Adaptive compression coding method and system for ocean buoy acoustic data

Abstract

The application discloses a kind of adaptive compression coding methods and systems for ocean buoy acoustic data.The method includes two parts of buoy end compression encoding and receiving end post-processing reconstruction.Buoy end blocks acoustic signal and carries out silence discrimination, and performs discrete wavelet transform on active block, optimally retains important coefficient under code rate budget constraint and quantizes and entropy encodes according to sub-band layering to generate compressed code stream, and is returned through wireless link.The receiving end is decoded and reconstructed, and the wavelet coefficient is recombined according to the sub-band by sub-band priority rearrangement, input double-domain cascade neural network is sequentially repaired by wavelet domain subnet and refined by time domain subnet, and physical constraint hard projection defined by quantization half step and discarded amplitude upper limit is applied, and the final reconstructed signal is output.The application realizes high compression ratio under very low encoding power, and guarantees that reconstruction quality statistics do not deteriorate by physical constraint, and is suitable for high-fidelity compression transmission of ocean buoy acoustic data and general acoustic signal.

Description

Technical Field

[0001] This invention relates to the field of marine data compression technology, and in particular to an adaptive compression encoding and decoding method and system for marine buoy acoustic data. Background Technology

[0002] Ocean buoys are important platforms for conducting marine environmental monitoring, underwater acoustic target detection, underwater acoustic communication, and marine biological acoustic observation. Buoys deployed at sea continuously collect acoustic data using sensors such as hydrophones. Due to the high sampling rate and the need for long-term continuous recording, the volume of raw acoustic data is very large.

[0003] Data collected by buoys needs to be transmitted back to shore-based, shipboard, or cloud-based systems for processing. Due to the limitations of the maritime deployment environment, buoys generally use satellite, drone relay, or 4G cellular wireless links for data transmission. These links typically suffer from limited bandwidth and high transmission costs (especially satellite links), and the buoys themselves often rely on batteries or solar power, resulting in tight energy budgets. Directly transmitting raw acoustic data is often difficult to complete within the constraints of limited bandwidth and energy consumption. Therefore, efficient compression of acoustic data at the buoy end and reconstruction at the receiving end are crucial for reducing the amount of data transmitted and saving bandwidth and energy.

[0004] Existing acoustic and audio compression technologies can be broadly categorized as follows: Perceptual audio coding (such as MP3 and AAC) is based on the human auditory masking model, and its psychoacoustic model is optimized for human hearing, which does not align with the statistical characteristics of underwater acoustic signals and the needs of downstream analysis; Compressed sensing methods often rely on iterative convex optimization for reconstruction, but when faced with complex signals, it is difficult to find a suitable sparse basis, affecting the reconstruction effect; Learning-based post-processing often adopts end-to-end direct regression, replacing traditional steps with neural networks. However, most learning-based post-processing methods directly perform end-to-end joint optimization, resulting in a huge computational load at the encoding end, which is not computationally friendly for edge computing devices.

[0005] In summary, the existing technology has the following main drawbacks: (1) Wasted bitrate in silent and low-energy segments: lack of robust mechanism for distinguishing between active and silent segments; (2) Enhanced blindness of objectives and lack of error budget analysis: lack of quantitative analysis on the proportion of two types of errors, namely quantification noise and discard coefficient, and their respective recoverable upper limits; (3) Mismatch between convolution structure and wavelet subband structure: When convolving directly on the coefficient sequence with block priority, adjacent positions of the convolution kernel span different subbands, destroying the correlation structure of the same subband across time blocks; (4) Lack of error budget analysis: The proportion of quantization noise and discard coefficient errors and their respective recoverable upper limits were not quantitatively analyzed, and the correction target was set blindly. Summary of the Invention

[0006] Therefore, the purpose of this invention is to provide an adaptive compression encoding and decoding method for acoustic data of marine buoys, which can adaptively allocate the code rate according to the signal characteristics and reliably correct the distortion caused by quantization and discarding at the receiving end with physically constrained learning-type post-processing, thereby achieving a high compression ratio and stable reconstruction quality under limited backhaul bandwidth and energy consumption.

[0007] To achieve the above objectives, this invention provides an adaptive compression encoding and decoding method for acoustic data from marine buoys, comprising the following steps: S1, acquire the acoustic signal received by the buoy end, perform block processing to form multiple signal blocks, perform silence discrimination estimation on each signal block, and divide the signal block into silent blocks and active blocks according to the silence discrimination estimation result; S2, apply M-level discrete wavelet transform to the active block to obtain a wavelet coefficient vector containing approximation coefficients and level-by-level detail coefficients; S3, based on the preset compression ratio parameter, determine the total bit rate budget for each signal block, determine the number of detail coefficients to be retained according to the total bit rate budget, retain all approximate coefficients to form the final retained set, and record the retained positions to form a retained bitmap; S4, the retained approximation coefficients and detail coefficients are quantized and dequantized in layers according to their respective subbands, and the quantized block type sequence and coefficients are entropy encoded to generate a compressed bitstream and sent to the receiving end; S5, the receiving end performs entropy decoding and dequantization on the compressed bitstream, executes inverse discrete wavelet transform, and performs adaptive smoothing on the transformed block boundaries to obtain the baseline reconstruction signal; S6. Based on the baseline reconstructed signal and the preserved bitmap, perform subband-first rearrangement of the coefficients of multiple consecutive signal blocks within the context window, and arrange the coefficient segments belonging to the same subband in each signal block consecutively to obtain the rearranged one-dimensional coefficient sequence. S7, The rearranged one-dimensional coefficient sequence is input into a dual-domain cascaded neural network for post-processing reconstruction; the dual-domain cascaded neural network includes a time-domain subnetwork and a wavelet-domain subnetwork; Specifically, the process includes: using a wavelet domain subnet to obtain the predicted coefficients after wavelet domain repair, and then applying physical constraint hard projection to obtain reconstructed coefficients that conform to the physical constraints; performing an inverse discrete wavelet transform on the reconstructed coefficients that conform to the physical constraints to convert them back to time-domain waveform signals; inputting the time-domain waveform signals into a time-domain subnet to output a time-domain residual signal; and superimposing the time-domain residual signal with the baseline reconstruction signal obtained in S5 to obtain the final reconstructed acoustic signal after dual-domain cascade repair.

