A method and system for reconstructing underwater acoustic signals

Abstract

The application relates to a kind of underwater acoustic signal reconstruction method and system, wherein the method comprises the following steps: receiving underwater acoustic signal by underwater acoustic sensor node;Signal decomposition is carried out on the underwater acoustic signal, and trend item signal and seasonal item signal are obtained;Feature of trend item signal is extracted based on one-dimensional convolution;Trend item signal reconstruction is carried out based on the extracted feature;Seasonal item signal reconstruction is carried out based on encoder-decoder;The reconstructed trend item signal and seasonal item signal are added element by element, and reconstructed underwater acoustic signal is obtained.Compared with prior art, the reconstructed signal has higher signal-to-noise ratio, lower root mean square error, and higher signal processing efficiency.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic signal processing, and in particular to an underwater acoustic signal reconstruction method and system. Background Technology

[0002] Underwater acoustic communication is a crucial technology in marine exploration, monitoring, and communication, relying on the propagation of acoustic signals in the underwater environment. However, the complexity of the underwater environment, including but not limited to variations in seawater temperature and salinity, currents, tides, and the variability of seabed topography, significantly impacts the stability of underwater acoustic signal propagation. These factors can cause signal attenuation, scattering, and reflection, and may even lead to signal loss under certain conditions, thereby reducing communication quality and reliability.

[0003] In existing technologies, although various signal reconstruction techniques have been developed, such as compressed sensing-based methods, which attempt to recover signals by reducing the number of required sampling points and thus achieving effective signal reconstruction, the characteristics of signals vary with seasons and environmental conditions. For example, seasonal variations in factors such as water temperature, salinity, and water flow can affect the propagation characteristics of sound waves. This variation means that compressed sensing models, which are based on steady-state assumptions, may not accurately reconstruct signals in practical applications.

[0004] On the other hand, traditional signal reconstruction methods, such as those based on Fourier transform or wavelet transform, often perform poorly when dealing with high noise levels or complex signal variations. These methods tend to introduce additional noise or distortion during reconstruction, especially when there is significant signal loss, which greatly reduces the accuracy of the reconstruction. Furthermore, these traditional methods rarely consider the seasonality of underwater acoustic signals, making it impossible to effectively reconstruct the signal components affected by seasonal changes.

[0005] Machine learning offers a novel approach to signal reconstruction using massive amounts of data. Theoretically, by training large neural network models, it's possible to capture the intrinsic features of complex signals. However, this method requires vast amounts of training data and expensive computational resources. Furthermore, in practical applications, it faces challenges in addressing overfitting and generalization issues, which are particularly difficult in resource-constrained underwater communication devices.

[0006] In summary, existing underwater acoustic signal reconstruction techniques have significant limitations when facing complex and variable underwater environments, especially in handling signal reconstruction problems affected by seasonal factors, where current methods often fail to achieve satisfactory results. Therefore, there is an urgent need for a new signal reconstruction technique that can overcome the shortcomings of existing technologies and effectively improve the reconstruction quality and communication reliability of underwater acoustic signals. Summary of the Invention

[0007] The purpose of this invention is to provide an underwater acoustic signal reconstruction method and system that effectively improves the quality of underwater acoustic signals, thereby ensuring the performance and reliability of underwater acoustic communication systems.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] A method for reconstructing underwater acoustic signals includes the following steps:

[0010] S1, using underwater acoustic sensor nodes to receive underwater acoustic signals;

[0011] S2, perform signal decomposition on the underwater acoustic signal to obtain the trend term signal and the seasonal term signal;

[0012] S3, extracts features of trend term signals based on one-dimensional convolution;

[0013] S4, Reconstruct the trend term signal based on the extracted features;

[0014] S5, seasonal term signal reconstruction based on encoder-decoder;

[0015] S6, the reconstructed trend signal and seasonal signal are added element by element to obtain the reconstructed underwater acoustic signal.

[0016] In step S1, the underwater acoustic signal received by the underwater acoustic sensor node is represented as follows:

[0017]

[0018] in, This represents the superposition of multiple signals, each component consisting of amplitude A. i Attenuation coefficient α i The distance r is defined as h(t-τ). i ) represents the oscillation part of the i-th component; N(t,r) is the noise term; t represents time, r represents distance; N represents the number of signals.

[0019] The oscillating component of the underwater acoustic signal is represented as: h(t-τ) i )=cos(2πfi i (t-τ i )-k i r+φ i ), where f i τ represents frequency. i k represents the time delay. i Represents the wave number, φ i Indicates the initial phase.

[0020] In step S2, the trend term signal is obtained by weighting the underwater acoustic signal and performing a softmax operation, including the following steps:

[0021] S211, Calculate the output of the autocorrelation function:

[0022] Considering the variation of the underwater acoustic signal at different time points, the autocorrelation function is defined as follows:

[0023]

[0024] R(τ) represents the degree of correlation of the original signal P under the relative offset τ between different time points, where τ represents the time delay;

[0025] For each time point, extract the underwater acoustic signal P(t,r) and add the time delay τ to the current time node t to obtain the underwater acoustic signal P(t+τ,r) at the new time node t+τ. Calculate the output of the autocorrelation function.

