Communication adaptive adjustment method and system based on underwater acoustic communication frame blank time slot

Abstract

The invention discloses a communication adaptive adjustment method and system based on an underwater acoustic communication frame blank time slot, and the method comprises the steps: obtaining an underwater acoustic communication frame, and collecting environment noise data based on the blank time slot of the underwater acoustic communication frame; performing time-frequency transformation on the environmental noise data to generate a time-frequency characteristic graph; the time-frequency feature map is input into a residual hybrid expert network, and the residual hybrid expert network is obtained by adopting joint loss function training and comprises a shared backbone network, a gating network and a plurality of expert networks; extracting a high-dimensional feature vector from the time-frequency feature map through a shared backbone network; inputting the high-dimensional feature vector into a gating network, and outputting a hard routing selection result by the gating network, which is used for indicating to activate a certain expert network; according to a hard routing selection result, inputting the high-dimensional feature vector into a corresponding expert network for processing to obtain an environment state classification result; and acquiring an underwater acoustic communication physical layer processing strategy matched with the communication environment according to an environment state sensing classification result, and acting on the current underwater acoustic communication frame.

Description

A Communication Adaptive Adjustment Method and System Based on Underwater Acoustic Communication Frame Blank Time Slots Technical Field

[0001] This invention relates to the field of underwater acoustic communication technology, and in particular to a communication adaptive adjustment method and system based on the time slot of the underwater acoustic communication frame blanking. Background Technology

[0002] The highly dynamic marine environment is a key bottleneck restricting the performance of underwater acoustic communication. In particular, the severe environmental noise generated by changes in sea conditions such as wind, waves, and rainfall can cause the characteristics of the underwater acoustic channel to change rapidly over time, seriously interfering with the stability of the communication link.

[0003] Existing underwater acoustic communication systems typically employ two methods to cope with changes in sea state: the first relies on external equipment such as surface weather buoys to acquire environmental data and transmit it to underwater nodes; the second treats environmental noise as interference and suppresses it from a signal processing perspective. The first method requires additional equipment deployment, resulting in strong system dependence and significant information transmission delays; the second method focuses on interference resistance but struggles to output environmental parameters for adaptive communication adjustments.

[0004] In recent years, research on retrieving sea surface environmental parameters based on underwater noise has made some progress. The literature [Taylor et al., "Machine learning methods to approximate rainfall and wind from acoustic underwater measurements", IEEE TGRS, 2021] uses hydrophones to collect long-term underwater acoustic data and establishes a mapping relationship between acoustic features and environmental parameters through power spectral density analysis and machine learning algorithms. The literature [Trucco et al., "Introducing Temporal Correlation in Rainfall and Wind Prediction From Underwater Noise", IEEE JOE, 2023] further introduces a long short-term memory network to improve prediction accuracy using temporal correlation.

[0005] However, the solutions described in the above literature have significant drawbacks if directly applied to underwater acoustic communication systems: on the one hand, the blank time slots in efficient communication protocols are usually only tens of milliseconds, which cannot meet the requirements of the above solutions for long-term continuous sampling; on the other hand, highly complex models such as long short-term memory networks have large latency when running on underwater nodes with limited computing resources, making it difficult to achieve millisecond-level real-time perception and unable to guide the instantaneous adjustment of communication parameters in the next frame.

[0006] In summary, existing technologies lack methods for rapidly achieving environmental perception and adaptively adjusting communication based on the perception results. Summary of the Invention

[0007] To enable rapid environmental perception and adaptive communication adjustment based on the perception results, this invention provides a method for adaptive communication adjustment based on the blank time slots of underwater acoustic communication frames, which includes the following steps:

[0008] Acquire underwater acoustic communication frames and collect environmental noise data based on the blank time slots of the underwater acoustic communication frames; perform time-frequency transformation on the environmental noise data to generate a time-frequency feature map; input the time-frequency feature map into a residual hybrid expert network, which is trained using a joint loss function and includes a shared backbone network, a gating network, and multiple expert networks; the joint loss function includes classification loss and load balancing loss; extract high-dimensional feature vectors from the time-frequency feature map through the shared backbone network; input the high-dimensional feature vectors into the gating network, which outputs a hard routing selection result to indicate the activation of a certain expert network; according to the hard routing selection result, input the high-dimensional feature vectors into the corresponding expert network for processing to obtain an environmental state classification result; obtain an underwater acoustic communication physical layer processing strategy matching the communication environment based on the environmental state perception classification result and apply it to the current underwater acoustic communication frame.

