast, STFT not only clearly extracts the characteristic line spectrum but also characterizes the transient characteristics of fluid noise through time-frequency energy distribution, as shown in the comparison. Figure 3 As shown. More importantly, the phase difference can directly reflect the multipath propagation delay. This time-frequency-complex joint characterization method makes STFT an ideal choice for ship radiated noise analysis and channel estimation. In summary, this invention selects STFT data of ship radiated noise as the input feature of the network.

[0046] In step two, a complex convolutional layer structure is designed. By using channel grouping constraints and orthogonal parameterization, the computational complexity is reduced while ensuring the phase-amplitude correlation of complex features. The specific implementation is as follows: Let the input tensor be ,in For an even number of channels, a complex representation is created through explicit channel partitioning:

[0047] Number of output channels Similarly, the number is forced to be even to ensure that the feature space satisfies... To ensure the integrity of complex dimensions and prevent dimensionality mismatch, grouped convolution (groups=2) is used to process real and imaginary components independently.

[0048] in Independent convolutional kernels, This indicates an effective convolution operation; to maintain rotation invariance in the complex space, a bi-orthogonal constraint is applied to the convolution kernel:

[0049] The initialization process employs a modified Gramm-Schmidt orthogonal model, where the gain factor... Used to compensate for activation decay caused by grouping operations; bias terms are uniformly set to zero vectors to maintain the zero-central symmetry of complex features: .

[0050] The multi-scale feature extraction includes a residual learning module. The complex domain residual learning module employs a pre-activation structure to construct complex domain pre-activated residuals, and its computational process is as follows:

[0051] Stable propagation of complex features is achieved using a pre-activation structure. Key characteristics of this process include... Figure 4 As shown, specifically: First, the pre-activation structure uses a sequential connection from instance normalization to GELU activation and then to convolution; instance normalization standardizes the real and imaginary parts separately, and the complex GELU activation function ensures that the phase changes continuously during the nonlinear transformation; second, identity mapping is performed; when the input and output dimensions are inconsistent, 1×1 convolution is used to adjust the dimensions; finally, considering the characteristics of complex features, instance normalization is chosen instead of batch normalization, which to some extent avoids the interference of small-batch statistics on phase information.

[0052] To achieve accurate joint estimation of time delay and amplitude in underwater multipath channels, a multipath regression head with a dual-branch structure was designed within a complex residual neural network model. This structure comprises a time delay prediction branch and an amplitude prediction branch. The time delay prediction branch employs a fully connected layer with a Softplus activation function to ensure the output strictly satisfies the non-negativity of the time delay parameter while avoiding the zero-gradient problem of ReLU-like functions. The amplitude prediction branch uses a simplified fully connected structure and maintains the algebraic properties of the amplitude through L2 regularization. Both branches share a feature extraction backbone network composed of complex convolutional layers, ultimately outputting a 6×2 dimensional estimation matrix. The first and second columns correspond to the time delay and amplitude estimates of the six resolvable paths, respectively. This dual-branch collaborative design considers both the differences in parameter characteristics and maintains the consistency of feature extraction.

[0053] In step two, the loss function is designed. The loss function of this invention is based on the Hungarian algorithm to simultaneously optimize the estimation accuracy of multiple channel parameters. The design process of this loss function is as follows: First, assume the model outputs the following prediction set:

[0054] The real tag set is:

[0055] in , , ; Calculate the Euclidean distance between the estimated value and the true value to obtain the distance matrix. Specifically:

[0056] The globally optimal matching relationship is found based on the distance matrix and using the Hungarian algorithm. This ensures that each estimated path corresponds one-to-one with the actual path, minimizing the matching cost.

[0057] Define permutation space For all from Select from the estimates The permutations and combinations of , where Represent the true target Estimated index of allocation; After matching is completed, the latency and amplitude are calculated using MSE and Smooth L1 loss respectively, and learnable weights are introduced. and To balance the effects of the two, so that .

[0058] This design fully considers the global optimality in the multi-path matching problem, while also taking into account the physical constraints of each parameter, thereby achieving high-precision parameter estimation.

[0059] Step 3: Model training and optimization; In step three, the AdamW optimizer is used to set the initial learning rate and weight decay for network parameter training; Dropout and instance normalization techniques are used to prevent overfitting and improve training stability; the learning rate is dynamically adjusted using a cosine annealing scheduling strategy, combined with gradient accumulation to achieve large-scale training; and the model is repeatedly trained on a large-scale simulation dataset until it converges to obtain the best estimation performance.