[0026] S212, Define the Gaussian weight vector:

[0027] Define Gaussian function The weight vector used as the values ​​of the weighted autocorrelation function, where τ is the time delay and σ is the standard deviation;

[0028] S213, weighted autocorrelation vector:

[0029] The output of the autocorrelation function is weighted using a Gaussian weight vector to obtain the weighted autocorrelation value R. w (τ)=w(τ)·R(τ);

[0030] S214, applying the softmax function:

[0031] The weighted autocorrelation value R w (τ) is taken as input and transformed into a set of normalized weights S(τ) by the softmax function.

[0032]

[0033] S215, Extract trend signal:

[0034] Using the output S(τ) of the softmax function as weights, the original underwater acoustic signal P(t,r) is weighted and averaged to extract the trend term signal.

[0035] In step S2, the seasonal term signal is obtained by subtracting the extracted trend term signal from the original underwater acoustic signal.

[0036] Step S3 includes the following steps:

[0037] S31, preprocess the trend signal;

[0038] S32, according to the preset time window size, traverse the trend term signals in chronological order, and extract the trend term signals in each window into a vector S[t], which is used as the input signal of the convolution module;

[0039] S33 defines the underwater acoustic convolution kernel, which performs a one-dimensional convolution operation on the input signal:

[0040] Define a one-dimensional underwater acoustic convolution kernel K[n], where n = 0, 1, ..., L-1, and L is the length of the convolution kernel. The weight distribution of the underwater acoustic convolution kernel is K[n] = acos(b·n+c)+d, where a is the amplitude, controlling the weight range; b is the modulation factor, adjusting the sensitivity of the underwater acoustic convolution kernel to different paths; c is the phase offset, adjusting the weight of the underwater acoustic convolution kernel to match the path effect in the signal; and d is the DC offset, ensuring that the weight of the underwater acoustic convolution kernel is positive. Apply the underwater acoustic convolution kernel K[n] to the input signal S[t] to perform a one-dimensional convolution operation, obtaining...

[0041] S34, the ReLU activation function is applied to the output of the one-dimensional convolution operation to obtain the feature vector of the trend term signal.

[0042] Step S4 includes the following steps:

[0043] S41, calculate the mean and variance of the feature data y of the trend term signal obtained in step S3, and perform real-time layer normalization based on the mean and variance to obtain the preprocessed signal Y = [Y (1) ,Y (2) ,…,Y (t) ];

[0044] S42, reconstructing the trend signal, specifically includes the following steps:

[0045] S421, given the initial hidden state h (0) and memory unit m (0) ;

[0046] S422, Calculating the forget gate:

[0047] f (t) =σ(W f ·[h (t-1) ,Y (t) ]+b f +γ·D)

[0048] Among them, f (t) It is the forget gate for the current time step, σ is the sigmoid function, and W f It is the weight matrix of the forget gate, h (t-1) It is the hidden state of the previous time step, Y(t) It is the input signal at the current time step, b f γ is the bias term of the forget gate, D is a factor related to the attenuation of the underwater acoustic signal, and γ is an adjustment coefficient used to control the degree of influence of the underwater acoustic signal attenuation factor on the forget gate.

[0049] S423, Calculation Input Gate:

[0050] i (t) =σ(W i ·[h (t-1) ,Y (t) ]+b i +δ·M)

[0051] Among them, i (t) This represents the input gate for the current time step. The input gate determines how much information will be written into the memory cell based on the hidden state of the previous time step and the input signal of the current time step. W i It is the weight matrix of the input gate, b i δ is the bias term of the input gate, M represents the delay metric obtained based on the multipath effect, and δ is the adjustment coefficient.

[0052] S424, Calculate candidate memory units:

[0053]

[0054] in, It represents the candidate memory unit at the current time step. tanh is the hyperbolic tangent function, which converts the input value to a value between -1 and 1, representing the possible range of values ​​for the memory unit. W m It is the weight matrix of the candidate memory units, b m It is the bias term for candidate memory units;

[0055] S425, Calculate the output gate:

[0056] o (t) =σ(W o ·[h (t-1) ,Y (t) ]+b o +η·N)

[0057] Among them, o (t) It is the output gate for the current time step. The amount of information output depends on the hidden state of the previous time step and the input signal of the current time step. W o It is the weight matrix of the input gate, b o η is the bias term of the input gate, N represents the parameter of the ambient noise level, and η is the adjustment coefficient that balances the effect of noise on the output gate.

[0058] S426, Update the memory unit based on the forget gate and the input gate:

[0059]

[0060] Where, m (t) The memory unit represents the current time step. ⊙ represents element-wise multiplication. The forget gate controls the proportion of memory information passed down from the previous time step, and the input gate controls the proportion of new candidate memory units. By combining the two parts of information through element-wise multiplication and addition, the memory unit value of the current time step is obtained.

[0061] S427, Update hidden status:

[0062] h (t) =o (t) ⊙tanh(m (t) )

[0063] The output gate controls how information in the memory cell is output to the hidden state. Element-wise multiplication applies the value of the output gate to the value of the memory cell to obtain the hidden state value at the current time step.

[0064] S43 reconstructs the signal Xtrend by outputting the final trend term from the output layer.

[0065] The delay metric is constructed based on the delay difference, and the delay metric is obtained by summing the reciprocals of the delay differences:

[0066]

[0067] in, This indicates the difference in delay.