[0009] Optionally, performing time-frequency transformation on the environmental noise data to generate a time-frequency feature map includes: performing frame segmentation and windowing processing on the environmental noise data to obtain a frame sequence; performing a short-time Fourier transform on the frame sequence to obtain the spectrum of each frame; stacking the spectra of each frame in chronological order to obtain an STFT spectrum map; or, filtering the spectrum of each frame through a Mel filter bank or a Barker filter bank and taking the logarithm to obtain the nonlinear scaling features of each frame, and then stacking the nonlinear scaling features in chronological order to obtain a nonlinear scaling spectrum map; or, convolving the spectrum of each frame with a CQT kernel to obtain the CQT features of each frame, and then stacking the CQT features in chronological order to obtain a CQT spectrum map; using the STFT spectrum map, the nonlinear scaling spectrum map, or the CQT spectrum map as the time-frequency feature map.

[0010] Optionally, the shared backbone network includes convolutional layers, normalization layers, activation function layers, a first pooling layer, multiple residual layers, and a second pooling layer.

[0011] Optionally, a high-dimensional feature vector can be extracted from the time-frequency feature map through the shared backbone network. Specifically, the time-frequency feature map is subjected to feature extraction, dimensionality reduction, and flattening operations to obtain a high-dimensional feature vector.

[0012] Optionally, the high-dimensional feature vector is input into the gating network, and the gating network outputs a hard routing result, including: inputting the high-dimensional feature vector into the gating network, and after the high-dimensional feature vector is linearly transformed, obtaining the routing score of each expert network; after the routing score of each expert network is mapped by an activation function, obtaining the confidence score of each expert network; selecting the index of the expert network corresponding to the maximum confidence score as the hard routing result, and outputting it.

[0013] Optionally, the load balancing loss is calculated based on the dot product of the sample allocation ratio of each expert network in the current training batch and the average routing probability, and is used to constrain the number of samples allocated to each expert network to tend to be balanced.

[0014] Optionally, the output dimension of the expert network is the same as the number of preset environmental state categories; the environmental state categories include rainfall levels of different intensities, and the rainfall levels include at least two of the following: no rain, light rain, moderate rain, heavy rain, and rainstorm.

[0015] Optionally, after acquiring environmental noise data based on the blank time slots of the underwater acoustic communication frame, the method further includes: segmenting the original audio data within the blank time slots based on the duration constraint of the blank time slots to obtain audio segments; performing spectral analysis on the audio segments and suppressing narrowband interference components through filtering.

[0016] Corresponding to the aforementioned adaptive communication adjustment method based on the blank time slots of underwater acoustic communication frames, this invention provides an adaptive communication adjustment system based on the blank time slots of underwater acoustic communication frames, comprising: an environmental noise data acquisition module for acquiring underwater acoustic communication frames and acquiring environmental noise data based on the blank time slots of the underwater acoustic communication frames; a time-frequency transformation module for performing time-frequency transformation on the environmental noise data, generating a time-frequency feature map and inputting it into a residual hybrid expert network; and a residual hybrid expert network comprising: a shared backbone network for extracting high-dimensional feature vectors from the time-frequency feature map and inputting them into the gated network; and a gated network, which... The system is used to output a hard routing selection result based on the high-dimensional feature vector, which is used to indicate the activation of a certain expert network; multiple expert networks are respectively connected to the gating network, wherein the activated expert network is used to process the high-dimensional feature vector and output an environmental state classification result; the residual hybrid expert network is trained using a joint loss function, which includes classification loss and load balancing loss; the communication adaptive adjustment module is used to obtain an underwater acoustic communication physical layer processing strategy that matches the communication environment based on the environmental state perception classification result and apply it to the current underwater acoustic communication frame.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) Environmental noise data is collected by using the blank time slot of the underwater acoustic communication frame, and environmental information can be obtained without relying on external equipment such as sea surface meteorological buoys; by performing time-frequency transformation on the environmental noise data to generate time-frequency feature maps, the noise data collected in a short time is converted into feature representations suitable for deep learning processing, overcoming the limitation that the blank time slot is only tens of milliseconds and cannot meet the long-term continuous sampling requirements; the residual hybrid expert network adopts a design architecture of shared backbone network, gated network and multiple expert networks, and combined with the hard routing selection mechanism, only activates a single expert network for inference, which significantly reduces the computational complexity of the model, enabling underwater nodes with limited computing resources to achieve millisecond-level real-time environmental perception; the underwater acoustic communication physical layer processing strategy matching the communication environment is obtained according to the environmental state perception classification results, and the environmental noise is transformed from interference items into a processing strategy that can guide the adaptive adjustment of the underwater acoustic communication physical layer configuration, and applied to the current underwater acoustic communication frame, thereby realizing the adaptive adjustment of the underwater acoustic communication physical layer configuration based on the perception results. As can be seen, the present invention solves the technical problems in the prior art, such as reliance on external devices, the contradiction between long-term sampling requirements and short-term blank time slots, large inference delays in high-complexity models, and difficulty in outputting environmental parameters for adaptive adjustment of communication.

