A method and system for completing missing data of underwater acoustic signals based on a conditional diffusion model

Abstract

The application provides a kind of underwater acoustic signal missing data completion method and system based on conditional Diffusion model, comprising: obtaining complete underwater acoustic signal sample, generating missing observation signal and missing mask according to multiple missing modes, and constructing training sample set;Condition Diffusion model including forward and reverse diffusion process is constructed;The model is trained using the training sample set, and the missing observation signal and the mask are used as the condition to guide the reverse diffusion;The missing signal to be recovered is input into the trained model, the condition-guided reverse diffusion process is executed, and the complete completion signal is output.The application improves the generalization ability by multi-mode missing modeling, improves the recovery accuracy by using the high-order distribution learning ability of the conditional Diffusion model, and introduces cross attention, asymmetric loss and fast sampling mechanism, solves the problem of poor recovery effect and low inference efficiency of continuous large-area missing.

Description

Technical Field

[0001] This invention relates to the fields of underwater acoustic signal processing, artificial intelligence and deep learning technology, and specifically to a method and system for completing missing underwater acoustic signal data based on a conditional Diffusion model. Background Technology

[0002] Underwater acoustic signals are an important carrier for marine information acquisition and transmission, and have significant application value in underwater communication, target identification, sonar detection, and other fields. However, factors such as noise interference, multipath propagation, sensor failure, and network packet loss in complex marine environments often lead to data loss in the acquired underwater acoustic signals.

[0003] Existing technologies for missing data completion mainly include: First, traditional interpolation methods, such as linear interpolation and spline interpolation, which are simple to implement but only utilize local information and are difficult to recover the long-term dependency structure of complex signals. Second, statistical modeling methods, such as autoregressive models and Kalman filtering, which rely on pre-defined model assumptions and are insufficient in describing the distribution of nonlinear signals. Third, deep learning methods, such as convolutional neural networks and generative adversarial networks, which can improve accuracy to some extent, but still have problems such as poor ability to model the distribution of real data, poor recovery effect of continuous large-area missing data, overly smooth recovery results, and insufficient ability to maintain spectral structure.

[0004] In recent years, diffusion models have demonstrated excellent data distribution learning capabilities in image generation. However, when directly applied to underwater acoustic signal completion, they face challenges such as insufficient utilization of conditional information, low sampling efficiency, and limited recovery accuracy. Therefore, how to accurately recover missing underwater acoustic signals using observational information in complex marine environments has become a pressing technical problem to be solved in this field. Summary of the Invention

[0005] To address the problems of low recovery accuracy, insufficient long-range dependency modeling ability, poor recovery performance of continuous missing regions, and low inference efficiency of traditional interpolation methods, statistical modeling methods, and ordinary deep learning recovery methods in complex marine environments, this invention proposes a technical solution combining multi-modal missing data modeling, conditional Diffusion recovery networks, cosine noise scheduling strategies, Cross-Attention conditional injection mechanisms, asymmetric weighted loss functions, and DDIM consistency replacement recovery mechanisms. Specifically, this invention provides a method for completing missing underwater acoustic signal data based on a conditional Diffusion model, including: This invention provides a method for completing missing underwater acoustic signal data based on a conditional diffusion model, comprising the following steps: Step S1: Construct an underwater acoustic signal database and a missing sample library.

[0006] Step S2: Establish a modeling mechanism for missing data in complex marine environments.

[0007] Step S3: Construct an underwater acoustic signal completion network based on a conditional Diffusion model.

[0008] Step S4: Model training, system deployment, and signal completion output.

[0009] The above method enables the model to adapt to complex engineering scenarios by constructing a multi-mode missing sample set; it fundamentally improves the modeling accuracy of non-stationary signals by utilizing the powerful distributed learning capability of the conditional Diffusion model; and it achieves accurate information interaction between the observation area and the missing area by explicitly introducing the missing mask as a condition, thereby systematically solving the problems of low recovery accuracy and poor continuous missing recovery capability of existing technologies.

[0010] In some preferred embodiments, multiple missing patterns include random missing, consecutive missing, and marginal missing patterns. This preferred approach can more comprehensively simulate various data corruption scenarios in real-world engineering, significantly improving the model's generalization ability and robustness.