[0068] Step S5 includes the following steps:

[0069] S51 converts the seasonal signal into a frequency domain signal and processes the frequency domain signal in the encoder. The specific steps are as follows:

[0070] S511 converts the time-domain representation of the seasonal term signal into its frequency-domain representation;

[0071] S512 utilizes a feedforward network module to propagate and process the frequency domain seasonal term signal:

[0072]

[0073] in, It is a frequency domain seasonal term signal. This represents a coded signal with rich feature representation; FeedForward indicates feedforward processing; and Add indicates connection processing.

[0074] S52 uses a decoder to decode the encoded signal, thereby reconstructing the seasonal signal. The specific steps are as follows:

[0075] S521, Reverse processing: The decoder is used to parse the encoded signal and the inverse Fourier transform is used to convert it back to the time domain representation;

[0076] S522, Seasonal Term Signal Reconstruction:

[0077] The parsed coded signal is propagated and processed using a feedforward network module to obtain...

[0078] The propagated signal is summed and concatenated with the analyzed coded signal to obtain the reconstructed seasonal term signal.

[0079] An underwater acoustic signal reconstruction system, comprising:

[0080] The signal acquisition module is used to receive underwater acoustic signals using underwater acoustic sensor nodes;

[0081] The signal decomposition module is used to decompose the underwater acoustic signal to obtain the trend term signal and the seasonal term signal.

[0082] The feature extraction module is used to extract features of the trend term signal based on one-dimensional convolution;

[0083] The trend term signal reconstruction module is used to reconstruct trend term signals based on the extracted features.

[0084] The seasonal term signal reconstruction module is used to reconstruct seasonal term signals based on the encoder-decoder.

[0085] The signal reconstruction module is used to add the reconstructed trend signal and seasonal signal element by element to obtain the reconstructed underwater acoustic signal.

[0086] Compared with the prior art, the present invention has the following beneficial effects:

[0087] (1) This invention utilizes signal decomposition to divide the signal into trend terms and seasonal terms for analysis, thereby enabling better capture of the dynamic changes in the trend and seasonal characteristics of underwater acoustic signals.

[0088] (2) The present invention analyzes the decomposed underwater acoustic trend term signal using a one-dimensional convolution processing method, including performing a sliding window operation on the input signal or data sequence, and performing convolution operation with the data in the sliding window through a set of learnable filters to generate an output sequence. This helps to extract the data processing features of the input signal sequence, thereby improving the accuracy of the subsequent reconstructed signal.

[0089] (3) In the seasonal signal processing, the present invention adopts an encoder and decoder structure. The encoder uses a frequency domain attention mechanism to capture the frequency domain relationship in the underwater acoustic signal, while the decoder uses the frequency domain features extracted by the encoder to make predictions. This structure has higher accuracy and reliability in predicting underwater acoustic signals.

[0090] (4) The reconstructed signal of the present invention has a higher signal-to-noise ratio and a lower root mean square error. The reconstructed signal has high quality and less signal loss during the reconstruction process, and is closer to the original signal.

[0091] (5) The reconstruction method used in this invention has low model complexity and high signal processing efficiency. Attached Figure Description

[0092] Figure 1 This is a flowchart of the method of the present invention;

[0093] Figure 2 This is a flowchart of signal decomposition.

[0094] Figure 3 A flowchart for one-dimensional convolution processing;

[0095] Figure 4 A structural diagram of the long short-term memory network used for reconstructing the trend term signal;

[0096] Figure 5 Flowchart for reconstructing seasonal signals;

[0097] Figure 6 This is a schematic diagram of the reconstructed signal in one embodiment;

[0098] Figure 7 This is a comparison chart of the signal-to-noise ratio of the reconstructed signal of the present invention compared with the prior art in one embodiment;

[0099] Figure 8 This is a root mean square error plot of the present invention at one time point in one embodiment;

[0100] Figure 9 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0101] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0102] This embodiment provides a method for underwater acoustic signal reconstruction, such as... Figure 1 As shown, it includes the following steps:

[0103] S1 uses an underwater acoustic sensor node to receive underwater acoustic signals.

[0104] Considering the effect of multipath propagation on underwater acoustic signals, the signal is represented as:

[0105]

[0106] in, This represents the superposition of multiple signals, each component consisting of amplitude A. i Attenuation coefficient α i The distance r is defined as h(t-τ). i ) represents the oscillation part of the i-th component, and N represents the number of signals.

[0107] The signal is subject to random scattering effects, further representing the underwater acoustic signal as follows:

[0108]

[0109] Where N(t,r) is the noise term; t represents time and r represents distance.

[0110] In one embodiment, the oscillating portion of the underwater acoustic signal is represented as: h(t-τ) i )=cos(2πf i (t-τ i )-k i r+φ i ), where f i τ represents frequency. i k represents the time delay. i φ represents the wave number. i Indicates the initial phase.

[0111] S2 decomposes the underwater acoustic signal to obtain the trend term signal and the seasonal term signal.

[0112] In step S2, the trend term signal is obtained by weighting the underwater acoustic signal and performing a softmax operation, such as... Figure 2 As shown, it includes the following steps:

[0113] S21, Extract trend signal

[0114] S211, Calculate the output of the autocorrelation function:

[0115] Considering the variation of the underwater acoustic signal at different time points, the autocorrelation function is defined as follows:

[0116]

[0117] R(τ) represents the degree of correlation of the original signal P under the relative offset τ between different time points, where τ represents the time delay.