[0018] (2) By performing frame segmentation and windowing on the environmental noise data, a frame sequence is obtained. The frame sequence is then subjected to short-time Fourier transform to obtain the STFT spectrum, or filtered by the Mel filter bank / Barker filter bank to obtain the nonlinear scale spectrum, or convolved with the CQT kernel to obtain the CQT spectrum. This achieves multi-scale feature extraction from the original audio data to the time-frequency feature map, and can capture acoustic features with different frequency resolutions from short-time noise data, thus improving the effective characterization ability of environmental noise features under the condition of limited blank time slots.

[0019] (3) The shared backbone network adopts a hierarchical structure design of convolutional layer, normalization layer, activation function layer, first pooling layer, multiple residual layers and second pooling layer, which realizes deep feature extraction and dimensionality reduction of time-frequency feature map. The introduction of residual layer alleviates the gradient vanishing problem in deep network training and improves the stability and accuracy of feature extraction.

[0020] (4) By performing a linear transformation on the high-dimensional feature vector through a gated network to obtain the routing score, and then obtaining the confidence score through the activation function mapping, the expert network index corresponding to the maximum value is selected as the hard routing selection result. This realizes a sparse routing mechanism that activates only a single expert network. Compared with the soft routing method that activates all expert networks, this significantly reduces the amount of inference computation and improves real-time performance.

[0021] (5) The load balancing loss is calculated based on the dot product of the sample allocation ratio of each expert network in the current training batch and the average routing probability. This constrains the number of samples allocated to each expert network to tend to be balanced, avoiding the problem that the gated network tends to allocate samples to a few expert networks, resulting in insufficient training of other expert networks. This improves the overall generalization ability and classification accuracy of the residual hybrid expert network.

[0022] (6) The output dimension of the expert network is set to be the same as the number of preset environmental state categories, and the environmental state categories include at least one of different intensities of rainfall (no rain, light rain, moderate rain, heavy rain, and rainstorm), which realizes the fine classification and identification of rainfall, a key environmental parameter in the marine environment, and provides a specific decision basis for obtaining the physical layer processing strategy of underwater acoustic communication.

[0023] (7) By acquiring the underwater acoustic communication physical layer processing strategy that matches the communication environment and applying it to the current underwater acoustic communication frame, the environmental perception results are quickly fed back to the underwater acoustic communication physical layer configuration for adaptive adjustment, ensuring the underwater acoustic communication physical layer's reception and demodulation processing of underwater acoustic signals, thus improving the reliability of the underwater acoustic communication link.