[0011] In some preferred embodiments, the Conditional Diffusion model employs a Conditional U-Net as the noise prediction network and introduces a multi-head self-attention mechanism in its bottleneck layer. This approach effectively captures long-range dependencies of the signal through self-attention, enhancing the model's ability to understand the context of continuously missing regions.

[0012] In some preferred embodiments, the forward diffusion process employs a cosine noise scheduling strategy. This strategy can slow down the rate of loss of signal structure information in the early stages of diffusion, better preserving the spectral and periodic characteristics of the original signal.

[0013] In some preferred embodiments, a cross-attention mechanism is used to fuse missing observation signals and missing masks as conditional information with diffusion features. This scheme establishes a global correlation between the missing region and the observed region, significantly improving the efficiency of utilizing known information.

[0014] In some preferred embodiments, an asymmetric weighted loss function is used for training, assigning higher weights to missing regions. This approach allows the model training focus to be more concentrated on difficult regions, effectively improving the recovery accuracy for continuous and large-area missing parts.

[0015] In some preferred embodiments, backdiffusion is performed using DDIM deterministic sampling and a skip-step strategy. This scheme significantly reduces the number of inference steps while ensuring recovery quality, thus improving the real-time performance of engineering applications. Furthermore, a consistency replacement operation is performed in each backdiffusion step to preserve the true data of the observed area and prevent distortion.

[0016] The present invention also provides a system for completing missing underwater acoustic signal data based on a conditional diffusion model, comprising: a data preparation module, a model building module, a model training module, and a signal recovery module, each module being used to implement the steps of the method in the first aspect above.

[0017] Compared with the prior art, the present invention has the following technical effects: By constructing a multi-modal missing data modeling mechanism, this invention significantly improves the model's adaptability and generalization ability to various complex scenarios, including random, continuous, and marginal missing data. Employing a conditional diffusion model as the core framework, it effectively learns and reconstructs the high-order statistical features of complex non-stationary underwater acoustic signals, avoiding the over-smoothing problem of traditional methods. By introducing a Cross-Attention conditional injection mechanism, a global correlation between the missing region and the observation region is established, greatly improving long-distance dependency modeling capabilities and information utilization efficiency. Through the design of a cosine noise scheduling strategy and an asymmetric weighted loss function, the preservation of signal detail structure and missing region recovery accuracy is optimized from the perspectives of diffusion process and training objectives, respectively. By adopting DDIM sampling and consistency replacement mechanisms, inference efficiency is significantly improved while ensuring high recovery accuracy, meeting the real-time requirements of practical engineering deployments. Compared with existing methods, this invention can achieve high-precision, high-stability, and high-efficiency underwater acoustic signal missing data recovery in complex marine environments, providing reliable data support for underwater communication, sonar detection, and intelligent marine sensing. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall process of the present invention; Figure 2 This is a flowchart of the underwater acoustic signal completion process of the present invention; Figure 3 This is a flowchart of the missing data modeling process for this invention; Figure 4 This is a flowchart of the Diffusion-based network training process for the present invention. Figure 5 This is a flowchart illustrating the system deployment and underwater acoustic signal completion output of the present invention. Figure 6 This is a diagram of the overall architecture of the present invention. Detailed Implementation

[0019] This invention provides a method and system for completing missing underwater acoustic signal data based on a conditional diffusion model, aiming to solve the problem of missing underwater acoustic signals caused by noise, interference, equipment failure, etc. in complex marine environments, and to achieve high-precision and high-fidelity signal recovery.

[0020] Reference Figure 1The method of the present invention mainly includes the following steps: Step S1: Construct an underwater acoustic signal database and a missing sample library. Collect raw underwater acoustic signals and artificially construct various missing patterns such as random, continuous, and edge-based signals to generate missing observation signals and corresponding missing masks.

[0021] Step S2: Establish a missing data modeling mechanism for complex marine environments. Learn the signal distribution through a forward diffusion process and design a Conditional U-Net as a denoising network.

[0022] Step S3: Construct an underwater acoustic signal completion network based on a conditional Diffusion model. Using the complete signal as the supervision target and the missing signal and mask as conditions, optimize the asymmetric weighted loss function.

[0023] Step S4: System Deployment and Underwater Acoustic Signal Completion Output Process. The missing signal to be recovered is input, and DDIM-accelerated sampling and consistency replacement are used to output the complete signal.