[0118] For each time point, the underwater acoustic signal P(t,r) is extracted, and the time delay τ is added to the current time point t to obtain the new underwater acoustic signal P(t+τ,r) at time point t+τ. The autocorrelation function is then calculated. By calculating the autocorrelation function, the periodicity characteristics of the signal can be understood.

[0119] S212, Define the Gaussian weight vector:

[0120] Define Gaussian function The weight vector used as the values ​​of the weighted autocorrelation function, where τ is the time delay and σ is the standard deviation.

[0121] S213, weighted autocorrelation vector:

[0122] The output of the autocorrelation function is weighted using a Gaussian weight vector to obtain the weighted autocorrelation value R. w (τ)=w(τ)·R(τ).

[0123] S214, applying the softmax function:

[0124] The weighted autocorrelation value R w (τ) is taken as input and transformed into a set of normalized weights S(τ) by the softmax function.

[0125]

[0126] S215, Extract trend signal:

[0127] Using the output S(τ) of the softmax function as weights, the original underwater acoustic signal P(t,r) is weighted and averaged to extract the trend term signal.

[0128] S22, Extract seasonal term signal

[0129] The seasonal term signal is obtained by subtracting the extracted trend term signal from the original underwater acoustic signal, expressed as Sseasonal = P - Strend. This design enables the effective acquisition of seasonal information from the signal in practice, thereby better addressing the needs of periodic signal changes and trend analysis.

[0130] S3 extracts features of the trend term signal based on one-dimensional convolution.

[0131] like Figure 3 As shown, step S3 includes the following steps:

[0132] S31 preprocesses the trend signal by normalizing the data to ensure that the data format meets the requirements of the model.

[0133] S32, In this embodiment, the time window is set to 1000 sampling points. The required number of time windows is obtained by dividing the total number of sampling points by the length of the time window. The trend signal is traversed in chronological order, and the trend signal in each window is extracted into a vector S[t], which is used as the input signal of the convolution module.

[0134] S33 defines the underwater acoustic convolution kernel, which performs a one-dimensional convolution operation on the input signal:

[0135] Define a one-dimensional underwater acoustic convolution kernel K[n], where n = 0, 1, ..., L-1, and L is the length of the convolution kernel. The weight distribution of the underwater acoustic convolution kernel is K[n] = acos(b·n+c)+d, where a is the amplitude, controlling the weight range; b is the modulation factor, adjusting the sensitivity of the underwater acoustic convolution kernel to different paths; c is the phase offset, adjusting the weight of the underwater acoustic convolution kernel to match the path effect in the signal; and d is the DC offset, ensuring that the weight of the underwater acoustic convolution kernel is positive. Apply the underwater acoustic convolution kernel K[n] to the input signal S[t] to perform a one-dimensional convolution operation, obtaining...

[0136] S34, the ReLU activation function is applied to the output of the one-dimensional convolution operation to obtain the feature vector of the trend term signal.

[0137] S4, reconstruct the trend term signal based on the extracted features.

[0138] Step S4 uses a Long Short-Term Memory (LSTM) network to reconstruct the trend term signal, the structure of which is as follows: Figure 4 As shown, it includes the following steps:

[0139] S41, calculate the mean and variance of the feature data y of the trend term signal obtained in step S3, and perform real-time layer normalization based on the mean and variance to improve the network's adaptability to changes in input data, thereby reducing internal covariate bias and improving the stability of network training.

[0140] S411, Calculate the mean μ c :

[0141]

[0142] Where T is the signal length and y(t,c) represents the value of signal y at time t and channel c.

[0143] S412, Calculate variance

[0144]

[0145] S413, real-time layer normalization of underwater acoustic signals:

[0146]

[0147] ∈ is a constant used to prevent overflow of the square root. γ c and β c These are learnable parameters. Real-time layer normalization is implemented in the software part of the data preprocessing module to ensure real-time processing and normalization of the underwater acoustic signal data. After real-time layer normalization, Y = [Y...] (1) ,Y (2) ,...,Y (t) ].

[0148] S42, reconstructing the trend signal, specifically includes the following steps:

[0149] S421, given the initial hidden state h (0) and memory unit m (0) Set it to the zero vector.

[0150] S422, Calculating the forget gate:

[0151] f (t) =σ(W f ·[h (t-1) ,Y (t) ]+b f +γ·D)

[0152] Among them, f (t) It is the forget gate for the current time step, σ is the sigmoid function, and W f It is the weight matrix of the forget gate, h (t-1) It is the hidden state of the previous time step, Y (t) It is the input signal at the current time step, b f γ is the bias term of the forget gate, D is a factor related to the attenuation of the underwater acoustic signal, and γ is an adjustment coefficient used to control the degree of influence of the underwater acoustic signal attenuation factor on the forget gate.

[0153] S423, Calculation Input Gate:

[0154] i (t) =σ(W i ·[h (t-1) ,Y (t) ]+b i +δ·M)

[0155] Among them, i (t) This represents the input gate for the current time step. The input gate determines how much information will be written into the memory cell based on the hidden state of the previous time step and the input signal of the current time step. W i It is the weight matrix of the input gate, b iδ is the bias term of the input gate, M represents the delay metric obtained based on the multipath effect, and δ is the adjustment coefficient.