[0024] (8) Based on the duration constraint of the blank time slot, the original audio data is segmented to obtain audio segments, and the audio segments are subjected to spectrum analysis and filtering to suppress narrowband interference components. This realizes the extraction of effective noise data within the blank time slot of limited duration. At the same time, the filtering process suppresses the influence of narrowband interference on noise analysis, thereby improving the quality of environmental noise data and the accuracy of environmental status classification. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of the invention and constitute a part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 is a simplified flowchart of an embodiment of the adaptive adjustment method for communication based on the blank time slots of underwater acoustic communication frames according to the present invention; Figure 2 is a schematic diagram of the residual hybrid expert network architecture of an embodiment of the adaptive adjustment method for communication based on the blank time slots of underwater acoustic communication frames according to the present invention; Figure 3 is a schematic diagram of experimental results of different time-frequency characteristic maps in the residual hybrid expert network of an embodiment of the adaptive adjustment method for communication based on the blank time slots of underwater acoustic communication frames according to the present invention; Figure 4 is a framework diagram of an embodiment of the adaptive adjustment system for communication based on the blank time slots of underwater acoustic communication frames according to the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. 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.

[0027] As shown in Figure 1, the present invention provides a communication adaptive adjustment method based on the blank time slots of underwater acoustic communication frames, comprising the following steps: acquiring underwater acoustic communication frames and collecting environmental noise data based on the blank time slots of the underwater acoustic communication frames; performing time-frequency transformation on the environmental noise data to generate a time-frequency feature map; inputting the time-frequency feature map into a residual hybrid expert network, which is trained using a joint loss function and includes a shared backbone network, a gating network, and multiple expert networks; the joint loss function includes classification loss and load balancing loss; extracting high-dimensional feature vectors from the time-frequency feature map through the shared backbone network; inputting the high-dimensional feature vectors into the gating network, which outputs a hard routing selection result, which is used to indicate the activation of a certain expert network; according to the hard routing selection result, inputting the high-dimensional feature vectors into the corresponding expert network for processing to obtain an environmental state classification result; obtaining an underwater acoustic communication physical layer processing strategy matching the communication environment based on the environmental state perception classification result and applying it to the current underwater acoustic communication frame.

[0028] This invention utilizes blank slots in underwater acoustic communication frames to collect environmental noise data, enabling the acquisition of environmental information without relying on external equipment such as surface weather buoys. By performing time-frequency transformation on the environmental noise data to generate time-frequency feature maps, the short-term noise data is converted into feature representations suitable for deep learning processing, overcoming the limitation that blank slots, which are only tens of milliseconds long, cannot meet the requirements for long-term continuous sampling. The residual hybrid expert network adopts a design architecture of shared backbone network, gated network, and multiple expert networks, combined with a hard routing selection mechanism that activates only a single expert network for inference, significantly reducing the computational complexity of the model and enabling underwater nodes with limited computing resources to achieve millisecond-level real-time environmental perception. Based on the environmental state perception classification results, an underwater acoustic communication physical layer processing strategy matching the communication environment is obtained, transforming environmental noise from an interference term into a processing strategy that can guide the adaptive adjustment of the underwater acoustic communication physical layer configuration, and applying it to the current underwater acoustic communication frame, thereby realizing adaptive adjustment of the underwater acoustic communication physical layer configuration based on the perception results. As can be seen, the present invention solves the technical problems in the prior art, such as reliance on external devices, the contradiction between long-term sampling requirements and short-term blank time slots, large inference delays in high-complexity models, and difficulty in outputting environmental parameters for adaptive adjustment of communication.

[0029] In this embodiment, time-frequency transformation is performed on environmental noise data to generate a time-frequency feature map, including: performing frame segmentation and windowing processing on the environmental noise data to obtain a frame sequence; performing a short-time Fourier transform on the frame sequence to obtain the spectrum of each frame; stacking the spectra of each frame in chronological order to obtain an STFT spectrum map; or, filtering the spectrum of each frame through a Mel filter bank or a Bark filter bank and taking the logarithm to obtain the nonlinear scaling features of each frame, and then stacking the nonlinear scaling features in chronological order to obtain a nonlinear scaling spectrum map; or, convolving the spectrum of each frame with a CQT kernel to obtain the CQT features of each frame, and then stacking the CQT features in chronological order to obtain a CQT spectrum map; using the STFT spectrum map, the nonlinear scaling spectrum map, or the CQT spectrum map as the time-frequency feature map.

