ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN

Abstract

This invention proposes an ADCP echo signal denoising method based on overlapping segmentation DRSN-LSTM-CNN, belonging to the field of deep learning technology. This invention addresses the inherent limitations of existing ADCP signal processing methods when dealing with nonlinear, non-stationary, and low signal-to-noise ratio signals, and the lack of deep adaptation and optimization of existing machine learning methods for ADCP signal characteristics. The steps of this invention include: S1: Segmenting the noisy ADCP signal into noisy segments using a sliding window; S2: Inputting each segment into a DRSN module; S3: Performing a weighted average on the signal output from the DRSN module; S4: Sequentially inputting the signal into a bidirectional LSTM network and a CNN network, outputting the denoised ADCP signal.

Description

Technical Field

[0001] This invention relates to an ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN, belonging to the field of deep learning technology. Background Technology

[0002] As a key instrument for hydrological measurement and oceanographic exploration, the accuracy of ADCP current measurement directly depends on the accurate estimation of the Doppler frequency shift in the echo signal. However, the complexity of the actual underwater acoustic environment and the characteristics of the ADCP signal itself pose serious challenges to traditional signal processing methods.

[0003] ADCP echo signals are inherently non-stationary and nonlinear. Their non-stationarity stems from the time-varying characteristics of the aquatic environment, such as water surface fluctuations caused by waves, changes in water flow due to sudden wind changes, and the motion of the ADCP carrier (e.g., a ship). These factors collectively cause the Doppler frequency shift to change dynamically over time. The nonlinearity manifests as the modulation of the echo signal by the complex motion trajectories of the scatterer caused by turbulence, eddies, etc., and the nonlinear distortion of signal amplitude and phase caused by sudden changes in scatterer concentration at interfaces such as thermoclines and turbidity strata.

[0004] In practical applications, ADCP echo signals are highly susceptible to environmental noise (such as water flow noise and biological noise), equipment noise (such as oscillator phase noise and amplifier thermal noise), and multipath interference, leading to a decrease in signal-to-noise ratio (SNR). Especially in long-distance measurements or in highly turbid water bodies, the effective signal is weak, and the frequency estimation accuracy and stability of traditional algorithms drop sharply under low SNR conditions.

[0005] In the traditional ADCP field, the Complex Autocorrelation Algorithm (CAAL) is commonly used as the core frequency estimation method for echo signal processing. This algorithm, based on time-domain phase analysis, has the advantages of low computational complexity and ease of real-time implementation. However, its inherent limitations are also quite obvious: 1) Sensitive to noise: In low signal-to-noise ratio environments, noise can severely interfere with phase estimation, leading to increased velocity estimation errors. 2) Limited frequency resolution: Limited by the correlation window length, its ability to resolve rapidly changing flow fields is insufficient. 3) Lack of long-term dependency modeling capability: As a local processing method based on a short-time window, the CAAL algorithm struggles to capture and utilize long-range temporal dependencies in the signal, and cannot effectively model the continuous evolution trend of the flow field. 4) Strong dependency assumptions: The algorithm's performance largely depends on the stability of the transmitting equipment, resulting in poor adaptability to non-ideal operating conditions.

[0006] In recent years, machine learning and deep learning technologies have demonstrated tremendous potential in the field of nonlinear and non-stationary signal processing, providing new approaches to solving complex signal processing problems. Various deep learning models have been successfully applied in related fields. CNNs, with their local feature extraction capabilities, excel in image and vibration signal processing, such as image recognition and fault diagnosis. LSTM and its variants effectively capture temporal dependencies through gating mechanisms and are widely used in temporal modeling tasks such as power load forecasting and speech enhancement. Some hybrid models, such as CNN-LSTM and CAE-M, further enhance their ability to model spatiotemporal features by combining the advantages of different networks, achieving performance superior to single models in tasks such as anomaly detection and wind speed prediction.

[0007] Despite the significant achievements of deep learning technology in many fields, its application in the specific domain of ADCP signal processing remains insufficient. Most existing research directly applies general-purpose deep learning models to ADCP signals, lacking targeted modeling of the unique physical mechanisms (such as acoustic scattering and the Doppler effect) and signal characteristics (such as nonlinear coupling and spatiotemporal nonstationarity) of ADCP echo signals. Most methods focus on end-to-end mapping learning, failing to fully utilize the deep physical features and long-range temporal context information in ADCP signals, making it difficult to handle signals under complex water flow conditions. Deep learning models typically require large amounts of labeled data for training, but obtaining a large number of high-precision ADCP noisy signals and their corresponding clean signals in real-world environments is extremely costly and difficult, casting doubt on the model's generalization ability in complex real-world environments. Some complex deep models (such as deep Transformers) have high computational costs, making it difficult to meet the stringent real-time signal processing requirements of ADCP systems. Summary of the Invention

[0008] This invention addresses the inherent limitations of existing ADCP signal processing methods when dealing with nonlinear, non-stationary, and low signal-to-noise ratio signals, as well as the lack of deep adaptation and optimization of existing machine learning methods to ADCP signal characteristics. Therefore, it proposes an ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN.

[0009] The technical solution adopted by the present invention to solve the above problems is as follows: The present invention includes the following steps: Step 1: Receive the noisy ADCP signal and use the sliding window method to perform overlapping segmentation on the noisy ADCP signal; Step 2: Input each ADCP signal after overlapping segmentation into the DRSN module for feature extraction, feature learning, residual connection and dimensionality reduction in sequence; Step 3: Perform a weighted average of the ADCP signals of each segment output by the DRSN module to obtain the complete ADCP signal after preliminary denoising; Step 4: Input the pre-denoised complete ADCP signal into a bidirectional LSTM network to obtain the time-series feature sequence; Step 5: Input the temporal feature sequence into the CNN network, perform local refinement processing on the temporal feature sequence to obtain the refined feature representation, and generate the final denoised ADCP signal based on the refined feature representation.