[0156] The latency metric is constructed based on the latency difference, and the latency metric is obtained by summing the reciprocals of the latency differences:

[0157]

[0158] in, This represents the delay difference; to avoid the denominator being 0, each delay difference is smoothed by adding 1.

[0159] S424, Calculate candidate memory units:

[0160]

[0161] in, It represents the candidate memory unit at the current time step. tanh is the hyperbolic tangent function, which converts the input value to a value between -1 and 1, representing the possible range of values ​​for the memory unit. W m It is the weight matrix of the candidate memory units, b m It is the bias term for candidate memory units.

[0162] S425, Calculate the output gate:

[0163] o (t) =σ(W o ·[h (t-1) ,Y (t) ]+b o +η·N)

[0164] Among them, o (t) It is the output gate for the current time step. The amount of information output depends on the hidden state of the previous time step and the input signal of the current time step. W o It is the weight matrix of the input gate, b o η is the bias term of the input gate, N represents the parameter of the ambient noise level, and η is the adjustment coefficient that balances the effect of noise on the output gate.

[0165] S426, Update the memory unit based on the forget gate and the input gate:

[0166]

[0167] Where, m (t) The memory cell represents the current time step. ⊙ represents element-wise multiplication. The forget gate controls the proportion of memory information passed down from the previous time step, and the input gate controls the proportion of new candidate memory cells. By combining the two parts of information through element-wise multiplication and addition, the memory cell value of the current time step is obtained.

[0168] S427, Update hidden status:

[0169] h (t) =o (t) ⊙tanh(m (t) )

[0170] The output gate controls how information in the memory cell is output to the hidden state. Element-wise multiplication applies the value of the output gate to the value of the memory cell to obtain the hidden state value at the current time step.

[0171] S43 reconstructs the signal Xtrend by outputting the final trend term from the output layer.

[0172] Table 1 shows the impact of the number of nodes and the training period on the Long Short-Term Memory network. It can be seen that as the number of nodes and the training period increase, the BER decreases continuously, eventually dropping to 0.0817, indicating that the signal quality is significantly improved.

[0173] Table 1. Impact of Node Count and Training Period on Long Short-Term Memory Network Output

[0174]

[0175] S5 reconstructs seasonal terms based on encoder-decoder.

[0176] like Figure 5 As shown, step S5 includes the following steps:

[0177] S51 converts the seasonal signal into a frequency domain signal and processes the frequency domain signal in the encoder. The specific steps are as follows:

[0178] S511 converts the time-domain representation of the seasonal term signal into its frequency-domain representation.

[0179] Given a sequence x in the time domain n Where n = 1, 2, ..., N. The Discrete Fourier Transform is: Where X(t) represents the frequency index, i is the imaginary unit, and u represents the frequency index. Discrete Fourier transforms are performed on the query vector q, key vector k, and numerical vector v, resulting in Q[u] = F(q[n]), K[u] = F(k[n]), and V[u] = F(v[n]), yielding the frequency domain representation.

[0180] S512 utilizes a feedforward network module to propagate and process the frequency domain seasonal term signal:

[0181]

[0182] in, It is a frequency domain seasonal term signal. This represents a coded signal with rich feature representation; FeedForward indicates feedforward processing; and Add indicates connection processing.

[0183] S52 uses a decoder to decode the encoded signal, thereby reconstructing the seasonal signal. The specific steps are as follows:

[0184] S521, Reverse Processing: The decoder is used to parse the encoded signal, and the inverse Fourier transform is used to return it to the time domain representation.

[0185]

[0186] Where, d k W is the dimension of the key vector. path [u] is a weighting factor in the frequency domain, mapped from path information. Activation is performed using the softmax function to ensure the effectiveness of the attention weights.

[0187] S522, Seasonal Term Signal Reconstruction:

[0188] The parsed coded signal is propagated and processed using a feedforward network module to obtain...

[0189] The propagated signal is summed and concatenated with the analyzed coded signal to obtain the reconstructed seasonal term signal.

[0190] S6, the reconstructed trend term signal and seasonal term signal are added element-by-element to obtain the reconstructed underwater acoustic signal, expressed as: like Figure 6 As shown, the noise in the signal was removed by the system reconstruction, and the missing signal was successfully predicted.