[0030] The Mel / Barker filter bank has a higher density in the low-frequency region to achieve higher frequency resolution. To simulate the non-linear perception of sound frequencies by the human ear, this invention uses the Mel and Barker scales to map linear frequencies; the corresponding conversion formulas are as follows: ;in, Indicates Mel frequency, It represents linear frequency (unit: Hz) and is the basic frequency measurement unit for each frame of the spectrum; ;in, This indicates the Buck frequency.

[0031] The CQT core consists of a set of bandpass filters spaced logarithmically across frequencies, and its spectrum exhibits higher time resolution in the high-frequency region. The conversion formula corresponding to the CQT core used in this invention is as follows: , ;in, Represents CQT features, Indicates the first The center frequency of the triangular filter, Indicates the minimum frequency to be processed. Indicates the maximum frequency to be processed. This indicates octave resolution.

[0032] This invention obtains a frame sequence by framing and windowing environmental noise data, and then performs a short-time Fourier transform on the frame sequence to obtain an STFT spectrum, or filters it with a Mel filter bank / Barker filter bank to obtain a nonlinear scale spectrum, or convolves it with a CQT kernel to obtain a CQT spectrum. This enables multi-scale feature extraction from raw audio data to time-frequency feature maps, and can capture acoustic features with different frequency resolutions from short-time noise data, thus improving the effective characterization capability of environmental noise features under space-constrained conditions.

[0033] As shown in Figure 2, in this embodiment, the shared backbone network uses an improved ResNet-18 as the backbone to extract high-dimensional feature vectors. The shared backbone network specifically includes convolutional layers, normalization layers, activation function layers, a first pooling layer (3x3 pooling), multiple residual layers (base blocks), and a second pooling layer (global average pooling).

[0034] In this embodiment, the first convolutional layer of ResNet is reconstructed. The standard ResNet's 3-channel input is modified to a 1-channel input, the number of output channels is 64, the kernel size is 7x7, the stride is 2, and the padding is 3.

[0035] In this embodiment, a high-dimensional feature vector is extracted from the time-frequency feature map by sharing a backbone network. Specifically, the time-frequency feature map is subjected to feature extraction, dimensionality reduction, and flattening operations to obtain a high-dimensional feature vector. Preferably, after the residual layer, an adaptive average pooling layer is used to fix the feature map size to (1,1) and flatten it to obtain a high-dimensional feature vector with a dimension of 512.

[0036] The shared backbone network of this invention adopts a hierarchical structure design of convolutional layers, normalization layers, activation function layers, first pooling layers, multiple residual layers (preferably 4), and second pooling layers, which realizes deep feature extraction and dimensionality reduction of time-frequency feature maps. The introduction of residual layers alleviates the gradient vanishing problem in deep network training and improves the stability and accuracy of feature extraction.

[0037] In this embodiment, a high-dimensional feature vector is input into a gating network, and the gating network outputs a hard routing result. This includes: inputting the high-dimensional feature vector into the gating network, and after a linear transformation, obtaining the routing score for each expert network, as shown in the following formula: ;in, Indicates the route score; The learnable weight matrix of the gated network has dimensions of . , Represents the set of real numbers. This represents the dimension of the high-dimensional feature vector extracted and flattened by the shared backbone network. This represents the total number of expert networks. Represents the composition of real numbers OK A matrix of columns; Represents a high-dimensional feature vector; This represents the learnable bias vector of the gated network, with dimension . , Indicates length is A real row vector.

[0038] The routing scores of each expert network are mapped by an activation function to obtain the confidence level of each expert network. Preferably, the sigmoid activation function is used to map the routing scores to the (0, 1) interval, which represents the confidence level of each expert network. The index of the expert network corresponding to the maximum confidence level is selected as the hard routing selection result and output.

[0039] This invention obtains routing scores by performing linear transformation on high-dimensional feature vectors through a gated network, and then obtains confidence scores by mapping through an activation function. Finally, it selects the expert network index corresponding to the maximum value as the hard routing selection result, thus realizing a sparse routing mechanism that activates only a single expert network. Compared with the soft routing method that activates all expert networks, this significantly reduces the amount of inference computation and improves real-time performance.