[0060] Step 4: Channel parameter estimation and performance evaluation; In step four, the received signals from the test set are input into the network to obtain multipath delay and amplitude estimates. Error metrics between the estimates and the true values ​​are calculated, including MAE, MSE, NMSE, and dual R² evaluation metrics, to comprehensively assess the estimation accuracy. Comparison with traditional second-order statistical subspace algorithms and LMS algorithms highlights the robustness and accuracy of the proposed method under low signal-to-noise ratio conditions. Analysis of error distribution, system bias, and path difference performance guides subsequent model optimization.

[0061] Step 5: Robustness testing and model application.

[0062] In step five, the simulation environment parameters are changed to generate diverse test sets to verify the model's adaptability under different environmental conditions.

[0063] Example The deep-sea munk environment acoustic channel was simulated using the BELLHOP ray theory model, and the parameter configuration is shown in Table 1.

[0064] Table 1 Dataset Parameter Settings

[0065] For each receiving point, the delay and amplitude of the six strongest multipath paths are extracted as labels.

[0066] The ship's radiated noise signal synthesis encompasses the fundamental shaft frequency and its harmonics, characteristic high-frequency components, and a Gaussian pulse sequence modulated by a simulated four-bladed propeller, with a duration of 5 seconds. Gaussian white noise is added after discretized convolution of the signal and the channel impulse response to form a realistic perceived signal. A Short-Time Fourier Transform (STFT) is applied to the signal, using a Hamming window of length 256, a frame shift of 128, and an FFT point of 512, to extract complex time-frequency features while preserving amplitude and phase information. The extracted feature tensors are used as network input to ensure the model can learn the phase differences and amplitude changes caused by multipath delays.

[0067] Training uses the AdamW optimizer with an initial learning rate of Weight decay Dropout 0.5, batch size 32, equivalent to 64 batches after gradient accumulation. Cosine annealing learning rate scheduling was used in a GPU environment. Experiments showed that the model reached optimal performance after 30 training epochs. Figure 5 As shown.

[0068] The loss function is based on the Hungarian algorithm to achieve globally optimal multipath matching, combining MSE and Smooth L1 loss with weight adjustment to promote coordinated optimization of delay and amplitude. Finally, the performance of the trained complex residual neural network is compared with that of traditional channel estimation algorithms, and the obtained channel impulse response is compared with the actual channel impulse response. Figure 6 , 7 As shown in Figure 8.

[0069] The evaluation was conducted using a comprehensive approach, incorporating MAE, MSE, NMSE, and dual R² indicators. The results are shown in Tables 2 and 3.

[0070] Table 2 Comparison of delay estimation accuracy of different algorithms

[0071] Table 3 Comparison of Amplitude Estimation Accuracy of Different Algorithms

[0072] Comparative experimental results show that the CRN-CE method proposed in this invention significantly outperforms traditional methods in channel estimation accuracy. The delay estimation MAE is only 0.10679 seconds, NMSE is 0.0739, and R² reaches 0.92608, demonstrating extremely strong delay resolution capabilities. The amplitude estimation NMSE is 0.32969, and R² reaches 0.67031, showing good overall amplitude variation capture, but there is still room for improvement in weak path amplitude estimation. Compared with traditional subspace algorithms based on second-order statistics and the LMS method, the model is significantly superior in all metrics.

[0073] according to Figure 9 Robustness analysis under different signal-to-noise ratios showed that the CRN-CE method maintained stable performance under low signal-to-noise ratio (-5dB~10dB) conditions and had good noise resistance.

[0074] In summary, this invention provides a passive radiated noise multipath channel estimation method based on complex residual neural networks, belonging to the fields of underwater acoustic signal processing and passive sonar technology. This method fully utilizes the complex time-frequency characteristics of ship radiated noise, accurately estimating the delay and amplitude parameters of the multipath channel through a deep neural network in non-cooperative scenarios without prior information. This overcomes the problems of unstable performance and low estimation accuracy of traditional methods under low signal-to-noise ratio conditions. Simulation experiments systematically analyze the estimation accuracy, robustness, and path resolution capability of the proposed algorithm from multiple perspectives, including different signal-to-noise ratios and path strengths. The results show that this invention has stable and reliable performance under ship radiated noise backgrounds, verifying the effectiveness and potential for widespread application of the proposed method.

[0075] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the passive radiated noise multipath channel estimation method based on complex residual neural networks.

[0076] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the passive radiated noise multipath channel estimation method based on a complex residual neural network.

[0077] The memory in this application embodiment can be volatile memory or non-volatile memory, or it can include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0078] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0079] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0080] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0081] The present invention provides a detailed description of a passive radiated noise multipath channel estimation method based on a complex residual neural network. Specific examples are used to illustrate the principle and implementation of the invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the invention. At the same time, those skilled in the art will know that there will be changes in the specific implementation and application scope based on the idea of ​​the invention. Therefore, the content of this specification should not be construed as a limitation of the invention.