[0191] This invention mainly includes several steps: signal decomposition, one-dimensional convolution processing, trend term signal reconstruction, seasonal term signal reconstruction, and overall signal reconstruction. The signal decomposition step pre-processes the received underwater acoustic signal, decomposing it into trend and seasonal terms. This step is fundamental to the entire reconstruction process, laying the foundation for subsequent precise processing by distinguishing between long-term trends and periodic changes in the signal. For the decomposed underwater acoustic signal trend term, a one-dimensional convolution processing method is used for analysis. A sliding window operation is performed on the input signal or data sequence, and then a set of learnable filters (convolution kernels) are used to perform convolution operations with the data within the sliding window to generate the output sequence. This step is used for data processing of the input signal sequence, including: an input module for receiving the input signal sequence; a convolution module for determining the size and parameters of the underwater acoustic convolution kernel and performing convolution operations on the sliding underwater acoustic convolution kernel across the input signal sequence; and an output module for generating the output sequence. This step performs in-depth analysis of the signal through the designed underwater acoustic convolution kernel, effectively extracting the temporal features of the signal. This process enhances the system's ability to capture signal changes, especially when processing non-periodic, long-term changing signal components. This invention reconstructs trend term signals based on Long Short-Term Memory (LSTM) networks after one-dimensional convolution processing. Leveraging the powerful temporal data processing capabilities of LSM networks, trend term signals are predicted and reconstructed, especially in cases of missing signal data, enabling accurate prediction of missing signal values ​​based on existing data. It includes a data preprocessing unit and a signal reconstruction module. The data preprocessing unit normalizes the output sequence generated by convolution to improve model training performance. Using historical data of the input signal, the reconstruction module learns and models the temporal characteristics of the signal. During training, the module automatically learns long-term dependencies and patterns in the sequence data, enabling accurate signal prediction. During reconstruction, the input module is known historical signal data, and the signal reconstruction module predicts missing signal values ​​based on the patterns and features learned from the historical data. For seasonal signals, this invention combines an encoder-decoder architecture and a frequency domain attention mechanism. In the encoding stage, the frequency domain attention mechanism deeply analyzes the periodic features in the signal; in the decoding stage, the encoded feature information is used to accurately reconstruct the seasonal signal. This process is specifically optimized for periodic changes in the signal, improving the reconstruction accuracy of periodic signal components. Finally, the prediction results of the trend term and the seasonal term are combined to obtain the final reconstructed signal.

[0192] To evaluate the performance of this invention in effectively reconstructing seasonally affected underwater acoustic signals, we propose the following evaluation criteria: signal-to-noise ratio (SNR), root mean square error (RMSE), and model time complexity. The advantages of this invention compared to other methods include:

[0193] 1. Higher Signal-to-Noise Ratio (SNR): SNR is a metric for measuring signal quality. It describes the relative strength between the signal and noise, a crucial indicator that directly impacts the accuracy of extracting useful information from data. Therefore, this embodiment uses SNR to evaluate signal quality. SNR is calculated by comparing the power of the signal with the power of the noise. A higher SNR value corresponds to a clearer, more reliable signal, meaning the signal has stronger strength or higher quality relative to the noise. SNR is defined in decibels (dB):

[0194]

[0195] Where P s It is the power of the signal, P n It's the power of the noise, the signal-to-noise ratio (SNR) of the signal, for example... Figure 7 . Figure 7 The signal-to-noise ratios (SNRs) of the original signal, the signal reconstructed using Fourier transform, and the signal reconstructed using a reconstruction system are represented. The original signal has an SNR of 0.26, the reconstructed signal has an SNR of 3.29, and the underwater acoustic signal reconstructed using the reconstruction system has an SNR of 8.12, indicating a significant improvement in signal quality.

[0196] 2. Lower Root Mean Square Error: The root mean square error (RMSE) represents the square root of the mean of the squared differences between the estimated and actual values. Because it is the square root of the mean of the squared errors, RMSE considers both the absolute values ​​of the errors and their average values. This makes RMSE very useful in assessing the overall difference between the reconstructed signal and the original signal. In the case of a reconstructed signal, a lower RMSE value means that the difference between the reconstructed signal and the original signal is smaller, the reconstructed signal is closer to the original signal, and less information is lost during the reconstruction process. Conversely, a higher RMSE value indicates that there is a larger difference between the reconstructed signal and the original signal, and the reconstruction effect may be less than ideal.

[0197]

[0198] Where N represents the number of samples, s i This represents the i-th sample point in the original signal. This represents the i-th sample point in the reconstructed signal. The total root mean square error of this invention is 0.19. Figure 8 The root mean square error plot representing different time points can be obtained from... Figure 8 The root mean square error of the reconstructed underwater acoustic signal fluctuates around 0.19, indicating that the reconstructed signal is close to the original signal and very little signal is lost during the reconstruction process.

[0199] 3. Lower Model Time Complexity: This invention proposes a method that uses frequency domain attention instead of traditional time domain attention to improve the efficiency and accuracy of signal processing. Traditional time domain attention mechanisms perform calculations in the time domain, typically involving matrix multiplication and softmax operations, resulting in a time complexity of O(N). 2 Where N is the length of the sequence, the time overhead in the time domain increases with the increase in signal processing requirements. Therefore, this embodiment introduces a frequency domain attention mechanism, using Discrete Fourier Transform (DFT) to transform the underwater acoustic signal from the time domain to the frequency domain, and then calculating the attention in the frequency domain. The DFT is performed on the query vector q, key vector k, and numerical vector v, with a time complexity of O(N log N). The matrix multiplication and softmax operation for calculating the attention require calculating N attention weights, resulting in a matrix multiplication time complexity of O(N log N). 2 The time complexity of the softmax operation is O(N). Therefore, the total time complexity for calculating the attention weights is O(N). 2 After calculating the attention weights, the frequency domain representation of the attention weights needs to be converted back to the time domain representation. The time complexity of the inverse Fourier transform is also O(N log N). The overall time complexity of the frequency domain attention mechanism is mainly determined by the time complexity of the discrete Fourier transform and the inverse Fourier transform, which is O(N log N).