[0040] In this embodiment, to prevent some expert networks from overfitting while others remain unutilized during training, a joint loss function is designed for training. The classification loss uses the cross-entropy loss function to measure the difference between the predicted class and the true label, as shown in the following formula: ;in, express, Indicates the size of the training batch. The number of categories representing rainfall levels. This indicates the index of the sample in the current batch. Category index representing rainfall levels, This indicates that the residual hybrid expert network predicts the first... Each sample belongs to The probability, Indicator variable representing the true label (One-hot encoding), if the first... The true category of each sample is ,but ,otherwise .

[0041] In this embodiment, the load balancing loss is calculated based on the dot product of the sample allocation ratio of each expert network in the current training batch and the average routing probability. This is used to constrain the number of samples allocated to each expert network to tend towards a balanced distribution. The specific calculation formula is as follows: ;in, This indicates load balancing losses. The weighting coefficients representing the balancing loss. Indicates the number of expert networks. Indicates the first A network of experts, Indicates the number actually allocated in the current batch. The sample proportion of each expert network (obtained by averaging the one-hot encoded expert network IDs). Indicates the first in the current batch The average routing probability of an expert network (obtained by averaging the routing scores).

[0042] The load balancing loss of this invention is calculated based on the dot product of the sample allocation ratio of each expert network in the current training batch and the average routing probability. This constrains the number of samples allocated to each expert network to tend to be balanced, avoiding the problem that gating networks tend to allocate samples to a few expert networks, resulting in insufficient training of other expert networks. This improves the overall generalization ability and classification accuracy of residual hybrid expert networks.

[0043] In this embodiment, the output dimension of the expert network is the same as the number of preset environmental state categories; the environmental state categories include rainfall levels of different intensities, with at least two of the following rainfall levels: no rain, light rain, moderate rain, heavy rain, and torrential rain. Traditional ResNet uses a single shared fully connected layer to process radiated underwater acoustic noise under all rainfall intensities, but it is difficult to take into account the differences in feature distribution of radiated underwater acoustic noise under light rain and torrential rain conditions. This invention, however, achieves feature decoupling by dynamically selecting the most suitable linear projection space through hard routing.

[0044] The output dimension of the expert network in this invention is set to be the same as the number of preset environmental state categories, and the environmental state categories include at least one of different intensities of rainfall (no rain, light rain, moderate rain, heavy rain, and rainstorm). This enables refined classification and identification of rainfall, a key environmental parameter in the marine environment, providing a specific decision-making basis for obtaining the physical layer processing strategy for underwater acoustic communication. Figure 3 shows the experimental results of different time-frequency feature maps in the residual hybrid expert network. As can be seen from Figure 3, the environmental perception accuracy of the corresponding features of each time-frequency feature map in the residual hybrid expert network is as high as 80% or more. Obviously, this invention can not only quickly realize environmental perception, but also ensure the accuracy of environmental perception.

[0045] In this embodiment, a pre-defined mapping rule between environmental state classification results and underwater acoustic communication physical layer processing strategies is used. Preferably, the underwater acoustic communication physical layer processing strategy is preset for the environmental state classification results, and can be specifically set according to actual needs. In one implementation, when the environmental state classification result indicates a high rainfall level and degraded signal-to-noise ratio, the corresponding underwater acoustic communication physical layer processing strategy guides the receiver to switch or select an underwater acoustic signal processing method suitable for low signal-to-noise ratio conditions for signal detection and demodulation. Conversely, when the environmental state classification result indicates a relatively good environmental state (e.g., low rainfall level), the underwater acoustic communication physical layer processing strategy employs a relatively low-complexity or high-efficiency underwater acoustic signal processing method. Furthermore, after parsing the underwater acoustic communication physical layer processing strategy, the receiver applies it to the current underwater acoustic communication frame to adaptively adjust the corresponding physical layer configuration, thereby improving the reliability and real-time performance of the underwater acoustic communication system in complex marine environments without increasing additional interaction overhead.

[0046] It is understandable that each underwater acoustic communication physical layer processing strategy is specifically an underwater acoustic signal processing method that matches the classification result of a certain environmental state. This method has a corresponding physical layer configuration preset in order to adaptively adjust the corresponding physical layer configuration according to the classification result of the environmental state.