[0008] Furthermore, in S1, silence discrimination estimation is performed on each signal block, and the signal blocks are divided into silence blocks and active blocks based on the silence discrimination estimation results. Specifically, this includes: The root mean square energy of all signal blocks is calculated, the maximum value is taken, and the dynamic range threshold is determined by combining it with the preset minimum dynamic range. The noise floor is estimated by taking the low percentile of the root mean square energy of all signal blocks, and the active threshold is adaptively determined based on the comparison between the noise floor and the dynamic range threshold. Signal blocks with root mean square energy below the active threshold are classified as silent blocks; When the proportion of the number of silent blocks to the total number of blocks exceeds the preset upper limit, the minimum dynamic range is relaxed and the judgment is re-evaluated. If the proportion of silent blocks still exceeds the limit after re-evaluation, no silent blocks are judged.

[0009] Furthermore, the total bitrate budget mentioned in S3 is calculated according to the following formula: Among them, B total The total bitrate budget for each block is expressed in bits; ρ is the compression ratio parameter; N is the block length. This is for rounding down; 16 represents the 16-bit PCM quantization depth of the original signal. When determining the number of detail coefficients to retain based on the total bitrate budget, the number of detail coefficients to retain is calculated using the following formula: In the formula: K is the number of detail coefficients retained per block; γ is the redundancy factor; B rem The remaining budget for detail coefficients; N represents the average quantization bit depth of the detail subband; a This represents the number of approximation coefficients.

[0010] Furthermore, in S4, the retained approximation coefficients and detail coefficients are subjected to hierarchical quantization and dequantization according to their respective subbands using the following formula: In the formula: q i wavelet coefficients c b The quantized integer value of [i]; c b [i] represents the i-th wavelet coefficient in the b-th block, where b is the block number and the coefficient number is i; m j L represents the maximum absolute value of the retention coefficients in the j-th sub-band, where j is the sub-band index; j Let b be the quantization level of the j-th subband, which is determined by the fixed quantization bit depth b pre-configured for that subband. j Determine; round(·) rounds to the nearest integer; clip((·), -L j ,L j() is for amplitude limiting and truncation; These are the inverse quantization reconstruction coefficients.

[0011] More preferably, in S4, the entropy encoding of the quantized block type sequence and coefficients includes: Run-length encoding is performed on the block type sequence to obtain the block type run-length segment; A standard Huffman coding table is constructed for the quantization coefficients of each sub-band of each signal block, and variable-length coding is performed to obtain the Huffman code stream. Write the sampling rate, total length of the original signal, block length, total number of blocks, and quantization bit depth of each subband into the header of the compressed file; Write the Huffman coding table for each subband, the run-length segment of the block type, the maximum subband amplitude of each active block, the reserved bitmap, and the Huffman code stream into the compressed file in sequence to generate the compressed code stream.

[0012] More preferably, in S7, the physical constraint hard projection specifically includes: Obtain the prediction coefficients output by the wavelet domain subnet, and extract the normalized reconstruction coefficients, quantization half-step size, and upper bound of the amplitude of the discarded detail coefficients; wherein, the normalized reconstruction coefficients are the coefficients after normalizing the inverse quantization reconstruction coefficients in S4 according to the scale of the subband, the quantization half-step size is determined by the quantization level of the subband, and the upper bound of the amplitude of the discarded detail coefficients is determined according to the minimum quantization amplitude among the retained detail coefficients; For the reserved position indicated by the reserved bitmap, the prediction coefficients are projected into a closed interval centered on the normalized reconstruction coefficient at that position and with the quantization half-step size at that position as the radius. For the discard position indicated by the retained bitmap, the prediction coefficients are projected into a closed interval with a lower bound of the negative upper bound of the magnitude of the discarded detail coefficient at that position and an upper bound of the positive upper bound of the magnitude of the discarded detail coefficient at that position. Based on the reserved bitmap, the reserved positions are projected using the reserved position projection results, and the discarded positions are projected using the discarded position projection results, and a reconstruction coefficient that conforms to physical constraints is obtained.

[0013] More preferably, in S7, the dual-domain cascaded neural network follows learnable Lagrangian duality constraints during the training phase: A learnable Lagrange multiplier network with backbone network features as input is introduced, and the penalty weights for each coefficient position are output after Softplus activation, which constitute the penalty term for the deviation of the predicted coefficients of the dual-domain cascaded neural network from the physical feasible region. A gradient reversal layer is inserted between the penalty term and the multiplier parameter. The main network of the dual-domain cascaded neural network receives the positive gradient to minimize the violation, and the multiplier parameter receives the reverse gradient to maximize the penalty term, forming a dual-ascending Lagrangian optimization, which drives the dual-domain cascaded neural network to constrain the prediction coefficients within the physical feasible region during training.

[0014] In a further preferred embodiment, in S7, the wavelet domain subnet is used to obtain the prediction coefficients repaired by the wavelet domain, including the following steps: The rearranged one-dimensional coefficient sequence is input into the input convolutional layer of the wavelet domain subnet and projected onto the high-dimensional feature space. The high-dimensional features are processed sequentially through multiple gated temporal convolutional blocks, with the dilation rate of each gated temporal convolutional block increasing layer by layer. The features of each sub-band are averaged and pooled, and each sub-band is aggregated into a token to form a time-frequency token matrix. The time-frequency token matrix is ​​input into the dual-axis attention module: multi-head self-attention calculation is performed on multiple signal block tokens in the context window in the time dimension to capture long-range time context dependencies; cross-attention calculation is performed in the sub-band dimension using the tokens of the low-frequency sub-band as query vectors and the tokens of the high-frequency sub-band as key vectors, and the global structure prior of the low-frequency sub-band is used to guide the refinement of details in the high-frequency sub-band. The refined time-frequency token is broadcast back to the original coefficient length of each sub-band and superimposed on the high-dimensional feature as a context residual feature. The superimposed features are passed through the output convolutional layer to generate wavelet domain residual increments, which are then superimposed on the input features of the wavelet domain subnet to obtain the wavelet domain-repaired prediction coefficients.

[0015] The present invention also provides an adaptive compression encoding and decoding system for ocean buoy acoustic data, comprising the steps of implementing the above-described adaptive compression encoding and decoding method for ocean buoy acoustic data, including: A buoy-end compression encoder is used to perform steps S1-S4 of the method to encode the acoustic signal into a compressed bitstream. The receiver base decoder is used to perform the operation of step S5 of the method to reconstruct the compressed bitstream into a baseline reconstructed signal. The receiving end neural network post-processing reconstruction module is used to perform the operations of steps S6 to S7 of the above method, to post-process and reconstruct the baseline reconstruction signal, and output the final reconstructed acoustic signal.