[0024] Through the above steps, this invention organically combines multi-mode data augmentation, conditional diffusion generation, and an efficient sampling mechanism. Compared to traditional methods, this approach can more accurately model the distribution of complex underwater acoustic signals, exhibits superior recovery performance in challenging scenarios such as continuous data loss, and significantly improves inference speed, providing a feasible technical solution for practical engineering applications.

[0025] The following is a detailed explanation of each step.

[0026] Step S1: Construct an underwater acoustic signal database and a missing sample library Step S11: Acquire raw underwater acoustic signals In practice, the raw underwater acoustic signals originate from data collected by underwater communication equipment, sonar detection systems, underwater target monitoring platforms, and marine environmental monitoring equipment. The collected data includes active sonar echo signals, passive sonar listening signals, underwater communication modulation signals, marine environmental background noise signals, and underwater target radiated noise signals. The sampling frequency is set to 2000Hz, and each signal segment has 800 sampling points. A sliding window method is used for sample segmentation, with a window length of 800 and a sliding step size of 200. This windowing process yields a large number of fixed-length samples.

[0027] To improve the stability of model training, all underwater acoustic signals were normalized. Z-score normalization was used to ensure the signal mean was 0 and the variance was 1, eliminating amplitude differences caused by different sampling environments and equipment.

[0028] Step S12: Establish a missing sample generation module To simulate data loss phenomena in real marine environments, this invention constructs three types of missing data patterns: The first type is the random missing mode, which randomly selects 30% of the sampling points in the signal as missing locations to simulate marine environmental noise interference, instantaneous communication interruption, and sampling errors.

[0029] The second type is the continuous missing mode, which randomly selects a continuous time interval for missing data. The missing length accounts for 10% to 70% of the total signal length, and 20% in the preferred embodiment. It is used to simulate situations such as underwater communication link interruption, equipment failure, and long-term signal obstruction.

[0030] The third type is the edge missing mode, which constructs a continuous missing region at the beginning or end of the signal, with the missing length accounting for 10% to 30% of the total length. It is used to simulate abnormal device startup and communication interruption scenarios.

[0031] Furthermore, this invention employs a hybrid missing mechanism: random missing, continuous missing, and marginal missing are used together in a 3:4:3 ratio during training, enabling the model to adapt to more complex engineering application scenarios.

[0032] Step S13: Construct the training set, validation set, and test set The generated dataset is divided into training, validation, and test sets. The training set accounts for 70% of the total samples, the validation set accounts for 15%, and the test set accounts for 15%. At the same time, the three missing data patterns are evenly distributed across the datasets to avoid training bias.

[0033] By constructing a standardized underwater acoustic signal database and a multi-mode missing sample library, a highly realistic simulation of missing data scenarios in complex marine environments was achieved. Compared with traditional single random missing data modeling methods, the hybrid missing data system established in this invention can more realistically reflect the data corruption that occurs in actual underwater acoustic communication and sonar detection processes, providing a reliable data foundation for subsequent Diffusion model training and effectively improving the model's generalization ability and engineering applicability.

[0034] Step S2: Establish a missing data modeling mechanism for complex marine environments Step S21: Establish a mathematical model for missing underwater acoustic signals To ensure the model accurately describes the problem of recovering missing underwater acoustic signals, a mathematical model for the missing data is first established. Let the complete underwater acoustic signal be represented as... ; Where L represents the signal length, x i This represents the amplitude of the i-th sampling point.

[0035] Introducing a binary mask matrix: , Where m i ∈{0,1} is defined as follows: When m i When m = 1, it means that the data at the i-th position is observable; when m i When =0, it means that the data at position i is missing.

[0036] The observed signal can be represented as: ; in, This represents element-wise multiplication. After the above processing, the original data values ​​are retained in the observation area, while the missing regions are set to zero, thus forming missing underwater acoustic signal samples.

[0037] Furthermore, to improve the model's adaptability in complex marine environments, this invention establishes a multi-mode missing data mechanism. The missing data mechanism includes: random missing data mode; continuous missing data mode; and marginal missing data mode. These three types of missing data modes can cover most data corruption scenarios in real-world engineering situations.

[0038] Step S22: Construct a random missing pattern Random missing data is mainly used to simulate situations such as marine environmental noise interference, transient communication interruptions, and sampling errors.