[0047] This invention obtains a underwater acoustic communication physical layer processing strategy that matches the communication environment and applies it to the current underwater acoustic communication frame. This allows for adaptive adjustment of the corresponding physical layer configuration based on the underwater acoustic communication physical layer processing strategy, thereby achieving rapid feedback of environmental perception results to the underwater acoustic communication physical layer configuration adjustment. This ensures the underwater acoustic communication physical layer's reception and demodulation of underwater acoustic signals, improving the reliability of the underwater acoustic communication link.

[0048] In this embodiment, after collecting environmental noise data based on the blank time slots of the underwater acoustic communication frame, the method further includes: segmenting the original audio data within the blank time slot based on the duration constraint of the blank time slot to obtain audio segments, thereby simulating the noise signal collected from the blank time slots of the underwater acoustic communication frame; and labeling and classifying the rainfall radiation underwater acoustic data in combination with the characteristics of underwater acoustic communication applications and on-site observation records; for example, the duration constraint is 0.5 seconds and the overlap ratio of adjacent segments is 50%.

[0049] Spectral analysis is performed on audio clips, and narrowband interference introduced by electronic devices is suppressed through filtering.

[0050] This invention segments the original audio data into audio segments based on the duration constraint of blank time slots, and performs spectral analysis and filtering on the audio segments to suppress narrowband interference components. This enables the extraction of effective noise data within a limited duration of blank time slots. At the same time, the filtering process suppresses the impact of narrowband interference on noise analysis, thereby improving the quality of environmental noise data and the accuracy of environmental state classification.

[0051] As shown in Figure 4, the present invention also provides a communication adaptive adjustment system based on the blank time slots of underwater acoustic communication frames, which includes: an environmental noise data acquisition module 10, used to acquire underwater acoustic communication frames and acquire environmental noise data based on the blank time slots of the underwater acoustic communication frames; a time-frequency transformation module 20, used to perform time-frequency transformation on the environmental noise data, generate a time-frequency feature map and input it into a residual hybrid expert network 30; the residual hybrid expert network 30 includes: a shared backbone network 31, used to extract high-dimensional feature vectors from the time-frequency feature map and input them into a gating network 32; the gating network 32, used to output hard routing selection results based on the high-dimensional feature vectors, the hard routing selection results being used to indicate the activation of a first expert network 33a or a second expert network 33b. The network consists of either a third expert network 33c or a fourth expert network 33d; a first expert network 33a, a second expert network 33b, a third expert network 33c, and a fourth expert network 33d, respectively connected to the gating network. The activated first expert network 33a, the second expert network 33b, the third expert network 33c, or the fourth expert network 33d is used to process the high-dimensional feature vector and output the environmental state classification result. The residual hybrid expert network 30 is trained using a joint loss function, which includes classification loss and load balancing loss. The communication adaptive adjustment module 40 is used to obtain the underwater acoustic communication physical layer processing strategy that matches the communication environment based on the environmental state perception classification result and apply it to the current underwater acoustic communication frame.

[0052] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0053] Furthermore, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0054] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A communication adaptive adjustment method based on the time slot gap in underwater acoustic communication frames, characterized in that, Includes the following steps: Acquire underwater acoustic communication frames and collect environmental noise data based on the blank time slots of the underwater acoustic communication frames; perform time-frequency transformation on the environmental noise data to generate a time-frequency feature map; input the time-frequency feature map into a residual hybrid expert network, which is trained using a joint loss function and includes a shared backbone network, a gating network, and multiple expert networks; the joint loss function includes classification loss and load balancing loss; extract high-dimensional feature vectors from the time-frequency feature map through the shared backbone network; input the high-dimensional feature vectors into the gating network, which outputs a hard routing selection result to indicate the activation of a certain expert network; according to the hard routing selection result, input the high-dimensional feature vectors into the corresponding expert network for processing to obtain an environmental state classification result; obtain an underwater acoustic communication physical layer processing strategy matching the communication environment based on the environmental state perception classification result and apply it to the current underwater acoustic communication frame.