[0016] The adaptive compression encoding and decoding method and system for acoustic data of marine buoys disclosed in this application have the following effects.

[0017] 1. The buoy-end only performs traditional operations such as wavelet transform, coefficient sorting, and entropy coding, completely without neural networks, resulting in extremely low computational load and making it suitable for battery-powered edge devices such as DSPs / microcontrollers. By eliminating invalid silence segments through dual-constraint silence discrimination and combining rate budget-driven coefficient optimization, the limited bitrate is accurately allocated to core coefficients, achieving superior rate-distortion performance at the same compression ratio.

[0018] 2. By embedding the quantization half-step and the upper bound of the dropout amplitude as hard projection operators into the network's forward propagation, the output coefficients are forcibly locked within the uniquely determined physical feasible region of the bitstream, fundamentally eliminating out-of-bounds distortion and ensuring that, statistically, it is no worse than baseline reconstruction. Error budgeting analysis decomposes the total error into quantization noise and dropout error, making the setting of correction targets more reliable.

[0019] 3. Subband-priority rearrangement aligns the temporal structure of the convolutional receptive field with that of the wavelet subbands. Dual-domain cascading repairs missing frequency bands and smooths temporal aliasing artifacts in the wavelet domain, resulting in clear division of labor and high parameter efficiency. Learnable Lagrange gradient inversion ensures that physical constraints do not collapse during training, and sign-symmetric flipping enhances systematic offsetting of nonlinear biases, further improving reconstruction fidelity. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the adaptive compression encoding and decoding method for acoustic data of marine buoys according to the present invention; Figure 2 This is a schematic diagram of the adaptive compression encoding and decoding system for acoustic data of marine buoys according to the present invention; Figure 3 This is a flowchart of the encoding process of this invention; Figure 4 This is a diagram of the dual-domain neural network processing architecture of the present invention; Figure 5 This is a structural diagram of the dual-axis neural network attention model of the present invention; Figure 6 This is a comparison of the full waveform overlay of the post-processing of the dual-domain cascaded neural network of this invention; Figure 7 This is a high-energy segment magnified superimposed comparison image of the post-processing of the dual-domain cascaded neural network of the present invention; Figure 8 This is a comparison of reconstruction errors within the magnified window of the post-processing of the dual-domain cascaded neural network of this invention; Figure 9 This is a frame-by-frame segmented signal-to-noise ratio (SegSNR) comparison chart for the post-processing of the dual-domain cascaded neural network of this invention; Figure 10 This is a distribution diagram of the frame-by-frame SegSNR improvement in the post-processing of the dual-domain cascaded neural network of the present invention; Figure 11This is a frame-by-frame comparison chart of the LSD moving average of the logarithmic spectral distance in the post-processing of the dual-domain cascaded neural network of this invention; Figure 12 This is a comparison diagram of the LSD cumulative distribution ECDF in the post-processing of the dual-domain cascaded neural network of this invention; Figure 13 This is a histogram showing the frame-by-frame improvement ΔLSD distribution of the dual-domain cascaded neural network post-processing of this invention. Figure 14 This is a graph showing the cumulative improvement ΔLSD of the post-processing of the dual-domain cascaded neural network in this invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, one embodiment of the present invention provides an adaptive compression encoding and decoding method for acoustic data of marine buoys, comprising the following steps: S1: Acquire the acoustic signal received by the buoy end, perform block processing to form multiple signal blocks, and perform silence discrimination estimation on each signal block. Based on the silence discrimination estimation result, the signal block is divided into silent blocks and active blocks. Specifically, this includes: dividing the input signal into blocks of fixed block length N; calculating the root mean square (RMS) energy of each block; taking the maximum value of the RMS of all blocks and the preset minimum dynamic range to determine the dynamic range threshold; estimating the noise floor using the lower percentile of the RMS of each block; if the noise floor is significantly lower than the dynamic range threshold, reserving a preset margin (e.g., about 3dB) as the active threshold; otherwise, taking the dynamic range threshold as the active threshold; blocks with RMS lower than the active threshold are judged as silent blocks. When the proportion of silent blocks exceeds the preset upper limit (typically 0.20), the minimum dynamic range is relaxed and re-discriminated; if it still exceeds the limit, no silent blocks are judged. Only active blocks are encoded; silent blocks are represented by block type markers in the bitstream and reconstructed as silent blocks at the decoding end, thereby saving bitrate.

[0023] S2, apply an M-level discrete wavelet transform to the active block to obtain a wavelet coefficient vector containing approximation coefficients and level-by-level detail coefficients; apply a multi-level discrete wavelet transform (preferably db16, 5 levels) to each active block, flattening the coefficients of each sub-band into a wavelet coefficient vector of length N in a fixed order (approximation first, level-by-level details); the 0th sub-band contains approximation coefficients (the number of which is denoted as N). a The remainder are detail coefficients. Wavelet transform is an orthogonal linear transform, and there exists an inverse transform operator W that allows the reconstructed signal to be recovered from the coefficients.

[0024] S3, based on the preset compression ratio parameter, determine the total bit rate budget for each signal block, determine the number of detail coefficients to be retained according to the total bit rate budget, retain all approximate coefficients to form the final retained set, and record the retained positions to form a retained bitmap; Furthermore, the total bitrate budget mentioned in S3 is calculated according to the following formula: Among them, B total The total bitrate budget for each block is expressed in bits; ρ is the compression ratio parameter; N is the block length. This is for rounding down; 16 represents the 16-bit PCM quantization depth of the original signal. All approximation coefficients are forcibly retained; after deducting the encoding cost of approximation coefficients, the overhead of retaining bitmaps, and the storage overhead of amplitude scales for each subband, the remaining budget B for detail coefficients is obtained. rem When determining the number of detail coefficients to retain based on the total bitrate budget, the number of detail coefficients to retain is calculated using the following formula: In the formula: K is the number of detail coefficients retained per block; γ is the redundancy factor (typically 1.15); B rem The remaining budget for detail coefficients; N represents the average quantization bit depth of the detail subband; a This represents the number of approximation coefficients. Subsequently, from all detail coefficients, K coefficients are selected in descending order of absolute amplitude value to form a retention set together with the approximation coefficients, and the retention positions are recorded in a bitmap of length N.