[0010] Furthermore, step 1 specifically includes: Receive a noisy ADCP signal, where the noisy ADCP signal is a one-dimensional time series, with a fixed window length of L and an overlap rate of [value missing]. The step size for each slide is The sliding window method is used to overlap and segment the noisy ADCP signal, and output multiple ADCP signal segments.

[0011] Furthermore, the DRSN module in step 2 includes a feature extraction submodule, an attention mechanism submodule, and a feature compression submodule connected in sequence; The feature extraction submodule includes two consecutive convolutional layers plus a batch normalization layer. The two consecutive convolutional layers plus the batch normalization layer contain a ReLU activation function layer. The feature extraction submodule is used to extract the features of each ADCP signal segment after overlapping segmentation to obtain a feature map. The attention mechanism submodule consists of an adaptive average pooling layer, a convolutional layer, a ReLU activation function layer, a convolutional layer, and a sigmoid activation function layer connected in sequence. The attention mechanism submodule is used to adaptively learn the weights of different positions in the extracted feature map, suppress noise-related features, and after completing feature learning, add the weighted features to the original input features of the ADCP signal segment through residual connections. The feature compression submodule includes a ReLU activation function layer and a convolutional layer. The feature compression submodule is used to reduce the dimensionality of the segmented features of the ADCP signal after residual connection and compress the number of output channels of each segment to 1.

[0012] Furthermore, step 3 specifically includes: Determine the overlapping region of the ADCP signals by analyzing the time positions of each segment of the ADCP signal output by the DRSN module. Based on the signal values ​​corresponding to each segment within the overlapping region and the number of times they have been processed, a weighted average is calculated to obtain the complete ADCP signal after preliminary denoising.

[0013] Furthermore, the bidirectional LSTM network and the CNN network are connected sequentially; The complete ADCP signal after initial denoising is input into a bidirectional LSTM network. The hidden layer size of the bidirectional LSTM network is set according to the length of the input ADCP signal. The bidirectional LSTM network captures the long-term time dependence of the complete ADCP signal after initial denoising. The dynamic pattern of the ADCP signal in the time dimension is learned by using a gating mechanism to correct the temporal inconsistency in the initial denoising result and obtain the temporal feature sequence.

[0014] Furthermore, the CNN network in step 5 includes a first convolutional layer, a normalization layer, a ReLU activation function layer, and a second convolutional layer connected in sequence. By performing dimensionality transformation on the temporal feature sequence, convolutional input data can be obtained; The first convolutional layer extracts features from the convolutional input data. A normalization layer accelerates the training of the CNN network. A ReLU activation function layer is used to introduce non-linear expressive power and alleviate the gradient vanishing problem. Finally, a second convolutional layer is used for feature mapping and dimensionality reduction to obtain a refined feature representation of the single-channel ADCP signal. Based on the refined feature representation, the final denoised ADCP signal is generated.

[0015] The beneficial effects of this invention are: 1. The DRSN-LSTM-CNN joint denoising model constructed in this invention exhibits leading overall performance in 0-20dB full-noise scenarios, with an average SNR improvement of 5.39dB and an SSIM of 0.95. It demonstrates significant advantages in low-noise scenarios without signal distortion, while requiring only 1367 parameters, balancing accuracy and engineering practicality. Ablation experiments verify that DRSN, LSTM, CNN, and the attention module are all indispensable, working synergistically to achieve noise suppression, temporal modeling, and signal fidelity preservation. Removing any of these modules leads to a significant performance degradation, highlighting the scientific rigor of the proposed solution.

[0016] 2. This invention addresses the core pain points of non-stationarity and low signal-to-noise ratio of ADCP signals through the innovative integration of overlapping segmentation mechanism and hybrid deep learning architecture. It breaks through the limitations of traditional algorithms and general models, combining creativity and uniqueness, and can efficiently adapt to the needs of ADCP signal processing. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN. Figure 2 This is a schematic diagram of the DRSN module. Detailed Implementation

[0018] In this embodiment, the DRSN module is a deep residual shrinking network module, LSTM is a long short-term memory network, CNN is a convolutional neural network, and ADCP is an acoustic Doppler current profiler. Figure 1 As shown, the steps of the ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN described in this embodiment include: S1: Perform sliding window segmentation on the noisy ADCP signal; The system receives a noisy ADCP signal as input, which is a one-dimensional time series. A sliding window method is used to segment the signal into overlapping segments, with the window length (i.e., the segment length) fixed. The overlap rate is set to The step size for each slide is . In the actual experiment, an overlap rate of 0.2 was set to achieve the best noise reduction effect. The value is determined by the actual signal length.

[0019] S2: Input each signal segment into the DRSN module; Each signal segment is input into a DRSN module for processing. The structure of the DRSN module is as follows: Figure 2 As shown, it includes a feature extraction submodule, an attention mechanism submodule, and a feature compression submodule connected in sequence. The feature extraction submodule consists of two consecutive convolutional layers plus a batch normalization layer. The two consecutive convolutional layers plus the batch normalization layer contain a ReLU activation function layer. The feature extraction submodule is used to extract features from each ADCP signal segment after overlapping segmentation, obtaining a feature map. The attention mechanism submodule consists of an adaptive average pooling layer, a convolutional layer, a ReLU activation function layer, another convolutional layer, and a sigmoid activation function layer connected in sequence. The attention mechanism submodule is used to adaptively learn the weights at different positions in the extracted feature map, suppressing noise-related features. After feature learning, the weighted features are added to the original input features of the ADCP signal segments through residual connections. The feature compression submodule includes a ReLU activation function layer and a convolutional layer. The feature compression submodule is used to reduce the dimensionality of the ADCP signal segment features after residual connections and compress the number of output channels for each segment to 1.