2. The adaptive adjustment method for communication based on the blank time slot of underwater acoustic communication frames according to claim 1, characterized in that, The process of performing time-frequency transformation on the environmental noise data to generate a time-frequency feature map includes: performing frame segmentation and windowing on the environmental noise data to obtain a frame sequence; performing a short-time Fourier transform on the frame sequence to obtain the spectrum of each frame; stacking the spectra of each frame in chronological order to obtain an STFT spectrum map; or, filtering the spectrum of each frame through a Mel filter bank or a Barker filter bank and taking the logarithm to obtain the nonlinear scaling features of each frame, and then stacking the nonlinear scaling features in chronological order to obtain a nonlinear scaling spectrum map; or, convolving the spectrum of each frame with a CQT kernel to obtain the CQT features of each frame, and then stacking the CQT features in chronological order to obtain a CQT spectrum map; and using the STFT spectrum map, the nonlinear scaling spectrum map, or the CQT spectrum map as the time-frequency feature map.

3. The adaptive adjustment method for communication based on the time slot gap in underwater acoustic communication frames according to claim 1, characterized in that, The shared backbone network includes convolutional layers, normalization layers, activation function layers, a first pooling layer, multiple residual layers, and a second pooling layer.

4. The adaptive adjustment method for communication based on the time slot gap in underwater acoustic communication frames according to claim 1, characterized in that, Through the shared backbone network, high-dimensional feature vectors are extracted from the time-frequency feature map. Specifically, the time-frequency feature map is subjected to feature extraction, dimensionality reduction, and flattening operations to obtain high-dimensional feature vectors.

5. The adaptive adjustment method for communication based on the blank time slot of underwater acoustic communication frames according to claim 1, characterized in that, The process of inputting the high-dimensional feature vector into the gating network and outputting a hard routing result includes: inputting the high-dimensional feature vector into the gating network; after linear transformation, obtaining the routing score of each expert network; mapping the routing score of each expert network through an activation function to obtain the confidence score of each expert network; selecting the index of the expert network corresponding to the maximum confidence score as the hard routing result and outputting it.

6. The adaptive adjustment method for communication based on the time slot gap in underwater acoustic communication frames according to claim 1, characterized in that, The load balancing loss is calculated based on the dot product of the sample allocation ratio of each expert network in the current training batch and the average routing probability, and is used to constrain the number of samples allocated to each expert network to tend to be balanced.

7. The adaptive adjustment method for communication based on the time slot gap in underwater acoustic communication frames according to claim 1, characterized in that, The output dimension of the expert network is the same as the number of preset environmental state categories; the environmental state categories include different intensities of rainfall levels, and the rainfall levels include at least two of the following: no rain, light rain, moderate rain, heavy rain, and rainstorm.

8. The adaptive adjustment method for communication based on the time slot gap in underwater acoustic communication frames according to claim 1, characterized in that, After acquiring environmental noise data based on the blank time slots of the underwater acoustic communication frame, the method further includes: segmenting the original audio data within the blank time slots based on the duration constraint of the blank time slots to obtain audio segments; performing spectral analysis on the audio segments and suppressing narrowband interference components through filtering.

9. A communication adaptive adjustment system based on the time slot gap of underwater acoustic communication frames, characterized in that, include: An environmental noise data acquisition module is used to acquire underwater acoustic communication frames and collect environmental noise data based on the blank time slots of the underwater acoustic communication frames. A time-frequency transformation module is used to perform time-frequency transformation on the environmental noise data, generate a time-frequency feature map, and input it into a residual hybrid expert network. The residual hybrid expert network includes: a shared backbone network, which is used to extract high-dimensional feature vectors from the time-frequency feature map and input them into a gating network; a gating network, which is used to output a hard routing selection result based on the high-dimensional feature vectors, the hard routing selection result being used to indicate the activation of a certain expert network; and multiple expert networks respectively connected to the gating network, wherein the activated expert network is used to process the high-dimensional feature vectors and output an environmental state classification result. The residual hybrid expert network is trained using a joint loss function, the joint loss function including classification loss and load balancing loss. A communication adaptive adjustment module is used to obtain a hydroacoustic communication physical layer processing strategy matching the communication environment based on the environmental state perception classification result and apply it to the current hydroacoustic communication frame.

Images