[0025] S4, the retained approximation coefficients and detail coefficients are quantized and dequantized layer by layer according to their respective subbands, and the quantized block type sequence and coefficients are entropy encoded to generate a compressed bitstream which is then sent to the receiving end; such as Figure 3 The diagram shows the overall flowchart of the encoding end: audio file - block segmentation - silence / active block discrimination - active block includes wavelet transform, selection of coefficients with maximum energy, layered quantization, Huffman + RLE lossless compression, and writing to .cbsin compressed file; silence block is discarded and processed into the next block without compression; this corresponds to the execution process of S1-S4.

[0026] Furthermore, in S4, the retained approximation coefficients and detail coefficients are subjected to hierarchical quantization and dequantization according to their respective subbands using the following formula: In the formula: q i wavelet coefficients c b The quantized integer value of [i]; c b [i] represents the i-th wavelet coefficient in the b-th block, where b is the block number and the coefficient number is i; m j L represents the maximum absolute value of the retention coefficients in the j-th sub-band, where j is the sub-band index; j Let b be the quantization level of the j-th subband, which is determined by the fixed quantization bit depth b pre-configured for that subband. jDetermine; round(·) rounds to the nearest integer; clip((·), -L j ,L j () is a limit cutoff; These are the inverse quantization reconstruction coefficients. From this, we can obtain the quantization half-step size q within the normalization domain. half j = 0.5 / L j This is used for post-processing physical constraint projection.

[0027] More preferably, in S4, the entropy encoding of the quantized block type sequence and coefficients includes: Run-length encoding is performed on the block type sequence to obtain the block type run-length segment; A standard Huffman coding table is constructed for the quantization coefficients of each sub-band of each signal block, and variable-length coding is performed to obtain the Huffman code stream. Write the sampling rate, total length of the original signal, block length, total number of blocks, and quantization bit depth of each subband into the header of the compressed file; Write the Huffman coding table for each subband, the run-length segment of the block type, the maximum subband amplitude of each active block, the reserved bitmap, and the Huffman code stream into the compressed file in sequence to generate the compressed code stream.

[0028] The compressed file includes a header containing Huffman tables for each subband, a block type run-length segment, the maximum amplitude of each subband in each active block data segment, a reserved bitmap, and a Huffman bitstream. The compression ratio is the ratio of the original 16-bit PCM bytes to the compressed file bytes.

[0029] S5, the receiving end performs entropy decoding and dequantization on the compressed bitstream, executes inverse discrete wavelet transform, and performs adaptive smoothing on the transformed block boundaries to obtain the baseline reconstruction signal; The receiver performs entropy decoding, subband dequantization, and inverse wavelet transform reconstruction. For each block boundary, the ratio of the transition amplitude δ between adjacent sampling points to the local root-mean-square energy σ of the boundary neighborhood is calculated. When δ > 0.25σ, uniform filtering smoothing is applied to the boundary neighborhood with a smoothing intensity α = min(0.3·δ / σ, 0.5), smoothing only at significant transitions to avoid excessive blurring. δ is the transition amplitude between adjacent sampling points at the block boundary; σ is the local root-mean-square energy of the boundary neighborhood; α is the adaptive smoothing intensity.

[0030] The global signal-to-noise ratio, segmented signal-to-noise ratio, and power spectral density deviation are respectively: Where: SNR is the global signal-to-noise ratio (dB); x[n] is the original signal, For the reconstructed signal, n is the sampling point number; log 10 It is a logarithm with base 10.

[0031] In the formula: SegSNR is the segmented signal-to-noise ratio (dB); M is the total number of frames, and m is the frame number; These are the original and reconstructed signals of the m-th frame, respectively; the signal-to-noise ratio of each frame is truncated to [-10, 50] dB.

[0032] Where: LSD is the logarithmic spectral distance deviation (dB); P f , These are the power spectral density estimates of the original and reconstructed signals (Welch method), respectively; f is the frequency point index, N f This represents the total number of frequency points.

[0033] By forming a three-dimensional evaluation of reconstruction quality from three orthogonal dimensions—overall time domain, time domain segmentation, and frequency domain spectral envelope—the influence of a single indicator is avoided, and the multi-dimensional consistency of the reconstructed signal at the physical level is objectively reflected.

[0034] like Figure 6-8 As shown, the difference in time-domain waveforms before and after post-processing is intuitively demonstrated. Figure 6 The image shows the effect of overlaying the original signal, the preliminary decoded signal, and the deep post-processed signal; it can be seen that the post-processed signal is closer to the original signal as a whole, especially in the area with strong energy, the deviation of the preliminary decoding is corrected to a certain extent. Figure 7 This is for high-energy band amplification and superposition effects; in the high-energy band, the post-processing curve matches the original signal better than the initial decoding curve. Figure 8 The reconstruction error curves within the magnification window are shown for preliminary decoding and post-processing, illustrating that post-processing is more effective in recovering strong signal regions. The overall post-processing error is lower than that of preliminary decoding, and the RMSE is smaller, indicating that post-processing reduces the temporal reconstruction error.

[0035] Figure 9 The frame-by-frame segmented signal-to-noise ratio (SegSNR) curves before and after post-processing show an average SegSNR of 19.32 dB during initial decoding and 19.62 dB during post-processing. Figure 10 The curve shows the change in SegSNR improvement per frame. In some time periods, ΔSegSNR>0, indicating that post-processing brings positive improvement in most frames, with only a slight decrease in a few frames.

[0036] Figure 11 Evaluating LSD frame-by-frame indicates that the post-processed LSD is smaller and its spectral envelope is closer to the original signal. Figure 12 Evaluating LSD from the cumulative distribution indicates that the overall distribution of post-processed LSD is smaller, meaning that the spectral distortion of most frames is lower; Figure 13The LSD was evaluated by the distribution of the improvement amount. The improvement amount was concentrated in the positive range, indicating that post-processing reduced the LSD in most frames.

[0037] Figure 14 The continuous upward trend of the LSD cumulative curve, as evaluated by the cumulative improvement, indicates that the LSD improvement brought about by post-processing is stable and positively cumulative, and there is no systematic degradation.

[0038] The lossy compression in the first part introduces two types of errors at the receiver: quantization noise caused by the quantization of retained coefficients, and discarding error caused by the discarding of unretained detail coefficients. This section provides a physically constrained neural network dual-domain post-processing method to correct these errors. It should be noted that this post-processing is an optional refinement step with limited improvement; its main value lies in the physical constraint mechanism ensuring reliable correction and preventing statistical degradation of reconstruction quality. During buoy-end compression, the total distortion faced by the receiver is the superposition of errors from two completely different physical mechanisms. Therefore, we first quantitatively assess the proportions of quantization noise and discarding error, formulate differentiated correction strategies accordingly, and then solidify these strategies into the network training through loss weights, thereby improving the relevance and stability of the post-processing.