[0020] The workflow of the DRSN module is as follows: First, preliminary features of the segmented signal are extracted through convolutional layers and batch normalization layers. Then, an attention mechanism module consisting of global pooling layers and fully connected layers adaptively learns the weights at different locations in the feature map, highlighting important features and suppressing noise-related features. After learning the features, the weighted features are added to the original input features through residual connections to alleviate the gradient vanishing problem and retain more original information. Finally, a... The convolutional layer reduces the dimensionality of the features, compressing the number of output channels for each segment to 1, resulting in a processed single-channel signal segment.

[0021] S3: Perform a weighted average on the signal output by the DRSN module; The overlapping region of the ADCP signal is determined by analyzing the time positions of each segment of the signal output from the DRSN module. Based on the signal values ​​of each segment within the overlapping region and the number of times they have been processed, a weighted average is calculated to obtain the complete ADCP signal after preliminary denoising.

[0022] S4: Input the signal sequentially into the bidirectional LSTM network and the CNN network, and output the denoised ADCP signal; The signal processed by S3 is fed into a bidirectional LSTM network for further processing. The hidden layer size of the bidirectional LSTM network is set by the length of the input ADCP signal. The LSTM network can capture long-term temporal dependencies in the signal and learn the dynamic patterns of the signal in the time dimension using its gating mechanism (input gate, forget gate, output gate), further correcting potential temporal inconsistencies in the initial denoising results. After dimensionality transformation, the output of the LSTM network is fed into a CNN module. This CNN consists of two convolutional layers, activation function layers, and batch normalization layers. The first convolutional layer extracts features from the convolutional input data, the normalization layer accelerates the training of the CNN network, and the ReLU activation function layer introduces non-linear expressive power and alleviates the gradient vanishing problem. Then, the second convolutional layer performs feature mapping and dimensionality reduction to obtain a refined feature representation of the single-channel ADCP signal. Based on the refined feature representation, the final denoised ADCP signal is generated. Finally, local refinement processing is performed at the feature level to further filter out residual noise and enhance the consistency of the signal waveform.

[0023] To fully verify the noise reduction effect of the present invention, this embodiment conducts system testing through comparative experiments and ablation experiments.

[0024] 1. Overall experimental condition design; The core objectives of the experiment are: first, to verify the performance advantages of this model compared to existing mainstream technologies through comparative experiments; and second, to verify the necessity and synergistic effect of each core component module of this model through ablation experiments. To ensure the objectivity, fairness, and reliability of the experimental results, the following overall experimental conditions and parameters are uniformly set: (1) Experimental samples and data sources; The experiment used a simulation dataset based on the principle of spherical echo superposition as the test sample, totaling 36,000 valid samples. The training, validation, and test sets were divided in a 7:2:1 ratio. All samples underwent preprocessing to ensure the consistency of the input data. Gaussian noise was used for the noise data, and test samples with different noise intensities were artificially added to simulate complex noise environments in real-world applications.

[0025] (2) Experimental evaluation index system; The experiment employs a multi-dimensional evaluation index system to comprehensively measure model performance. The main evaluation indexes are as follows: 1) Signal-to-noise ratio (SNR) improvement: defined as the difference between the SNR value of the signal after model processing and the SNR value of the original noisy signal. It is used to evaluate the noise suppression and signal recovery capabilities of the model. The larger the difference, the better the performance. (2) Pearson correlation coefficient: used to evaluate the linear correlation between the model output signal and the original clean signal. The value range is [-1, 1]. The closer it is to 1, the higher the signal fidelity. (3) Structural Similarity Index (SSIM): Used to evaluate the ability of a signal or image to preserve its structure. The value range is [0,1], and the closer it is to 1, the better the structure preservation effect. (4) Residual spectrum flatness: used to evaluate the stationarity of the residual signal and the degree of noise residue. The closer the value is to 1, the more stable the signal and the less noise residue. (3) Experimental model and parameter setting; 1) Comparative Experimental Models: Five mainstream models in the current field were selected as comparison objects, including CNN, LSTM, U-Net (encoder-decoder network), DnCNN (deep denoising convolutional neural network), and TCN (temporal convolutional network). All comparison models adopted standard structures and optimal parameter configurations obtained through grid optimization. Furthermore, wavelet denoising was also compared. [8] The methods were compared.

[0026] 2) Ablation Experiment Model: Based on the core structure of this model, four ablation variant models were constructed, including: No-DRSN (removing the DRSN module of this model), No-LSTM (removing the LSTM module of this model), No-CNN (removing the CNN feature extraction module of this model), and No-Attention (removing the attention mechanism module in the DRSN block), to verify the necessity of each core module in terms of improving the signal-to-noise ratio and preserving the signal structure.

[0027] 3) Unified parameter settings: All models (including this model, the comparison model, and the ablation variant model) use the same training and testing parameters: the Adam optimizer is used, the learning rate is set to 1e-4, the number of training iterations is 1000, and the following early stopping strategy is set: if the loss on the validation set does not improve after 20 rounds of training, then training is stopped. The batch size is set to 32. In order to optimize the denoising effect and signal fidelity at the same time, a hybrid loss function combining scale-invariant signal-to-noise ratio loss (SiSnrLoss) and logarithmic spectral distance loss (LsdLoss) is used to guide the model training. Among them, SiSnrLoss is the main loss function, and its expression is shown in formula (1). (1); In formula (1), The denoised signal output by the DRSN-LSTM-CNN model. To ensure a true and pure signal, this loss function is insensitive to changes in signal amplitude and effectively adapts to the amplitude fluctuation characteristics of ADCP flow velocity signals. LsdLoss, as an auxiliary loss function, is calculated... With real signals The Euclidean distance between the logarithmic power spectra, constraining the residual noise in the frequency domain, is expressed as formula (2): (2); In formula (2), , These are the power spectra of the denoised signal and the real signal, respectively. The final expression of the hybrid loss function is shown in equation (3): (3); In formula (3), during the training process of the DRSN-LSTM-CNN model, take , By weighting and balancing the optimization objectives in the time and frequency domains, the denoising performance of the model is improved, and the denoised waveform is guaranteed not to be severely distorted.