[0200] For temporal attention, the query vector q, key vector k, and numerical vector v need to be stored. The dimensions of these three vectors are typically proportional to the length N of the input sequence. Therefore, the space complexity of temporal attention is mainly determined by the storage space required for these three vectors, which is O(N×D), where D is the dimension of the numerical vector. In frequency domain attention, for an input sequence of length N, the Discrete Fourier Transform produces the same frequency domain representation as the input sequence; therefore, its space complexity remains O(N×D). Table 2 clearly shows the complexity of the model using the frequency and temporal domains.

[0201] Table 2 Model Complexity

[0202] Frequency domain attention model O(N log N) O(N×D) Temporal attention model <![CDATA[O(N 2 )]]> O(N×D)

[0203] This embodiment also provides an underwater acoustic signal reconstruction system, such as Figure 9 As shown, it includes:

[0204] 1) Signal acquisition module, used to receive underwater acoustic signals using underwater acoustic sensor nodes.

[0205] In underwater acoustic communication missions, hydrophones are devices within underwater acoustic sensor nodes. At the start of an underwater communication mission, the hydrophone is lowered to a depth of approximately 50 meters to capture transmitted acoustic signals in the underwater environment. Once positioned correctly, the hydrophone begins receiving acoustic communication signals from the water's surface. These signals propagate through the water via sound waves.

[0206] The underwater data logger converts these acoustic signals into corresponding electrical signals. The received electrical signals are recorded by the underwater data logger and transmitted via an underwater cable to the signal decomposition module for signal decomposition processing.

[0207] The following is a detailed description of the underwater data logger:

[0208] An underwater data logger is a device specifically designed to record and store data in an underwater environment. The data storage unit is the core component of the underwater data logger, used to record and store received signal data. It has sufficient storage capacity to store long-term data recordings and transmits the stored data to other devices for further analysis and processing via a data interface.

[0209] 2) Signal decomposition module, used to decompose underwater acoustic signals to obtain trend term signals and seasonal term signals.

[0210] 3) Feature extraction module, used to extract features of trend term signals based on one-dimensional convolution.

[0211] In this module, parameters including weights and biases are stored in the surface base station. For one-dimensional convolution operations, the convolution computation unit adopts a pipelined architecture, loading the input signal and underwater acoustic convolution kernel parameters into different computation units, and implementing the convolution operation through parallel computation. Furthermore, the feature data processed by the ReLU activation function is transmitted to the surface base station for storage.

[0212] 4) Trend signal reconstruction module, used to reconstruct trend signals based on extracted features.

[0213] 5) Seasonal signal reconstruction module, used for reconstructing seasonal signals based on encoder-decoder.

[0214] 6) Signal reconstruction module, used to add the reconstructed trend signal and seasonal signal element by element to obtain the reconstructed underwater acoustic signal.

[0215] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0216] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for reconstructing underwater acoustic signals, characterized in that, Includes the following steps: S1, using underwater acoustic sensor nodes to receive underwater acoustic signals; S2, perform signal decomposition on the underwater acoustic signal to obtain the trend term signal and the seasonal term signal; S3, extracts features of trend term signals based on one-dimensional convolution; S4, Reconstruct the trend term signal based on the extracted features; S5, seasonal term signal reconstruction based on encoder-decoder; S6, add the reconstructed trend signal and seasonal signal element by element to obtain the reconstructed underwater acoustic signal; Step S4 includes the following steps: S41, The characteristic data of the trend term signal obtained in step S3. y The mean and variance are calculated separately, and real-time layer normalization is performed based on the mean and variance to obtain the preprocessed signal. ; S42, reconstructing the trend signal, specifically includes the following steps: S421, Given the initial hidden state and memory unit ; S422, Calculating the forget gate: in, It is the forgetting gate of the current time step. It is the sigmoid function. It is the weight matrix of the forget gate. It is the hidden state of the previous time step. It is the input signal at the current time step. It is the bias term of the forgetting gate. It is a factor related to underwater acoustic signal attenuation. It is an adjustment coefficient used to control the degree of influence of the underwater acoustic signal attenuation factor on the forgetting gate; S423, Calculation Input Gate: in, This represents the input gate for the current time step. The input gate determines how much information will be written into the memory cell based on the hidden state of the previous time step and the input signal of the current time step. It is the weight matrix of the input gate. It is the bias term of the input gate. This represents a delay metric based on the multipath effect. It is the adjustment coefficient; S424, Calculate candidate memory units: in, It is a candidate memory unit at the current time step. It is the hyperbolic tangent function, which converts the input value into a value between -1 and 1, representing the possible range of values ​​for the memory unit. It is the weight matrix of the candidate memory units. It is the bias term for candidate memory units; S425, Calculate the output gate: in, It is the output gate for the current time step. The amount of information to be output depends on the hidden state of the previous time step and the input signal of the current time step. It is the weight matrix of the input gate. It is the bias term of the input gate. Parameters representing environmental noise levels, It is the adjustment coefficient, which balances the impact of noise on the output gate; S426, Update the memory unit based on the forget gate and the input gate: in, The memory unit representing the current time step. It is element-wise multiplication. The forget gate controls the proportion of memory information passed down from the previous time step, and the input gate controls the proportion of new candidate memory units. By combining the two parts of information through element-wise multiplication and addition, the memory unit value of the current time step is obtained. S427, Update hidden status: The output gate controls how information in the memory cell is output to the hidden state. Element-wise multiplication applies the value of the output gate to the value of the memory cell to obtain the hidden state value at the current time step. S43 reconstructs the signal using the final trend term output from the output layer. .