[0039] (1) Error budget analysis and correction objectives. The total error is quantized and decomposed using a prediction method: Prediction A takes the inverse quantization coefficient at the retained position and the true value at the discard position (perfect filling of the discard coefficient); prediction B takes the true value at the retained position and zero at the discard position (perfect inverse quantization retains the coefficient). The energy proportions of the two types of errors are: In the formula: η quant η drop These represent the energy proportions of quantization noise and dropout coefficient errors, respectively; x represents the original signal, For baseline reconstruction signal; x A x B ...

[0040] (2) Upper bound of triplet extraction and discard amplitude. For each active block, extract the true value of the coefficients (supervisory target), the dequantized reconstructed coefficients, the retained bitmap, and the scale of each subband. The amplitude of the discarded detail coefficients is less than the smallest among the retained detail coefficients, so there is an upper bound. The upper bound is taken on the set of retained detail coefficients as follows: In the formula: T is the upper bound of the magnitude of the discarded coefficients, and its minimum value is taken on the set of retained detail coefficients; Let be the inverse quantization reconstruction coefficient at position i; j(i) be the subband to which position i belongs. The coefficients and their upper bounds are normalized to the subband scale before being input into the network (normalized reconstruction coefficient x). rThe "normalized discard upper bound T" and "discard amplitude upper bound" together constitute a key data bridge connecting the physical constraints at the encoding end and the neural network inference at the receiving end. They are not simple data processing, but rather provide the necessary parameter basis and operational boundaries for the subsequent "physical constraint hard projection".

[0041] S6. Based on the baseline reconstructed signal and the preserved bitmap, subband-priority rearrangement is performed on the coefficients of multiple consecutive signal blocks within the context window. The coefficient segments belonging to the same subband in each signal block are arranged consecutively to obtain the rearranged one-dimensional coefficient sequence. The overall architecture of the dual-domain cascaded neural network is adopted. After obtaining the input features, the subband-priority rearrangement is performed first. After wavelet domain recovery in stage A and time domain recovery in stage B, the final reconstructed signal is obtained.

[0042] Subband-first rearrangement. Rearrange the coefficients of several consecutive blocks within the context window: for each subband, arrange the corresponding subband segments of each block consecutively to obtain the permutation π and its inverse π. -1 Thus, adjacent positions in a one-dimensional convolution correspond to the coefficients of adjacent signal blocks within the same subband, enabling the convolution kernel to directly model the cross-block temporal structure within the subband; the main network operates on a subband-priority layout, and after projection, it passes through π... -1 Prioritize the layout of back-blocks for inverse wavelet transform. Reshape the data arrangement to match the spatial processing logic of the neural network with the physical structure of the wavelet transform.

[0043] S7, The rearranged one-dimensional coefficient sequence is input into a dual-domain cascaded neural network for post-processing reconstruction; the dual-domain cascaded neural network includes a time-domain subnetwork and a wavelet-domain subnetwork; Specifically, the process includes: using a wavelet domain subnet to obtain the predicted coefficients after wavelet domain repair, and then applying physical constraint hard projection to obtain reconstructed coefficients that conform to the physical constraints; performing an inverse discrete wavelet transform on the reconstructed coefficients that conform to the physical constraints to convert them back to time-domain waveform signals; inputting the time-domain waveform signals into a time-domain subnet to output a time-domain residual signal; and superimposing the time-domain residual signal with the baseline reconstruction signal obtained in S5 to obtain the final reconstructed acoustic signal after dual-domain cascade repair.

[0044] The physical constraint hard projection specifically includes: Obtain the prediction coefficients output by the wavelet domain subnet, and extract the normalized reconstruction coefficients, quantization half-step size, and upper bound of the amplitude of the discarded detail coefficients; wherein, the normalized reconstruction coefficients are the coefficients after normalizing the inverse quantization reconstruction coefficients in S4 according to the scale of the subband, the quantization half-step size is determined by the quantization level of the subband, and the upper bound of the amplitude of the discarded detail coefficients is determined according to the minimum quantization amplitude among the retained detail coefficients; For the reserved position indicated by the reserved bitmap, the prediction coefficients are projected into a closed interval centered on the normalized reconstruction coefficient at that position and with the quantization half-step size at that position as the radius. For the discard position indicated by the retained bitmap, the prediction coefficients are projected into a closed interval with a lower bound of the negative upper bound of the magnitude of the discarded detail coefficient at that position and an upper bound of the positive upper bound of the magnitude of the discarded detail coefficient at that position. Based on the reserved bitmap, the reserved positions are projected using the reserved position projection results, and the discarded positions are projected using the discarded position projection results, and a reconstruction coefficient that conforms to physical constraints is obtained.

[0045] In S7, the wavelet domain subnet is used to obtain the prediction coefficients after wavelet domain repair, including the following steps: The rearranged one-dimensional coefficient sequence is input into the input convolutional layer of the wavelet domain subnet and projected onto the high-dimensional feature space. The high-dimensional features are processed sequentially through multiple gated temporal convolutional blocks, with the dilation rate of each gated temporal convolutional block increasing layer by layer. The features of each sub-band are averaged and pooled, and each sub-band is aggregated into a token to form a time-frequency token matrix. The time-frequency token matrix is ​​input into the dual-axis attention module as follows: Figure 5 This is the internal structure diagram of the dual-axis attention module: In the time dimension, multi-head self-attention calculation is performed on multiple signal block tokens within the context window to capture long-range temporal context dependencies; in the sub-band dimension, cross-attention calculation is performed using the tokens of the low-frequency sub-band as query vectors and the tokens of the high-frequency sub-band as key vectors, using the global structural prior of the low-frequency sub-band to guide the detailed refinement of the high-frequency sub-band; the input time-frequency token matrix is ​​first tokenized into sub-bands, then multi-head self-attention is performed on the time axis, and cross-band cross-attention is performed on the sub-band axis with low-frequency guiding high-frequency. After passing through a feedforward network and residual connections, the refined tokens are output for subsequent wavelet domain coefficient repair.