[0028] (4) Environment and equipment; The experiment was implemented using the PyTorch 2.0.1 + cu118 deep learning framework, with an Intel Core i7-7700K CPU and an NVIDIA GTX 1070 GPU.

[0029] (5) Experimental testing scenarios; To fully verify the robustness of the model, six different noise intensity scenarios were set up for the experiment, corresponding to SNR levels of 0dB, 1dB, 2dB, 5dB, 10dB, and 20dB, respectively, covering the full range of scenarios from low noise (20dB), medium noise (5dB, 10dB), to high noise (0dB, 1dB, 2dB). Each scenario was tested three times, and the average value was taken as the final experimental result to reduce the impact of random errors on the experimental conclusions.

[0030] 2. Performance comparison and analysis with other methods; (1) Improvement in signal-to-noise ratio; Signal-to-noise ratio (SNR) refers to the ratio of signal strength to noise strength, typically the ratio of signal power to noise power. Its calculation formula is shown in formula (4): (4); In formula (4), The power of the pure signal, Let be the power of the noise signal. In this model, the signal-to-noise ratio before denoising is determined by the signal-to-noise ratio set during simulation.

[0031] The signal-to-noise ratio (SNR) improvement test results are shown in Table 1. As can be seen from Table 1, the average SNR improvement of this invention reaches 5.39 dB, significantly better than all comparative models, demonstrating the best overall noise reduction performance. In low-noise intensity scenarios (0 dB, 1 dB, 2 dB), the advantages of this model are particularly prominent: the SNR improvement reaches 7.03 dB at 0 dB, and 6.85 dB and 6.57 dB at 1 dB and 2 dB respectively, far exceeding comparative models such as CNN (5.82 dB, 5.72 dB, 5.50 dB) and U-Net (5.64 dB, 5.60 dB, 5.45 dB), exhibiting extremely strong noise suppression capabilities. In medium-to-high noise intensity scenarios (5dB, 10dB, 20dB), this invention maintains a stable positive SNR improvement: 5.73dB at 5dB, 4.29dB at 10dB, and 1.66dB at 20dB. In contrast, LSTM, DnCNN, and wavelet denoising models show negative improvements in the 20dB scenario, indicating that they can damage the original signal in weak noise environments. It is worth noting that U-Net's 2.10dB improvement at 20dB is slightly higher than this invention, but considering the overall performance in low-to-medium noise scenarios, the DRSN-LSTM-CNN model demonstrates a significant overall advantage.

[0032] Table 1

[0033] (2) Pearson correlation coefficient; In this invention, the Pearson correlation coefficient is used to measure the waveform similarity between the denoised waveform and the clean waveform, reflecting whether there is phase distortion in the denoised signal. The Pearson correlation coefficient is a statistic that measures the degree and direction of the linear correlation between two variables. Its value ranges from... The calculation formula for this is shown in formula (5): (5); In formula (5), For the sample size, and The first The value of each sample and These are the means of the two samples, respectively.

[0034] The Pearson correlation coefficient test results are shown in Table 2. As shown in Table 2, the average Pearson correlation coefficient of the DRSN-LSTM-CNN model reaches 0.95, second only to the LSTM model (0.92), and significantly better than the negatively correlated comparison models such as CNN (-0.94) and DnCNN (-0.93), demonstrating excellent overall signal fidelity performance. Under all noise intensity scenarios, the DRSN-LSTM-CNN model maintains high positive correlation characteristics: the correlation coefficient is 0.91 at 0dB, 0.93 and 0.94 at 1dB and 2dB respectively, and further increases to 0.96, 0.98, and 1.00 at 5dB, 10dB, and 20dB. As the noise intensity decreases, the correlation gradually approaches 1, indicating a continuous improvement in signal recovery accuracy.

[0035] Table 2

[0036] Among the comparative models, LSTM's average correlation coefficient (0.92) is close to that of our model, but its correlation coefficient in low-noise scenarios (0dB) is 0.86, lower than our model's 0.91. U-Net's average correlation coefficient is 0.94, slightly lower than our model, and it is 0.89 in high-noise scenarios (0dB), indicating slightly inferior fidelity performance. CNN and DnCNN both show significant negative correlation in fully noisy scenarios, indicating a large deviation between their output signals and the original signals, resulting in extremely poor fidelity performance. TCN and wavelet denoising have average correlation coefficients of 0.86 and 0.87 respectively, with generally average fidelity performance. (3) SSIM; SSIM was originally used to measure the structural similarity between two images. In one-dimensional signal denoising tasks, it can be used to measure whether a signal has severe temporal distortion from a structural perspective. Generally, its calculation formula is shown in equation (6): (6); In formula (6), and The mean of the signal. and The standard deviation of the signal. Let be the covariance of the two signals, and take . .