[0039] Set missing ratio Its value range is In the preferred embodiment, the value is 30%. A random sampling algorithm is used to randomly select from a signal sequence of length L. For each sampling position, assign the corresponding position mask value to m. i =0, keep m at other positions i =1, ultimately forming a randomly discrete missing region. The random missing pattern is characterized by uncertain location distribution and a large number of missing points. It can simulate random bit error scenarios in complex marine environments.

[0040] Step S23: Constructing a continuous missing pattern Continuous loss is mainly used to simulate situations such as underwater communication link interruption, equipment failure, and long-term signal obstruction.

[0041] A random starting position s (1≤s≤L) is selected for continuous missing data. The missing length l is set to a range of 0.1L≤l≤0.7L. In a preferred embodiment, the continuous missing length accounts for 20% of the total signal length. All sampling points in the interval [s, s+l] are assigned the value m. i=0, forming a continuous missing region. Continuous missing patterns are more difficult to complete than random missing patterns because there is no observational information within the missing region. The model can only infer the structure of the missing part based on the surrounding context. Therefore, continuous missing patterns are an important test scenario for evaluating the model's completion ability.

[0042] During training, the proportion of consecutive missing samples is increased, enabling the model to focus on learning long-distance temporal dependencies.

[0043] Step S24: Construct the edge missing pattern Edge missing is mainly used to simulate real-world problems such as data loss during device startup, data truncation at the end of transmission, and cache corruption.

[0044] Edge-related missing data includes front-end missing data and back-end missing data. For front-end missing data, the first k sampling points of the signal are set to the missing state, i.e., m1, m2, ..., mk. k =0; For back-end missing data, set the last k sampling points of the signal to the missing state, i.e., m L-k+1 ,...,m L =0. Here, k accounts for 10%–30% of the total length. Edge-missing patterns lack contextual information from one side compared to continuous-missing patterns, making recovery more difficult. Therefore, this allows for further verification of the model's generalization performance.

[0045] Step S25: Establish a unified missing mask representation mechanism To enable different missing patterns to be input into the model in a unified manner, this invention also constructs a standardized mask encoding method.

[0046] The construction of a standardized mask encoding method is specifically represented as follows: ; Where L is the length of the aquatic signal, Mask represents a 1-row, L-column vector of real numbers, with the length of the Mask matching that of the original signal.

[0047] Input the Mask as conditional information into the Diffusion model to explicitly tell the model which locations are real observation data and which locations need to be regenerated.

[0048] Compared to traditional missing completion methods that only input the missing signal, this invention introduces Mask constraints. This allows the model to more accurately distinguish between known and unknown regions, effectively reducing the problem of erroneous generation during the restoration process.

[0049] Step S26: Forming conditional Diffusion model input data Construct the complete sample into triplet form: ; Where x represents the complete true signal, x obs denoted by , where M represents the missing observation signal.

[0050] During the training phase, the model uses x as the supervision target, and... As conditional input, the missing regions are learned to restore the mapping relationship. Only input is used during the inference phase. The complete signal can then be recovered.

[0051] Through the above steps, this invention establishes a multi-mode missing information modeling system for complex marine environments. This system not only realistically simulates engineering scenarios such as random bit errors, communication interruptions, equipment failures, and data truncation, but also achieves explicit decoupling of missing information from observational information through a unified mask representation mechanism, providing a standardized input foundation for subsequent Cross-Attention conditional injection mechanisms and conditional Diffusion model training.

[0052] Compared with existing methods that only use single random missing data modeling, the multi-mode missing data modeling mechanism proposed in this invention can significantly improve the model's adaptability to complex missing data scenarios, thereby improving the accuracy and robustness of underwater acoustic signal completion results.

[0053] Step S3: Construct an underwater acoustic signal completion network based on a conditional Diffusion model Step S31: Construct the conditional Diffusion completion network architecture The network includes a missing observation signal input module, a missing mask input module, a time step encoding module, a Conditional U-Net backbone network, a Cross-Attention conditional fusion module, a noise prediction output module, and a DDIM recovery module.

[0054] Conditional U-Net employs an Encoder-Decoder structure, including a four-level encoder, an intermediate bottleneck layer, a four-level decoder, and a skip connection module. The encoder uses a multi-level residual convolutional structure, and the decoder uses a symmetric upsampling structure. Feature fusion is performed between the encoder and decoder through skip connections, and the bottleneck layer introduces a multi-head self-attention mechanism to establish long-distance temporal dependencies.