[0046] The refined time-frequency token is broadcast back to the original coefficient length of each sub-band and superimposed on the high-dimensional feature as a context residual feature. The superimposed features are passed through the output convolutional layer to generate wavelet domain residual increments, which are then superimposed on the input features of the wavelet domain subnet to obtain the wavelet domain-repaired prediction coefficients.

[0047] like Figure 4Specifically: Stage A performs wavelet domain restoration: the 6-dimensional feature vector is projected onto a high-dimensional space, and then local features are extracted through 8 TCN blocks. Each sub-band is averaged and pooled into 1 token to form a 16×6 time-frequency map. Self-attention is performed on 16 time blocks to capture long-distance dependencies. Cross-attention is performed between 6 sub-bands, with low frequency (cA5) guiding the global distribution of high frequency (cD1). The refined time-frequency tokens are broadcast back to the original coefficient length and added to the original features as strong context residuals. The wavelet predictor can generate residual increments. The network's prediction coefficients are constrained within the quantization half-step and the upper limit of the discard, and then projected. Stage B performs time-domain restoration: time-domain restoration uses 4 TCN blocks to improve and restore the waveform output from Stage A, based on the residual signal. Finally, the enhanced signal is output.

[0048] In S7, the dual-domain cascaded neural network follows learnable Lagrangian duality constraints during the training phase: A learnable Lagrange multiplier network with backbone network features as input is introduced, and the penalty weights for each coefficient position are output after Softplus activation, which constitute the penalty term for the deviation of the predicted coefficients of the dual-domain cascaded neural network from the physical feasible region. A gradient reversal layer is inserted between the penalty term and the multiplier parameter. The main network of the dual-domain cascaded neural network receives the positive gradient to minimize the violation, and the multiplier parameter receives the reverse gradient to maximize the penalty term, forming a dual-ascending Lagrangian optimization, which drives the dual-domain cascaded neural network to constrain the prediction coefficients within the physical feasible region during training.

[0049] Specifically, this includes setting the network's prediction value for the i-th coefficient as follows: Its corresponding quantized reconstruction value is Quantization half-step size is The upper limit of the drop amplitude is Preserve bitmap as The physically feasible region is defined as: When K=1, the coefficients are retained; when K=0, the coefficients are discarded. The true value of the retained coefficients must fall within the interval centered on the quantization value and with a radius equal to half the step size. The absolute value of the true value of the discarded coefficients does not exceed the minimum quantization amplitude among the selected coefficients.

[0050] Learnable penalty weights: Traditional approaches use fixed hyperparameters As a fixed penalty weight, it is difficult to adapt to the local characteristics of each coefficient. Therefore, a learnable Lagrange multiplier network is introduced. , To introduce the function after the Lagrange multiplier network, the backbone features are used. As input, the penalty weight at each coefficient is activated by Softplus at the output. : Dual Ascent and Gradient Reversal: Lagrangian relaxation transforms the constrained optimization problem into a min-max saddle point problem. in, For Lagrange relaxation calculations, the amount of violation , for The upper and lower bounds, directly minimizing the above penalty term for both sets of parameters simultaneously will make Constraint degradation fails. A gradient inversion layer is used to achieve dual ascent: Insert a GRL between the penalty term and the multiplier parameter, and the network body... It receives positive gradients to minimize violations, while the parameters of the Lagrange multiplier network can be learned. The gradient inversion layer is then received to maximize the penalty term, forming a Lagrangian optimization and avoiding collapse degradation. α is the gradient scaling factor during backpropagation in the gradient inversion layer. I It is an identity matrix.

[0051] Let p be the predicted value (normalized domain) of the network at position i. i : Project the preserved position onto an interval centered on the inverse quantization reconstructed value and with a radius equal to the quantization half-step. Within; project the drop position onto the upper limit of the drop amplitude; then synthesize using a mask: In the formula: p i x is the network's predicted value at position i; {r,i} q is the normalized reconstruction coefficient at this location; half where i is the quantization half-step size at that position; Discard the upper bound for normalization at this position; keep i To preserve the bitmap (1 for keep / 0 for discard); proj i These are the valid coefficients after projection. Since hard projection squeezes the output into a feasible region uniquely determined by the bitstream, regardless of how the network predicts, the output is physically consistent with the bitstream, fundamentally avoiding overcorrection and out-of-bounds distortion in unconstrained end-to-end regression. The projected coefficients are then refined using inverse wavelet transform and a time-domain subnet to output the corrected signal.

[0052] (7) Loss and Training / Inference. The total training loss is a weighted average of four complementary terms: in, , , Take values ​​of 0.3, 0.1, and 0.1 respectively. Main loss For stage A auxiliary loss; For coefficient supervision loss, This is a penalty for physical constraints.

[0053] The main loss uses negative SNR and time-domain loss. Combined with STFT frequency domain amplitude loss: The STFT frequency domain amplitude term is specifically designed to address aliasing artifacts in DWT reconstruction. Stage A auxiliary loss. Apply supervision to the intermediate output in the wavelet domain, and structure and same. The final reconstructed acoustic signal is represented by y, where y is the original clean acoustic signal, i.e., the training label; STFT(·) represents the short-time Fourier transform. and The weights are for each item. This is a smoothing term.

[0054] Coefficient monitoring loss Pull the predicted coefficients toward the true values ​​and reduce the weight of the discarded coefficients to avoid the errors introduced by forced fitting and truncation: Physical constraint penalty Differentiable Lagrange penalty: in, The true value of the i-th wavelet coefficient, i.e., the full-precision wavelet coefficient before encoding. For the initial smoothing term, w i The supervisory weights at each position are used to adjust the importance of the loss at that position; k i To preserve the bitmap; γ is the redundancy factor. The learnable Lagrange multiplier at position i, i.e., the penalty weight, ν i (): The violation amount at the i-th position, used to measure the degree to which the predicted coefficient deviates from the physical feasible region.

[0055] The algorithm employs multi-pass encoding enhancement, symbolic flip enhancement based on symbolic symmetry, sequence overlap sampling, AdamW, single-cycle scheduling, exponential moving average, gradient accumulation, mixed precision, gradient pruning, and early stopping. Gradient checkpointing is used to reduce memory usage to accommodate devices with limited memory. During inference, a sliding window (approximately half a context step) is used to concatenate the centers of each window, and the positive and negative inputs are inferred separately using symbolic symmetry and then averaged (enhanced during testing). Silence blocks and invalid blocks maintain baseline / silence.