[0037] The SSIM test results for each model are shown in Table 3. As can be seen from the table, the DRSN-LSTM-CNN model achieves an average SSIM of 0.95, significantly outperforming all the comparison models and demonstrating excellent signal structure preservation capabilities. Under all noise levels, this model maintains a high SSIM value: in the 0dB strong noise scenario, the SSIM is 0.90, increasing to 0.92 and 0.93 in the 1dB and 2dB scenarios, respectively; as the noise intensity decreases, the SSIM further improves to 0.95, 0.98, and 0.99 in the 5dB, 10dB, and 20dB scenarios, respectively, indicating that the DRSN-LSTM-CNN model can effectively preserve the structural features of the original signal under different noise interference conditions. Among the comparative models, U-Net has an average SSIM of 0.93, slightly lower than our model, and only 0.85 in the 0dB scenario, indicating slightly inferior structure preservation performance; LSTM has an average SSIM of 0.76, and 0.61 in the low-noise scenario (0dB), indicating significant structure distortion; CNN and DnCNN have average SSIMs of only 0.10 and 0.05 respectively, indicating extremely poor structure preservation ability; TCN and wavelet denoising have average SSIMs of 0.66 and 0.79 respectively, indicating average overall structure preservation performance, and none of them can reach the index of our model.

[0038] Table 3

[0039] (4) Residual spectrum flatness; Residual spectral flatness is used to evaluate the stationarity of residual signals. Its core idea is to utilize the characteristic that the closer the residual signal is to white noise (the more uniform the spectral distribution), the closer the flatness value is to 1. This measures the completeness of the model's preservation of effective information in the original signal and the thoroughness of noise removal, and can evaluate denoising performance from a frequency domain perspective. Residual spectral flatness is quantified by calculating the ratio of the standard deviation to the mean of the residual signal's power spectral density; the smaller the ratio, the higher the flatness. Its expressions are shown in formulas (7) and (8): (7); (8); In formulas (7) and (8), The signal after denoising. For a pure signal, For residual signals, This represents the power spectrum of the residual signal.

[0040] The residual spectrum flatness test results for each model are shown in Table 4. As can be seen from the table, the average residual spectrum flatness of the DRSN-LSTM-CNN model reaches 0.93, which is excellent overall. It is only lower than LSTM (0.97), TCN (0.98), and wavelet denoising (1.00), and significantly better than the comparative models such as CNN (0.80), U-Net (0.71), and DnCNN (0.18), demonstrating excellent residual signal stabilization capabilities.

[0041] Table 4

[0042] Under all noise levels, the residual spectral flatness of the DRSN-LSTM-CNN model remains stable. It is 0.94 in the 0dB strong noise scenario, 0.93 and 0.94 in the 1dB and 2dB scenarios respectively, and remains at 0.94, 0.93, and 0.93 in the 5dB, 10dB, and 20dB scenarios, with minimal fluctuations. This indicates that the DRSN-LSTM-CNN model can achieve a balance between effective noise removal and residual signal stabilization under different noise interference conditions.

[0043] In the comparison models, although LSTM, TCN, and wavelet denoising have slightly higher average flatness than this model, their denoising performance has significant shortcomings when considering the core indicators such as signal-to-noise ratio improvement and signal fidelity mentioned above. U-Net has an average residual spectral flatness of 0.71, which is lower than this model in all noise scenarios, indicating poor residual signal stability. CNN has an average flatness of 0.80, and the residual signal still has some spectral inhomogeneity. DnCNN has an average flatness of only 0.18, and the spectral distribution of the residual signal is extremely uneven, with prominent issues of noise residue and loss of effective information, far inferior to the residual processing effect of the DRSN-LSTM-CNN model.

[0044] (5) Computational complexity The test results for each model's parameter count are shown in Table 5, using the number of model parameters as an indicator of computational complexity.

[0045] Table 5

[0046] As shown in Table 5, the DRSN-LSTM-CNN model has only 1367 parameters. While maintaining excellent performance, this number is significantly lower than that of comparable models such as CNN (11137), LSTM (8865), U-Net (2903), and DnCNN (6464), and only slightly higher than TCN (1745). Specifically, the DRSN-LSTM-CNN model has only 12.28% of the parameters of CNN, 15.42% of LSTM, 47.09% of U-Net, and 21.15% of DnCNN. Even compared to TCN, which has the closest parameter count, it is reduced by 21.66%, while TCN has significant shortcomings in noise reduction performance and signal fidelity.

[0047] This result demonstrates that the DRSN-LSTM-CNN model achieves high-performance noise reduction while effectively controlling computational complexity, exhibiting a significant advantage in parameter quantity, and possessing superior feasibility for engineering applications. It can better meet the requirements of real-time performance and hardware resource consumption in practical scenarios.

[0048] (6) Summary of comparative experiments; Comparative experimental results show that the DRSN-LSTM-CNN model significantly outperforms six mainstream comparison models, including CNN, LSTM, and U-Net, in overall performance across all noise levels. Its core advantages are reflected in the following three aspects: First, outstanding noise reduction performance, with an average signal-to-noise ratio improvement of 5.39 dB, particularly significant in low-noise scenarios, maintaining a positive improvement across all scenarios without signal distortion; second, excellent signal fidelity and structure preservation capabilities, with an average Pearson correlation coefficient of 0.95 and an average SSIM of 0.95, both leading levels, accurately restoring the original signal characteristics; third, strong engineering practicality, with a total model parameter count of only 1367, lower than all comparison models, and controllable computational complexity, fully demonstrating the advanced nature of the technical solution of this invention.

[0049] 3. Ablation experiment analysis; To verify the necessity and synergistic effect of each core module (DRSN module, LSTM module, CNN feature extraction module, and attention mechanism module) of the DRSN-LSTM-CNN model, four ablation variant models were constructed: No-DRSN (DRSN module removed), No-LSTM (LSTM module removed), No-CNN (CNN feature extraction module removed), and No-Attention (attention mechanism module removed). According to the evaluation metrics mentioned above, tests were conducted in six noise scenarios of 0dB, 1dB, 2dB, 5dB, 10dB, and 20dB, and the test conditions were consistent with the comparative experiments.

[0050] (1) Improvement in signal-to-noise ratio; The signal-to-noise ratio improvement test results of each ablation variant model and this model are shown in Table 6.