Claims

1. A passive radiated noise multipath channel estimation method based on complex residual neural networks, characterized in that, The method includes: Step 1: Data Acquisition and Signal Preprocessing; Step 2: Construction of complex residual neural network model; a dual-input structure is adopted to process the real and imaginary parts separately, and the real and imaginary parts are cross-calculated through grouped complex convolution, thereby preserving complete phase information; the overall architecture of the complex residual neural network model includes four core modules, namely the input layer, multi-scale feature extraction, dual prediction branch and output layer; Step 3: Model training and optimization; Step 4: Channel parameter estimation and performance evaluation; Step 5: Robustness testing and model application.

2. The method according to claim 1, characterized in that, In step one, the BELLHOP sound field model is used to simulate and generate multipath channel data in a typical deep-sea munk environment to obtain the sound source frequency, depth and receiver array position parameters.

3. The method according to claim 1, characterized in that, In step two, input features are selected, and the STFT data of ship radiated noise is chosen as the input features of the network.

4. The method according to claim 1, characterized in that, In step two, a complex convolutional layer structure is designed. By using channel grouping constraints and orthogonal parameterization, the computational complexity is reduced while ensuring the phase-amplitude correlation of complex features. The specific implementation is as follows: Let the input tensor be ,in For an even number of channels, a complex representation is created through explicit channel partitioning: Number of output channels Similarly, the number is forced to be even to ensure that the feature space satisfies... To ensure the integrity of complex dimensions and prevent dimension mismatch; grouped convolution is used to achieve independent processing of real and imaginary components: in Independent convolutional kernels, This indicates an effective convolution operation; to maintain rotation invariance in the complex space, a bi-orthogonal constraint is applied to the convolution kernel: The initialization process employs a modified Gramm-Schmidt orthogonal model, where the gain factor... Used to compensate for activation decay caused by grouping operations; The bias term is uniformly set as a zero vector to maintain the zero central symmetry of complex features: 。 5. The method according to claim 1, characterized in that, The multi-scale feature extraction includes a residual learning module. The complex domain residual learning module employs a pre-activation structure to construct complex domain pre-activated residuals, and its computational process is as follows: Stable propagation of complex features is achieved using a pre-activation structure. Specifically, the pre-activation structure adopts a sequential connection from instance normalization to GELU activation and then to convolution. Instance normalization standardizes the real and imaginary parts separately, and the complex GELU activation function ensures that the phase changes continuously during the nonlinear transformation. Second, identity mapping is performed. When the input and output dimensions are inconsistent, 1×1 convolution is used to adjust the dimensions. Finally, for the characteristics of complex features, instance normalization is chosen instead of batch normalization.

6. The method according to claim 1, characterized in that, In the complex residual neural network model, a multi-path regression head based on a two-branch structure was designed, consisting of a time delay prediction branch and an amplitude prediction branch. The latency prediction branch adopts a structure of fully connected layers plus Softplus activation function, while the amplitude prediction branch adopts a simple fully connected structure. The algebraic properties of the amplitude are maintained by L2 regularization constraint. The two branches share a feature extraction backbone network composed of complex convolutional layers, and finally output an estimation matrix, where the first and second columns correspond to the latency and amplitude estimates of multiple distinguishable paths, respectively.

7. The method according to claim 1, characterized in that, In step two, the loss function is designed. The design process for the loss function is as follows: First, assume the model outputs the following prediction set: The real tag set is: in , , ; Calculate the Euclidean distance between the estimated value and the true value to obtain the distance matrix. Specifically: The globally optimal matching relationship is found based on the distance matrix and using the Hungarian algorithm. This ensures that each estimated path corresponds one-to-one with the actual path, minimizing the matching cost. Define permutation space For all from Select from the estimates The permutations and combinations of , where Represent the true target Estimated index of allocation; After matching is completed, the latency and amplitude are calculated using MSE and Smooth L1 loss respectively, and learnable weights are introduced. and To balance the effects of the two, so that 。 8. The method according to claim 1, characterized in that, In step three, the AdamW optimizer is used to set the initial learning rate and weight decay to train the network parameters; Dropout and instance normalization techniques are used to prevent overfitting and improve training stability. The learning rate is dynamically adjusted using a cosine annealing scheduling strategy, and gradient accumulation is used to achieve large-scale training effects. The model is repeatedly trained on a large-scale simulation dataset until it converges, thus obtaining the best estimation performance.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-8.

Images