2. The underwater acoustic signal reconstruction method according to claim 1, characterized in that, In step S1, the underwater acoustic signal received by the underwater acoustic sensor node is represented as follows: in, This represents the superposition of multiple signals, each component consisting of amplitude. attenuation coefficient and distance definition; Indicates the first The oscillating portion of each component; It is a noise term; Indicates time, Indicates distance; N Indicates the number of signals.

3. The underwater acoustic signal reconstruction method according to claim 2, characterized in that, The oscillating component of the underwater acoustic signal is represented as follows: ,in Indicates frequency, Indicates time delay. Indicates wave number, Indicates the initial phase.

4. The underwater acoustic signal reconstruction method according to claim 2, characterized in that, In step S2, the trend term signal is obtained by weighting the underwater acoustic signal and performing a softmax operation, including the following steps: S211, Calculate the output of the autocorrelation function: Considering the variation of the underwater acoustic signal at different time points, the autocorrelation function is defined as follows: Represents the original signal Relative offset between different points in time The degree of relevance, Indicates a time delay; For each time point, extract the underwater acoustic signal. and delay the time Add to the current time node To obtain a new time node underwater acoustic signals Calculate the output of the autocorrelation function; S212, Define the Gaussian weight vector: Define Gaussian function , which is used as the weight vector for the values ​​of the weighted autocorrelation function, where It's a time delay. It is the standard deviation; S213, weighted autocorrelation vector: The output of the autocorrelation function is weighted using a Gaussian weight vector to obtain the weighted autocorrelation value. ; S214, applying the softmax function: Weighted autocorrelation value As input, it is transformed into a set of normalized weights through the softmax function. , S215, Extract trend signal: Using the output of the softmax function As a weight, the original underwater acoustic signal Perform a weighted average to extract the trend signal. .

5. The underwater acoustic signal reconstruction method according to claim 1, characterized in that, In step S2, the seasonal term signal is obtained by subtracting the extracted trend term signal from the original underwater acoustic signal.

6. The underwater acoustic signal reconstruction method according to claim 1, characterized in that, Step S3 includes the following steps: S31, preprocess the trend signal; S32, according to the preset time window size, traverse the trend signal in chronological order, and extract the trend signal in each window into a vector. , as the input signal for the convolution module; S33 defines the underwater acoustic convolution kernel, which performs a one-dimensional convolution operation on the input signal: Define a one-dimensional underwater acoustic convolution kernel ,in , Where is the length of the convolution kernel, and the weight distribution of the underwater acoustic convolution kernel is as follows: ,in, It's the amplitude, which controls the weight range. It is a modulation factor that adjusts the sensitivity of the underwater acoustic convolution kernel to different paths. It involves phase shifting, adjusting the weights of the underwater acoustic convolution kernel to match path effects in the signal. This is a DC offset, ensuring the weights of the underwater acoustic convolution kernel are positive; the underwater acoustic convolution kernel... Applied to input signals Perform a one-dimensional convolution operation on the top to obtain... ; S34, the ReLU activation function is applied to the output of the one-dimensional convolution operation to obtain the feature vector of the trend term signal.

7. The underwater acoustic signal reconstruction method according to claim 1, characterized in that, The delay metric is constructed based on the delay difference, and the delay metric is obtained by summing the reciprocals of the delay differences: in, This indicates the difference in delay.

8. The underwater acoustic signal reconstruction method according to claim 1, characterized in that, Step S5 includes the following steps: S51 converts the seasonal signal into a frequency domain signal and processes the frequency domain signal in the encoder. The specific steps are as follows: S511 converts the time-domain representation of the seasonal term signal into its frequency-domain representation; S512 utilizes a feedforward network module to propagate and process the frequency domain seasonal term signal: in, It is a frequency domain seasonal term signal. This represents a coded signal with rich feature representation. Indicates feedforward processing. Indicates connection processing; S52 uses a decoder to decode the encoded signal, thereby reconstructing the seasonal signal. The specific steps are as follows: S521, Reverse processing: The decoder is used to parse the encoded signal and the inverse Fourier transform is used to convert it back to the time domain representation; S522, Seasonal Term Signal Reconstruction: The parsed coded signal is propagated and processed using a feedforward network module to obtain... ; The propagated signal is summed and concatenated with the analyzed coded signal to obtain the reconstructed seasonal term signal. , .

9. A system for implementing the underwater acoustic signal reconstruction method as described in any one of claims 1-8, characterized in that, include: The signal acquisition module is used to receive underwater acoustic signals using underwater acoustic sensor nodes; The signal decomposition module is used to decompose the underwater acoustic signal to obtain the trend term signal and the seasonal term signal. The feature extraction module is used to extract features of the trend term signal based on one-dimensional convolution; The trend term signal reconstruction module is used to reconstruct trend term signals based on the extracted features. The seasonal term signal reconstruction module is used to reconstruct seasonal term signals based on the encoder-decoder. The signal reconstruction module is used to add the reconstructed trend signal and seasonal signal element by element to obtain the reconstructed underwater acoustic signal.

Images