[0056] Multi-pass encoding augmentation: During training, the input coefficients x are simultaneously encoded with a 50% probability. r (Reconstruction coefficient) and label coefficient y f Multiplying the (full-precision coefficients) by -1, since DWT is a linear transformation and the quantization grid is symmetric about the origin, the flipped sample is still an exact feasible solution to the original problem. This forces the model to be invariant to changes in the sign of the coefficients. During inference, the same block is input with the original coefficients +x and the flipped coefficients -x, respectively, and forward inference is performed twice to obtain the flipped label coefficients y. pos and the label coefficient y after inference neg The final output is 0.5 * (y) pos - y neg This symmetrical averaging can effectively offset the slight sign bias introduced by nonlinear activation in the model, thereby improving the reconstruction SNR.

[0057] Symbol-symmetry-based symbol-flipping enhancement: During the dataset construction phase, for each input WAV file, the code generates 7 independent encoding attempts. The first attempt serves as the baseline, and the subsequent 6 attempts apply random perturbations. These include random offsets and random gains. The random offsets are uniformly sampled integers within the range [0, N). The raw audio data is truncated from this offset point, simulating time-phase misalignment or asynchronous sampling of the analog signal; the random gain is logarithmically uniformly sampled within the logarithmic interval to simulate volume changes at different distances.

[0058] like Figure 2 The present invention also provides an adaptive compression encoding and decoding system for ocean buoy acoustic data, comprising the steps of implementing the above-described adaptive compression encoding and decoding method for ocean buoy acoustic data, including: A buoy-end compression encoder is used to perform steps S1-S4 of the method, encoding the acoustic signal into a compressed bitstream; and transmitting it to the receiving end via satellite communication. The receiver base decoder is used to perform the operation of step S5 of the method, shore-based decoding and post-processing, to reconstruct the compressed bitstream into a baseline reconstructed signal; The receiving end neural network post-processing reconstruction module is used to perform the operations of steps S6 to S7 of the above method, to post-process and reconstruct the baseline reconstruction signal, and output the final reconstructed acoustic signal.

[0059] Table 1 below details the parameter configuration and implementation process of this application: Read a segment of underwater acoustic signal collected by an ocean buoy and divide it into N blocks → Obtain active / silent masks by dual constraint discrimination → Perform db16 five-level wavelet transform on the active blocks and flatten them to obtain coefficient vectors → Determine the code rate budget according to ρ, retain all approximate coefficients, and selectively retain detail coefficients according to amplitude → Quantize each sub-band in layers and generate compressed files using run-length encoding and standard Huffman coding, send them through the backhaul link and record the compression rate → Obtain baseline reconstruction by entropy decoding, inverse quantization, inverse wavelet transform and adaptive smoothing of block boundaries at the receiving end → Obtain corrected reconstruction by post-processing through a dual-domain neural network → Evaluate by comparison of SNR, SegSNR, and LSD deviation.

[0060] Table 1 is a parameter configuration table for embodiments of this application.

[0061] Table 2 shows the comparison results between the processed methods of this application and existing methods. As shown in Table 2, the implementation results are as follows: A high compression ratio and stable reconstruction quality can be achieved under limited backhaul bandwidth; the receiver-side neural network post-processing performs constrained correction of quantization and dropout errors under physical constraints, achieving a certain degree of signal-to-noise ratio improvement without degrading the reconstruction quality. This indicates that, at the same compression ratio, the objective quality indicators of the reconstructed signal are generally superior to existing audio compression methods, especially at medium compression ratios; it remains competitive even at extremely low bit rates.

[0062] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An adaptive compression encoding and decoding method for acoustic data from ocean buoys, characterized in that, Includes the following steps: S1, acquire the acoustic signal received by the buoy end, perform block processing to form multiple signal blocks, perform silence discrimination estimation on each signal block, and divide the signal block into silent blocks and active blocks according to the silence discrimination estimation result; S2, apply M-level discrete wavelet transform to the active block to obtain a wavelet coefficient vector containing approximation coefficients and level-by-level detail coefficients; S3, based on the preset compression ratio parameter, determine the total bit rate budget for each signal block, determine the number of detail coefficients to be retained according to the total bit rate budget, retain all approximate coefficients to form the final retained set, and record the retained positions to form a retained bitmap; S4, the retained approximation coefficients and detail coefficients are quantized and dequantized in layers according to their respective subbands, and the quantized block type sequence and coefficients are entropy encoded to generate a compressed bitstream and sent to the receiving end; S5, the receiving end performs entropy decoding and dequantization on the compressed bitstream, executes inverse discrete wavelet transform, and performs adaptive smoothing on the transformed block boundaries to obtain the baseline reconstruction signal; S6. Based on the baseline reconstructed signal and the preserved bitmap, perform subband-first rearrangement of the coefficients of multiple consecutive signal blocks within the context window, and arrange the coefficient segments belonging to the same subband in each signal block consecutively to obtain the rearranged one-dimensional coefficient sequence. S7, The rearranged one-dimensional coefficient sequence is input into a dual-domain cascaded neural network for post-processing reconstruction; the dual-domain cascaded neural network includes a time-domain subnetwork and a wavelet-domain subnetwork; Specifically, the process includes: using a wavelet domain subnet to obtain the predicted coefficients after wavelet domain repair, and then applying physical constraint hard projection to obtain reconstructed coefficients that conform to the physical constraints; performing an inverse discrete wavelet transform on the reconstructed coefficients that conform to the physical constraints to convert them back to time-domain waveform signals; inputting the time-domain waveform signals into a time-domain subnet to output a time-domain residual signal; and superimposing the time-domain residual signal with the baseline reconstruction signal obtained in S5 to obtain the final reconstructed acoustic signal after dual-domain cascade repair.

2. The adaptive compression encoding and decoding method for ocean buoy acoustic data according to claim 1, characterized in that, In S1, silence discrimination estimation is performed on each signal block. Based on the silence discrimination estimation results, the signal blocks are divided into silence blocks and active blocks, specifically including: The root mean square energy of all signal blocks is calculated, the maximum value is taken, and the dynamic range threshold is determined by combining it with the preset minimum dynamic range. The noise floor is estimated by taking the low percentile of the root mean square energy of all signal blocks, and the active threshold is adaptively determined based on the comparison between the noise floor and the dynamic range threshold. Signal blocks with root mean square energy below the active threshold are classified as silent blocks; When the proportion of the number of silent blocks to the total number of blocks exceeds the preset upper limit, the minimum dynamic range is relaxed and the judgment is re-evaluated. If the proportion of silent blocks still exceeds the limit after re-evaluation, no silent blocks are judged.