[0055] The encoder progressively extracts temporal features at different scales, while the decoder progressively reconstructs the signal structure. Skip connections are used to fuse shallow detail information with deep semantic information. A multi-head self-attention mechanism is introduced at the bottleneck layer to establish long-distance temporal dependencies, thereby enhancing the model's ability to model the global structural features of complex underwater acoustic signals.

[0056] Step S32: Establish a forward diffusion process that adapts to timing characteristics A cosine noise scheduling strategy is used to construct the cumulative retention coefficient: ; in: ; Where: t represents the diffusion time step; T represents the total number of time steps in the diffusion process; s represents the cosine scheduling offset parameter, which is used to avoid the cumulative retention coefficient changing too quickly in the initial stage.

[0057] By using cosine modulation, the rate of information loss in the early stages of diffusion is reduced, thereby improving the ability to preserve the structural features of underwater acoustic signals.

[0058] Step S33: Construct a Cross-Attention conditional injection mechanism Extracting the observed signal features F using a conditional encoder obs and missing mask features F mask Construct conditional features F cond The diffusion characteristic F diff As the Query, the conditional features serve as the Key and Value. The Query represents the content the current diffusion network needs to extract from the conditional information; the Key represents the index and positional relationship within the conditional information; and the Value represents the actual feature content carried within the conditional information. This is calculated using Cross-Attention. ; Where A represents the attention weight matrix; Q represents the query matrix; and K represents the key matrix. represents the transpose of the key matrix K; d represents the feature dimension of the query and key; Softmax represents the normalization function used to convert the relevance score into a probability distribution.

[0059] Obtaining conditional guided features: ; The final enhanced feature representation is formed as follows: ; This structure can establish a global correlation between the missing region and the observed region, improving the ability to model long-distance temporal dependencies.

[0060] Step S34: Design the asymmetric weighted loss function This step aims to enable the model to focus on learning the task of recovering missing regions.

[0061] Construct the global reconstruction loss: ; Further analysis of the reconstruction losses in the missing areas: ; The final joint loss is expressed as: ; Where L represents the length of the underwater acoustic signal; x i This represents the i-th sampling point of the complete real signal; i represents the sampling point index; M represents the binary missing mask matrix.

[0062] Other, This allows the model to focus on optimizing the recovery quality of missing regions.

[0063] Step S35: Construct a fast recovery mechanism based on DDIM and consistent replacement After model training, backdiffusion recovery is performed using the DDIM deterministic sampling method. A skip-sampling strategy is employed, with the number of sampling steps less than the number of training diffusion steps, preferably 20–50 steps. For each backdiffusion result x... t Perform observation replacement: ; In this model, positions with Mask=1 retain the actual observations, while positions with Mask=0 retain the model reconstruction results. High-precision and high-efficiency reconstruction is achieved through the synergistic effect of DDIM fast reconstruction and the consistency replacement mechanism.

[0064] Step S4: System Deployment and Underwater Acoustic Signal Completion Output Process Step S41: Establish the architecture of the underwater acoustic signal completion system The system mainly includes: a data input module, a missing data detection module, a data preprocessing module, a conditional Diffusion completion module, a result post-processing module, and a completion result output module. These modules communicate with each other via data interfaces.

[0065] The system comprises the following modules: a data input module to receive raw underwater acoustic data; a missing data detection module to identify signal gaps; a data preprocessing module for normalization and format conversion; a conditional Diffusion completion module for recovering missing data; a result post-processing module for restoring the original signal scale; and a completion result output module for outputting the final completion result. All modules communicate with each other via a data interface, forming a complete underwater acoustic signal completion processing flow.

[0066] Step S42: Construct the data input module The system receives data from various underwater acoustic devices, including underwater communication equipment, active sonar systems, passive sonar systems, marine environmental monitoring equipment, and underwater unmanned platforms. The system first detects the length of the input signal. If the input signal length is greater than a preset length, it uses a sliding window method for segmentation; if the input signal length is less than the preset length, it uses zero-padding to fill the gaps. Z-score normalization is then performed.