[0051] Table 6

[0052] As shown in Table 6, the proposed DRSN-LSTM-CNN model achieves an average SNR improvement of 5.39 dB, significantly higher than all ablation variant models, indicating that the introduction of each core module is crucial for improving denoising performance. Specifically, the No-DRSN model only improves the average SNR by 1.80 dB, a decrease of 66.61% compared to our model, demonstrating the key role of the DRSN module in enhancing noise separation capabilities; the No-LSTM model improves the average SNR by 3.02 dB, a decrease of 44.01%, verifying the LSTM module's effect on suppressing temporal signal noise; the No-CNN model improves the average SNR by 3.50 dB, a decrease of 35.07%, reflecting the importance of the CNN feature extraction module in capturing local noise features; and the No-Attention model improves the average SNR by 2.58 dB, a decrease of 52.13%, indicating that the attention mechanism module can effectively focus on key signal regions and improve denoising accuracy. Under all noise levels, the performance degradation trend of each ablation variant model is consistent, with particularly significant performance differences in low noise levels (0dB, 1dB, 2dB). In the 0dB scenario, the DRSN-LSTM-CNN model improves the signal-to-noise ratio by 7.03dB, while the No-DRSN, No-LSTM, No-CNN, and No-Attention models only achieve 3.99dB, 5.25dB, 5.02dB, and 5.58dB respectively, all lower than the DRSN-LSTM-CNN model.

[0053] (2) Pearson correlation coefficient The waveform similarity test results between the various ablation variant models and this model are shown in Table 7.

[0054] Table 7

[0055] As shown in Table 7, the average Pearson correlation coefficient of the DRSN-LSTM-CNN model proposed in this invention reaches 0.95, significantly higher than all ablation variant models, indicating that the synergistic effect of each core module is crucial to ensuring signal fidelity. The performance degradation of each variant model is as follows: the average Pearson correlation coefficient of the No-DRSN model is only -0.91, a significant decrease compared to the proposed model, indicating that the DRSN module plays an irreplaceable role in maintaining the linear correlation between the output signal and the original clean signal; the average Pearson correlation coefficient of the No-LSTM model is 0.93, a decrease of 2.11%, verifying the positive contribution of the LSTM module in improving the fidelity of temporal signals; the average Pearson correlation coefficient of the No-CNN model is -0.93, with extremely significant performance degradation, reflecting the key value of the CNN feature extraction module in accurately capturing signal features and reducing signal distortion; the average Pearson correlation coefficient of the No-Attention model is 0.93, a decrease of 2.11%, indicating that the attention mechanism module can further improve signal fidelity by focusing on key signal regions.

[0056] Under all noise levels, the performance shortcomings of each ablation variant model are evident. In a 0dB strong noise scenario, the Pearson correlation coefficient of our model is 0.91, while the No-DRSN and No-CNN models are as low as -0.85 and -0.87, respectively, showing a significant negative correlation and severe signal distortion. Although the No-LSTM and No-Attention models maintain a positive correlation, their correlation coefficients are 0.88 and 0.88, respectively, lower than the DRSN-LSTM-CNN model. As the noise intensity decreases, the performance gap between each ablation variant model and the DRSN-LSTM-CNN model continues to exist.

[0057] (3) SSIM; The SSIM test results of each ablation variant model and the DRSN-LSTM-CNN model are shown in Table 8. SSIM is used to measure the structural consistency between the output signal and the original clean signal. The closer the value is to 1, the better the structure preservation effect. The average SSIM of this model is 0.95, which is significantly higher than all ablation variant models, confirming the key value of the synergistic effect of each core module in ensuring the integrity of the signal structure.

[0058] Table 8

[0059] In terms of specific performance degradation, the No-DRSN model has an average SSIM of only -0.07, a significant decrease compared to our model, indicating that the DRSN module plays an irreplaceable role in maintaining signal structure stability. The No-LSTM model has an average SSIM of 0.86, a decrease of 9.47% compared to our model, verifying the contribution of the LSTM module to preserving the structural features of temporal signals. The No-CNN model has an average SSIM of 0.10, with extremely significant performance degradation, reflecting the core role of the CNN feature extraction module in accurately capturing the local structure of the signal and avoiding structural distortion. The No-Attention model has an average SSIM of 0.86, a decrease of 9.47%, indicating that the attention mechanism module can further improve the accuracy of signal structure preservation by focusing on key structural regions.

[0060] Under all noise levels, the performance gap between the various ablation variants and the DRSN-LSTM-CNN model persists, and is even more pronounced in low noise levels (0dB, 1dB, 2dB). In the 0dB high noise scenario, our model achieves an SSIM of 0.90, while the No-DRSN and No-CNN models are as low as -0.17 and 0.10, respectively, indicating severe signal structure corruption. Although the No-LSTM and No-Attention models maintain positive SSIM values, they are 0.76 and 0.80, respectively, which are lower than the DRSN-LSTM-CNN model.

[0061] (4) Residual spectrum flatness; The residual spectral flatness test results of each ablation variant model and the DRSN-LSTM-CNN model are shown in Table 9. Residual spectral flatness is used to evaluate the stationarity of the residual signal after denoising. The closer the value is to 1, the closer the residual signal is to white noise and the more thorough the noise removal. The average residual spectral flatness of this model reaches 0.93, which is significantly higher than all ablation variant models, confirming the key significance of the synergistic effect of each core module in improving the stationarity of the residual signal.