3. The adaptive compression encoding and decoding method for ocean buoy acoustic data according to claim 1, characterized in that, The total bitrate budget mentioned in S3 is calculated according to the following formula: Among them, B total The total bitrate budget for each block is expressed in bits; ρ is the compression ratio parameter; N is the block length. This is for rounding down; 16 represents the 16-bit PCM quantization depth of the original signal. When determining the number of detail coefficients to retain based on the total bitrate budget, the number of detail coefficients to retain is calculated using the following formula: In the formula: K is the number of detail coefficients retained per block; γ is the redundancy factor; B rem The remaining budget for detail coefficients; N represents the average quantization bit depth of the detail subband; a This represents the number of approximation coefficients.

4. The adaptive compression encoding and decoding method for ocean buoy acoustic data according to claim 1, characterized in that, In S4, the retained approximation coefficients and detail coefficients are subjected to hierarchical quantization and dequantization according to their respective subbands using the following formula. In the formula: q i wavelet coefficients c b The quantized integer value of [i]; c b [i] represents the i-th wavelet coefficient in the b-th block, where b is the block number and the coefficient number is i; m j L represents the maximum absolute value of the retention coefficients in the j-th sub-band, where j is the sub-band index; j Let b be the quantization level of the j-th subband, which is determined by the fixed quantization bit depth b pre-configured for that subband. j Determine; round(·) rounds to the nearest integer; clip((·), -L j ,L j () is for amplitude limiting and truncation; These are the inverse quantization reconstruction coefficients.

5. The adaptive compression encoding and decoding method for ocean buoy acoustic data according to claim 4, characterized in that, In S4, the entropy encoding of the quantized block type sequence and coefficients includes: Run-length encoding is performed on the block type sequence to obtain the block type run-length segment; A standard Huffman coding table is constructed for the quantization coefficients of each sub-band of each signal block, and variable-length coding is performed to obtain the Huffman code stream. Write the sampling rate, total length of the original signal, block length, total number of blocks, and quantization bit depth of each subband into the header of the compressed file; Write the Huffman coding table for each subband, the run-length segment of the block type, the maximum subband amplitude of each active block, the reserved bitmap, and the Huffman code stream into the compressed file in sequence to generate the compressed code stream.

6. The adaptive compression encoding and decoding method for ocean buoy acoustic data according to claim 5, characterized in that, In S7, the physical constraint hard projection specifically includes: Obtain the prediction coefficients output by the wavelet domain subnet, and extract the normalized reconstruction coefficients, quantization half-step size, and upper bound of the amplitude of the discarded detail coefficients; wherein, the normalized reconstruction coefficients are the coefficients after normalizing the inverse quantization reconstruction coefficients in S4 according to the scale of the subband, the quantization half-step size is determined by the quantization level of the subband, and the upper bound of the amplitude of the discarded detail coefficients is determined according to the minimum quantization amplitude among the retained detail coefficients; For the reserved position indicated by the reserved bitmap, the prediction coefficients are projected into a closed interval centered on the normalized reconstruction coefficient at that position and with the quantization half-step size at that position as the radius. For the discard position indicated by the retained bitmap, the prediction coefficients are projected into a closed interval with a lower bound of the negative upper bound of the magnitude of the discarded detail coefficient at that position and an upper bound of the positive upper bound of the magnitude of the discarded detail coefficient at that position. Based on the reserved bitmap, the reserved positions are projected using the reserved position projection results, and the discarded positions are projected using the discarded position projection results, and a reconstruction coefficient that conforms to physical constraints is obtained.

7. The adaptive compression encoding and decoding method for ocean buoy acoustic data according to claim 5, characterized in that, In S7, the dual-domain cascaded neural network follows learnable Lagrangian duality constraints during the training phase: A learnable Lagrange multiplier network with backbone network features as input is introduced, and the penalty weights for each coefficient position are output after Softplus activation, which constitute the penalty term for the deviation of the predicted coefficients of the dual-domain cascaded neural network from the physical feasible region. A gradient reversal layer is inserted between the penalty term and the multiplier parameter. The main network of the dual-domain cascaded neural network receives the positive gradient to minimize the violation, and the multiplier parameter receives the reverse gradient to maximize the penalty term, forming a dual-ascending Lagrangian optimization, which drives the dual-domain cascaded neural network to constrain the prediction coefficients within the physical feasible region during training.

8. The adaptive compression encoding and decoding method for ocean buoy acoustic data according to claim 7, characterized in that, In S7, the wavelet domain subnet is used to obtain the prediction coefficients after wavelet domain repair, including the following steps: The rearranged one-dimensional coefficient sequence is input into the input convolutional layer of the wavelet domain subnet and projected onto the high-dimensional feature space. The high-dimensional features are processed sequentially through multiple gated temporal convolutional blocks, with the dilation rate of each gated temporal convolutional block increasing layer by layer. The features of each sub-band are averaged and pooled, and each sub-band is aggregated into a token to form a time-frequency token matrix. The time-frequency token matrix is ​​input into the dual-axis attention module: multi-head self-attention calculation is performed on multiple signal block tokens in the context window in the time dimension to capture long-range time context dependencies; cross-attention calculation is performed in the sub-band dimension using the tokens of the low-frequency sub-band as query vectors and the tokens of the high-frequency sub-band as key vectors, and the global structure prior of the low-frequency sub-band is used to guide the refinement of details in the high-frequency sub-band. The refined time-frequency token is broadcast back to the original coefficient length of each sub-band and superimposed on the high-dimensional feature as a context residual feature. The superimposed features are passed through the output convolutional layer to generate wavelet domain residual increments, which are then superimposed on the input features of the wavelet domain subnet to obtain the wavelet domain-repaired prediction coefficients.

9. An adaptive compression encoding and decoding system for acoustic data from marine buoys, characterized in that, The steps for implementing the adaptive compression encoding and decoding method for ocean buoy acoustic data according to any one of claims 1-8 include: A buoy-end compression encoder is used to perform steps S1-S4 of the method as described in claim 1, encoding the acoustic signal into a compressed bitstream; The receiver base decoder is used to perform the operation of step S5 of the method as described in claim 1, reconstructing the compressed bitstream into a baseline reconstructed signal; The receiving end neural network post-processing reconstruction module is used to perform the operations of steps S6 to S7 of the method as described in claim 1, to post-process and reconstruct the baseline reconstruction signal, and output the final reconstructed acoustic signal.

Images