[0067] Step S43: Establish an automatic missing data detection module For scenarios with known missing locations, the system directly reads the corresponding missing mask; for scenarios with unknown missing locations, it identifies potential missing regions and constructs a missing mask matrix through methods such as outlier analysis, continuous zero value detection, energy mutation detection, and temporal continuity detection.

[0068] Step S44: Execute conditional Diffusion completion reasoning The completion module first loads the trained Conditional Diffusion model, including the parameters of the Conditional U-Net network, the parameters of the Cross-Attention module, the cosine noise scheduling parameters, and the asymmetric loss training parameters. Then, it executes the backdiffusion recovery process. First, it initializes random noise, then performs fast recovery according to the DDIM sampling mechanism established in step S35. Simultaneously, it uses a consistency replacement mechanism to keep the true data in the observation area unchanged, and continuously introduces contextual conditional information using the Cross-Attention module until the entire sampling process is completed.

[0069] Step S45: Generate a complete underwater acoustic signal After diffusion recovery, the recovered result is fused with the data from the observed area to form a complete signal. Subsequent signal smoothing and boundary correction processes are then performed to eliminate any minor discontinuities that may arise at the recovery boundaries. The system can simultaneously output time-domain waveform data, frequency-domain spectrum data, completion quality assessment indicators, and recovery error statistics for subsequent analysis and applications.

[0070] Step S46: Output the completion result and complete the application deployment. The final data completion results can be directly applied to underwater communication data recovery, sonar target detection, target identification preprocessing, marine environmental monitoring, ship radiated noise analysis, and underwater intelligent sensing systems. This invention can be deployed on high-performance server platforms, edge computing devices, embedded underwater acoustic processing terminals, unmanned underwater vehicle platforms, and underwater monitoring station systems. After system deployment, it can automatically detect, intelligently recover, and output missing underwater acoustic signal data in real time, forming a complete data completion process.

[0071] Compared with existing technologies, this invention provides a method for completing missing underwater acoustic signal data based on a conditional Diffusion model. Addressing the problems of low recovery accuracy, insufficient long-range dependency modeling ability, poor recovery of continuous missing regions, and low inference efficiency of traditional interpolation methods, statistical modeling methods, and ordinary deep learning recovery methods in complex marine environments, this invention proposes a technical solution combining multi-modal missing data modeling, a conditional Diffusion recovery network, a cosine noise scheduling strategy, a Cross-Attention conditional injection mechanism, an asymmetric weighted loss function, and a DDIM consistency replacement recovery mechanism. This approach offers the following advantages: This invention establishes a multi-modal missing data modeling mechanism for complex marine environments. By constructing various missing data modes, including random missing data, continuous missing data, and marginal missing data, it achieves realistic simulation of complex scenarios such as underwater communication link interruptions, sonar data loss, equipment failures, and cache corruption. Compared to traditional methods that only use random missing data for training, this invention can significantly improve the model's adaptability and generalization ability to different missing data modes, thereby enhancing recovery stability in practical engineering applications.

[0072] This invention employs a conditional diffusion model as the framework for underwater acoustic signal completion. It learns the latent probability distribution of the complete underwater acoustic signal through a forward diffusion process and gradually recovers information from missing regions using a backward diffusion process. Compared to traditional methods such as linear interpolation, spline interpolation, and autoregressive models, this invention can effectively learn higher-order statistical features in complex, non-stationary underwater acoustic signals, improving the realism and accuracy of missing region reconstruction and avoiding significant distortion and over-smoothing in the recovery results.

[0073] This invention constructs a diffusion recovery network based on Conditional U-Net and introduces a multi-head self-attention mechanism in the network bottleneck layer to achieve long-distance temporal dependency modeling of underwater acoustic signals. By establishing global contextual relationships, the model can simultaneously utilize structural information from the front and back sides of the missing region as well as from distant locations for recovery, effectively solving the problem of insufficient contextual information in continuous missing scenarios, thereby improving the recovery accuracy of long-term missing regions.

[0074] This invention proposes a Cross-Attention conditional injection mechanism that combines contextual information. It constructs conditional constraint information by combining observed signal features and missing mask features, and utilizes Cross-Attention to achieve dynamic interaction between diffusion features and conditional features. Compared to traditional conditional diffusion models that use simple feature concatenation or additive fusion, this invention can establish a more effective global correlation between the missing region and the observed region, improving the efficiency of utilizing known information and thus enhancing the recovery capability of complex missing regions.