[0062] Table 9

[0063] In terms of specific performance degradation, the average residual spectral flatness of the No-DRSN model is only 0.49, a decrease of 47.31% compared to the DRSN-LSTM-CNN model, indicating that the DRSN module plays a core role in promoting thorough noise separation and optimizing residual signal characteristics; the average residual spectral flatness of the No-LSTM model is 0.43, a decrease of 53.76%, verifying the positive contribution of the LSTM module to the stabilization of temporal residual signals; the average residual spectral flatness of the No-CNN model is 0.89, a decrease of 4.30%. Although the degradation is relatively small, it still reflects the auxiliary value of the CNN feature extraction module in refining noise removal and improving the residual spectral distribution; the average residual spectral flatness of the No-Attention model is 0.78, a decrease of 16.13%, indicating that the attention mechanism module can improve the residual spectral flatness by accurately focusing on the noise region and reducing the residual of effective signal.

[0064] Under full noise levels, the performance gap between the various ablation variant models and the model itself persists: In a 0dB high noise scenario, the residual spectral flatness of the DRSN-LSTM-CNN model is 0.94, while the No-DRSN and No-LSTM models are as low as 0.50 and 0.44, respectively, indicating extremely uneven residual signal spectral distribution and significant noise residue; although the No-CNN and No-Attention models maintain relatively high flatness, their values ​​of 0.89 and 0.78, respectively, are still lower than the model itself. As the noise intensity decreases, the performance shortcomings of the various ablation variant models are not improved.

[0065] (5) Summary of ablation experiments; Ablation experiments show that the DRSN module, LSTM module, CNN feature extraction module, and attention mechanism module of the DRSN-LSTM-CNN model are all indispensable and have a significant synergistic effect. Removing the DRSN module reduces the signal-to-noise ratio by 66.61% and the SSIM drops to -0.07, proving that it is the core for achieving noise separation and ensuring signal structural integrity. Removing the LSTM module reduces the residual spectral flatness by 53.76%, confirming its key role in capturing long-term time dependencies and optimizing the stationarity of residual signals. Removing the CNN module results in an SSIM of only 0.10 and a Pearson correlation coefficient as low as -0.93, highlighting its important value in refining local features and reducing signal distortion. Removing the attention mechanism (No-Attention) reduces the signal-to-noise ratio by 52.13% and significantly decreases the structure preservation accuracy, indicating that it can accurately focus on key signal regions, improving denoising targeting and signal fidelity. The ablation experiments described above show that each core module has functions such as noise suppression, timing modeling, signal fidelity, and residual optimization. This model achieves optimal performance enhancement through deep collaboration among the modules, further demonstrating the rationality and scientific nature of the technical solution design of this invention.

[0066] 4. Summary of model performance; The comparative experiments above demonstrate that this model outperforms other models in a 0-20dB noisy environment, achieving an average SNR improvement of 5.39dB and an SSIM of 0.95. It exhibits significant advantages in low-noise scenarios with no signal distortion, while maintaining a parameter count of only 1367, balancing accuracy and engineering practicality. The ablation experiments confirm that DRSN, LSTM, CNN, and the attention module are all indispensable, working synergistically to achieve noise suppression, temporal modeling, and signal fidelity. Removing any of these modules leads to a significant performance degradation, highlighting the scientific rigor of the design.

[0067] In summary, this model, through the innovative integration of an overlapping segmentation mechanism and a hybrid deep learning architecture, specifically addresses the core pain points of ADCP signals, such as non-stationarity and low signal-to-noise ratio. It breaks through the limitations of traditional algorithms and general models, possessing both creativity and uniqueness, and can efficiently adapt to the needs of ADCP signal processing.

[0068] Example To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with... Figure 1 and Figure 2 The invention will be further described in detail with reference to the embodiments.

[0069] (1) Signal input and overlapping segmentation; Acquire a one-dimensional noisy signal sequence from an Acoustic Doppler Current Profiler (ADCP). Set the segment length. If the overlap ratio is set to 20%, then the sliding step size is... The input long sequence signal is segmented using a sliding window with a step size of S and a sliding length of L, generating multiple overlapping signal segments. For portions of the sequence that are less than the length of a complete segment, a reverse alignment method is used. When the end position of the sliding window exceeds the total length of the sequence, a strategy of first fixing the end position and then deriving the start position in reverse is employed to ensure that the length of the last segment is strictly equal to the preset segment length. (2) DRSN feature extraction and dimensionality reduction; The shape of each segment obtained in step (1) is The input signal is fed into a pre-trained Deep Residual Shrinking Network (DRSN) module. The forward propagation process of this module is based on the DRSN architecture, enabling adaptive extraction of high-level features from multi-channel input to single-channel signals. Specifically, the input signal first undergoes basic feature transformation through two consecutive convolutional-batch normalization layers. Subsequently, a channel attention sub-network generates soft attention weights for evaluating the importance of each channel through global average pooling and two convolutional layers with a bottleneck structure. These weights are multiplied channel-wise with the feature map to adaptively suppress noise-related or redundant feature responses, and then summed with the original input residuals, thereby achieving feature calibration while preserving key original information. The fused features are non-linearized by the ReLU activation function and finally aggregated and compressed into a single-channel output through a pointwise convolutional layer. The module outputs a pre-denoised signal segment with the shape [image of segment]. .

[0070] (3) Segmented signal reconstruction; All signal segments processed by DRSN are spliced ​​together according to their original time sequence. For overlapping regions, a weighted average method is used for fusion, that is, the weighted average of the points corresponding to all segments in the overlapping region is calculated according to the number of times they have been processed, thereby generating a preliminary denoised complete-length signal, effectively avoiding distortion at the segment boundaries.

[0071] (4) LSTM timing modeling; Step (3) will yield a complete preliminary denoised signal, the shape of which is Where N is the total signal length. This initially denoised signal is input into a two-layer bidirectional LSTM network. The hidden layer size of this LSTM is 4, used to capture long-term temporal dependencies in the signal. After processing by the LSTM network, it outputs a feature sequence containing rich contextual information, with the shape of... (Due to the bidirectional connection, it is twice the size of the hidden layer, which is 8).