[0075] This invention designs a cosine noise scheduling strategy adapted to temporal characteristics. By dynamically controlling the noise growth rate during diffusion, it preserves more original underwater acoustic signal structure information in the early stages of diffusion and gradually increases the noise disturbance intensity in the later stages. Compared with traditional linear noise scheduling methods, this invention can slow down the degradation rate of important time-frequency features, better preserve the spectral structure, periodic features, and local texture information in the underwater acoustic signal, and improve the diffusion model's ability to model complex temporal signals.

[0076] This invention constructs an asymmetric weighted loss function that focuses on missing regions. By establishing separate global reconstruction losses and local reconstruction losses for missing regions, and assigning higher weights to missing regions, the focus of model training shifts from overall error optimization to the recovery of key missing regions. Compared to the traditional uniform mean squared error loss function, this invention can significantly improve the recovery quality of continuous missing regions, large-scale missing regions, and edge missing regions, while maintaining the consistency of the overall signal structure.

[0077] This invention proposes a fast sampling recovery mechanism based on DDIM and consistency replacement. By adopting the deterministic sampling method of DDIM instead of the traditional random sampling method of DDPM, the number of inference steps in the backdiffusion process is significantly reduced, improving the model recovery efficiency. Simultaneously, the consistency replacement strategy forces the retention of true data in the observed region in each backdiffusion step, ensuring that the model only recovers and generates data for the missing regions. This effectively avoids distortion of the observed region and boundary drift problems, improving the reliability and stability of the recovery results.

[0078] The conditional diffusion completion framework proposed in this invention can simultaneously balance restoration accuracy and computational efficiency. While maintaining high-quality restoration results, it significantly reduces the number of diffusion sampling steps, lowers computational resource consumption, and improves the model's real-time processing capabilities, thereby meeting the application needs of practical engineering scenarios such as underwater communication systems, sonar detection systems, underwater unmanned platforms, and marine environmental monitoring systems.

[0079] The underwater acoustic signal recovered by this invention can better maintain the consistency of the original signal's time-domain waveform structure, frequency-domain energy distribution, and statistical characteristics, effectively improving the data quality of subsequent target recognition, signal detection, feature extraction, and intelligent sensing tasks, and providing more reliable data support for underwater acoustic information processing systems.

[0080] This invention constructs a complete underwater acoustic signal missing data completion process, realizing fully automated processing from missing data modeling, conditional diffusion recovery, intelligent completion to result output. This method has strong versatility and scalability, applicable not only to underwater communication signal recovery but also to sonar signal processing, marine monitoring data repair, underwater target detection, and other time-series signal missing data completion tasks, demonstrating broad engineering application prospects and promotional value.

[0081] It should be understood that the above embodiments are merely exemplary. Those skilled in the art can make various non-substantial improvements and substitutions based on the technical concept of this invention, and all such improvements and substitutions should be considered to fall within the protection scope of this application.

Claims

1. A method for completing missing underwater acoustic signal data based on a conditional diffusion model, characterized in that, Includes the following steps: Step S1: Obtain complete underwater acoustic signal samples, generate missing observation signals and corresponding missing masks according to multiple missing modes and corresponding missing parameters, and construct a training sample set containing the complete underwater acoustic signal, the missing observation signal and the missing mask; Step S2: Construct a conditional diffusion model, which includes a forward diffusion process and a backward diffusion process. The forward diffusion process is used to convert the complete underwater acoustic signal into a noise distribution by gradually adding noise, and the backward diffusion process is used to gradually remove noise from the noise distribution to recover the signal. Step S3: Train the conditional diffusion model using the training sample set. During training, the missing observation signal and the missing mask are used as conditional information to guide the back diffusion process. Step S4: Input the missing underwater acoustic signal to be recovered and its corresponding missing mask into the trained conditional diffusion model, execute the reverse diffusion process, use the conditional information to guide the denoising at each step, and output the complete underwater acoustic signal.

2. The method according to claim 1, characterized in that, In step S1, the multiple missing patterns include random missing patterns, consecutive missing patterns, and marginal missing patterns; The random missing mode involves randomly selecting sampling points in the signal as missing locations according to a preset missing ratio, where the preset missing ratio ranges from 10% to 50%. The continuous missing pattern involves selecting a continuous time interval in the signal for missing data, where the length of the missing time interval accounts for 10% to 70% of the total signal length. The edge missing pattern involves constructing a continuous missing region at the beginning or end of the signal, wherein the length of the continuous missing region accounts for 10% to 30% of the total length of the signal.

3. The method according to claim 2, characterized in that, In step S1, the complete underwater acoustic signal sample comes from data collected by underwater communication equipment, sonar detection system, underwater target monitoring platform and marine environmental monitoring equipment; The underwater acoustic signal includes at least one of the following: active sonar echo signal, passive sonar listening signal, underwater communication modulation signal, marine environmental background noise signal, and underwater target radiated noise signal. The Z-score normalization method is used to normalize the segmented signal so that the mean is 0 and the variance is 1.

4. The method according to claim 1, characterized in that, In step S2, the conditional diffusion model uses Conditional U-Net as the noise prediction network; The Conditional U-Net includes a four-level encoder, an intermediate bottleneck layer, a four-level decoder, and skip connections; The encoder is used to progressively extract temporal features at different scales, the decoder is used to progressively recover the signal structure, and the skip connection is used to fuse the shallow detail information of the encoder with the deep semantic information of the decoder. The bottleneck layer introduces a multi-head self-attention mechanism to establish long-distance temporal dependencies of underwater acoustic signals.

5. The method according to claim 4, characterized in that, In step S2, the forward diffusion process employs a cosine noise scheduling strategy to construct the cumulative retention coefficient: ； in: ； Cosine scheduling can be used to slow down the rate of information loss in the early stages of diffusion.

6. The method according to claim 5, characterized in that, In step S2, the missing observation signal and the missing mask are used as conditional information to guide the backdiffusion process, specifically including: The missing observation signal and the missing mask are feature-encoded to obtain the conditional feature F. cond ; The diffusion characteristic F in the reverse diffusion process diff As a query, the condition feature F cond As Key and Value, they are fused using a cross-attention mechanism: ； Obtaining conditional guided features: ； The final enhanced feature representation is formed as follows: ； Where A represents the attention weight matrix; Q represents the query matrix; and K represents the key matrix; represents the transpose of the key matrix K; d represents the feature dimension of the query and key; Softmax represents the normalization function used to convert the relevance score into a probability distribution.

7. The method according to claim 1, characterized in that, In step S3, the loss function for training the conditional diffusion model is an asymmetric weighted loss function, which includes a global reconstruction loss term and a missing region reconstruction loss term. The global reconstruction loss is expressed as: ； The reconstruction loss of the missing region is expressed as: ； The joint loss is expressed as: ； in, .

8. The method according to claim 1, characterized in that, In step S4, the back diffusion process is executed, specifically using the DDIM deterministic sampling method, and the number of back diffusion steps is reduced to 20 to 50 steps through a skip sampling strategy; In each step of the reverse diffusion process, a consistent replacement operation is performed: ； In this context, the position with Mask=1 retains the actual observation value, while the position with Mask=0 retains the model recovery result.

9. The method according to claim 8, characterized in that, In step S4, for the signal to be recovered at an unknown missing location, the potential missing region is identified by at least one of the following methods: outlier analysis, continuous zero detection, energy mutation detection, and temporal continuity detection, and a corresponding missing mask is generated.

10. A system for completing missing underwater acoustic signal data based on a conditional diffusion model, characterized in that, include: The data preparation module is used to acquire complete underwater acoustic signal samples, generate missing observation signals and corresponding missing masks according to multiple missing modes and corresponding missing parameters, and construct a training sample set containing the complete underwater acoustic signals, the missing observation signals and the missing masks. The model building module is used to build a conditional diffusion model, which includes a forward diffusion process and a backward diffusion process. The forward diffusion process is used to convert the complete underwater acoustic signal into a noise distribution by gradually adding noise, and the backward diffusion process is used to gradually remove noise from the noise distribution to recover the signal. The model training module is used to train the conditional diffusion model using the training sample set. During training, the missing observation signal and the missing mask are used as conditional information to guide the back diffusion process. The signal recovery module is used to input the missing underwater acoustic signal to be recovered and its corresponding missing mask into the trained conditional diffusion model, execute the back diffusion process, use the conditional information to guide the denoising at each step, and output a complete and supplemented underwater acoustic signal.

Images