[0072] (5) CNN feature refinement; The feature sequence output by the LSTM is dimensionally transformed to change its shape. To accommodate one-dimensional convolution operations, it is then fed into a CNN refinement network, the structure of which is as follows: 1) First layer: one-dimensional convolution, 32 input channels, 16 output channels, kernel size 5, padding 2, followed by ReLU and batch normalization.

[0073] 2) Second layer: one-dimensional convolution, 16 input channels, 1 output channel, kernel size 3, padding 1.

[0074] The CNN network is responsible for further refining features at the local scale and filtering out residual noise.

[0075] (6) Signal output; The signal after CNN processing (shape is) This is the final denoised signal, which is used as the output of this model.

[0076] The detailed training steps of the model in this embodiment are described below: (1) Dataset preparation; ADCP signals are generated using computer simulation. By mixing pure water flow velocity echo signals with additive white Gaussian noise of different intensities (0dB, 5dB, 10dB, 15dB, and 20dB respectively during the training process of this invention), large-scale noisy-pure signal pairs are constructed as training, validation, and test sets.

[0077] (2) Loss function; The hybrid loss function defined above is adopted, which directly optimizes the waveform recovery quality and constrains the waveform recovery in the frequency domain to avoid over-smoothing and optimization.

[0078] (3) Optimizer and parameters; The Adam optimizer was used, with an initial learning rate of 1e-4, a batch size of 32, and a total of 1000 training epochs. Training was stopped to avoid overfitting when the validation set loss did not decrease for several consecutive epochs.

[0079] (4) Training strategies; An end-to-end training approach is adopted, which directly uses the SI-SNR of the denoised signal and the target clean signal as the optimization objective, enabling the model to automatically learn the complex mapping from noisy signal to clean signal.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for denoising ADCP echo signals based on overlapping segmented DRSN-LSTM-CNN, characterized in that, include: Step 1: Receive the noisy ADCP signal and use the sliding window method to perform overlapping segmentation on the noisy ADCP signal; Step 2: Input each ADCP signal after overlapping segmentation into the DRSN module for feature extraction, feature learning, residual connection and dimensionality reduction in sequence; Step 3: Perform a weighted average of the ADCP signals of each segment output by the DRSN module to obtain the complete ADCP signal after preliminary denoising; Step 4: Input the pre-denoised complete ADCP signal into a bidirectional LSTM network to obtain the time-series feature sequence; Step 5: Input the temporal feature sequence into the CNN network, perform local refinement processing on the temporal feature sequence to obtain the refined feature representation, and generate the final denoised ADCP signal based on the refined feature representation.

2. The ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN according to claim 1, characterized in that, Step 1 specifically includes: Receive a noisy ADCP signal, where the noisy ADCP signal is a one-dimensional time series, with a fixed window length of L and an overlap rate of [value missing]. The step size for each slide is The sliding window method is used to overlap and segment the noisy ADCP signal, and output multiple ADCP signal segments.

3. The ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN according to claim 1, characterized in that, The DRSN module in step 2 includes a feature extraction submodule, an attention mechanism submodule, and a feature compression submodule connected in sequence; The feature extraction submodule includes two consecutive convolutional layers + batch normalization layers. The two consecutive convolutional layers + batch normalization layers are equipped with ReLU activation function layers. The feature extraction submodule is used to extract the features of each ADCP signal segment after overlapping segmentation to obtain a feature map. The attention mechanism submodule includes an adaptive average pooling layer, a convolutional layer, a ReLU activation function layer, a convolutional layer, and a sigmoid activation function layer connected in sequence. The attention mechanism submodule is used to adaptively learn the weights of different positions in the extracted feature map, suppress noise-related features, and after completing feature learning, add the weighted features to the original input features of the ADCP signal segment through residual connection. The feature compression submodule includes a ReLU activation function layer and a convolutional layer. The feature compression submodule is used to reduce the dimensionality of the segmented features of the ADCP signal after residual connection and compress the number of output channels of each segment to 1.

4. The ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN according to claim 1, characterized in that, Step 3 specifically includes: Determine the overlapping region of the ADCP signals by analyzing the time positions of each segment of the ADCP signal output by the DRSN module. Based on the signal values ​​corresponding to each segment within the overlapping region and the number of times they have been processed, a weighted average is calculated to obtain the complete ADCP signal after preliminary denoising.

5. The ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN according to claim 1, characterized in that, The bidirectional LSTM network and the CNN network are connected sequentially; The complete ADCP signal after initial denoising is input into a bidirectional LSTM network. The hidden layer size of the bidirectional LSTM network is set according to the length of the input ADCP signal. The bidirectional LSTM network captures the long-term time dependence of the complete ADCP signal after initial denoising. The dynamic pattern of the ADCP signal in the time dimension is learned by using a gating mechanism to correct the temporal inconsistency in the initial denoising result and obtain the temporal feature sequence.

6. The ADCP echo signal denoising method based on overlapping segmented DRSN-LSTM-CNN according to claim 1, characterized in that, The CNN network in step 5 includes a first convolutional layer, a normalization layer, a ReLU activation function layer, and a second convolutional layer connected in sequence, wherein the first convolutional layer and the second convolutional layer are both one-dimensional convolutional layers; By performing dimensionality transformation on the temporal feature sequence, convolutional input data can be obtained; The first convolutional layer extracts features from the convolutional input data. A normalization layer accelerates the training of the CNN network. A ReLU activation function layer is used to introduce non-linear expressive power and alleviate the gradient vanishing problem. Finally, a second convolutional layer is used for feature mapping and dimensionality reduction to obtain a refined feature representation of the single-channel ADCP signal. Based on the refined feature representation, the final denoised ADCP signal is generated.

Images