An AMT signal denoising method and device based on adaptive multi-level U-Net

Abstract

This application relates to an AMT signal denoising and device based on adaptive multi-level U-Net. This method classifies noisy segments of noisy AMT signals and records the corresponding position masks of clean AMT signals, providing noise localization information for the training of the denoising model to improve the training effect of the denoising model and avoid the loss of non-noisy segments. This method also uses two U-Net networks to form a U-Net cascade structure. The first network performs preliminary denoising, and the second network processes the segment data output by the first network. It uses the residual to generate noise intensity probability and uses the noise intensity probability as a driving signal to assign corresponding weight values, so as to achieve dynamic control of the selective absorption and fusion of the segment data output by the second network to the first network. It can adaptively focus on regions with significant residual noise, achieve feature-level complementarity and optimization, and ultimately improve the denoising effect of the denoising model on noisy AMT signals.

Description

Technical Field

[0001] This application relates to the field of AMT signal denoising technology, and in particular to an AMT signal denoising method and apparatus based on adaptive multi-level U-Net. Background Technology

[0002] Audio magnetotelluric (AMT), as an important geophysical exploration method, plays an irreplaceable role in mineral resource exploration, geothermal surveys, groundwater detection, and geological structure research due to its advantages such as moderate detection depth, high resolution, and sensitivity to low resistivity layers. However, AMT signals are highly susceptible to interference from various types of human noise. This noise often exhibits characteristics such as non-stationarity, non-Gaussianity, strong abruptness, and significant overlap with the effective signal band, leading to a sharp decrease in the signal-to-noise ratio of measured data, severely distorting apparent resistivity and phase curves, and consequently causing deviations or even errors in geological interpretation. Therefore, developing efficient and robust AMT signal denoising techniques is a crucial prerequisite for improving data quality and ensuring interpretation accuracy, possessing significant theoretical value and widespread engineering application needs.

[0003] In recent years, deep learning (DL) has achieved revolutionary breakthroughs in denoising tasks across multiple fields, including images, speech, and biological signals, thanks to its powerful feature learning capabilities, nonlinear modeling advantages, and end-to-end optimization potential. This wave has also rapidly swept through the field of geophysical signal processing, opening up a completely new path for analog-to-mass denoising (AMT). However, existing AMT denoising methods based on deep learning (DL) still have limitations:

[0004] In AMT time series, the localization of noise is ambiguous, and the output or regression output does not clearly or accurately define the noise boundary, resulting in incomplete denoising or excessive signal smoothing; single-level networks have limited ability to process complex noise patterns (such as strong and weak noise mixed, abrupt noise and gradual noise coexisting). Summary of the Invention

[0005] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0006] The main objective of this disclosure is to propose an AMT signal denoising method and apparatus based on adaptive multi-level U-Net, which can improve the denoising effect of the denoising model on noisy AMT signals.

[0007] A first aspect of this application provides an AMT signal denoising method based on an adaptive multi-level U-Net, the method comprising:

[0008] In response to the training instructions of the denoising model, the noisy AMT signal is classified to determine the noise segments classified in the noisy AMT signal;

[0009] Select a clean AMT signal corresponding to the noisy AMT signal, and mark a first segment in the clean AMT signal; wherein the position of the first segment in the clean AMT signal corresponds to the position of the noise segment in the noisy AMT signal;

[0010] The denoising model is trained based on the data corresponding to the noise segment of the noisy AMT signal and the data corresponding to the first segment of the clean AMT signal; the denoising model includes a first U-Net network and a second U-Net network; wherein, the process of training the denoising model includes:

[0011] The first U-Net network is used to perform a first denoising process on the data corresponding to the noise segment to obtain the first denoised data.

[0012] The first denoised data is denoised a second time using the second U-Net network to obtain the second denoised data;

[0013] Calculate the residual between the data corresponding to the noise segment and the first denoised data, and determine the noise intensity probability based on the residual;

[0014] Based on the noise intensity probability, assign corresponding weights to the first denoised data and the second denoised data, and generate third denoised data based on the first denoised data, the second denoised data and their corresponding weights;

[0015] The data corresponding to the first segment is used as the regression target of the data corresponding to the noise segment, and the denoising model is backpropagated to finally obtain the trained denoising model.

[0016] In response to the denoising instruction of the target AMT signal, the target AMT signal is input into the trained denoising model to obtain the denoised target AMT signal.

[0017] The AMT signal denoising method based on adaptive multi-level U-Net provided in this embodiment has at least the following beneficial effects:

[0018] This method performs binary classification on noisy AMT signals (distinguishing between noisy and non-noisy segments) and records the corresponding positional label mask of the clean AMT signal sample. This provides precise noise mask guidance for the subsequent training process of the denoising model, offering accurate noise localization information to improve the accuracy of the subsequent denoising model training. Simultaneously, it avoids affecting the non-noisy segments of the noisy AMT signal. The method uses two U-Net networks in a cascaded structure. The first U-Net network receives the noisy segments classified as containing noise from the noisy AMT signal and performs initial noise suppression. The second U-Net network processes the segment data output by the first U-Net network, using the residual estimated after denoising by the first U-Net network to generate noise intensity probabilities. These probabilities are used as driving signals, and corresponding weight values ​​are assigned to dynamically control the selective absorption and fusion of the segment data output by the second U-Net network. This adaptively focuses on regions with significant residual noise, achieving precise feature complementarity and optimization, ultimately improving the denoising effect of the denoising model on AMT signals.

[0019] A second aspect of this application provides an AMT signal denoising device based on an adaptive multi-level U-Net, the device comprising:

[0020] The signal classification module is used to classify the noisy AMT signal in response to the training instructions of the denoising model, so as to determine the noise segments classified in the noisy AMT signal.

[0021] The regression target acquisition module is used to select a clean AMT signal corresponding to the noisy AMT signal and mark a first segment in the clean AMT signal; wherein the position of the first segment in the clean AMT signal corresponds to the position of the noise segment in the noisy AMT signal;

[0022] A denoising model training module is used to train the denoising model based on data corresponding to the noise segment of the noisy AMT signal and data corresponding to the first segment of the clean AMT signal; the denoising model includes a first U-Net network and a second U-Net network; wherein, the process of training the denoising model includes:

[0023] The first U-Net network is used to perform a first denoising process on the data corresponding to the noise segment to obtain the first denoised data.

[0024] The first denoised data is denoised a second time using the second U-Net network to obtain the second denoised data;

[0025] Calculate the residual between the data corresponding to the noise segment and the first denoised data, and determine the noise intensity probability based on the residual;

[0026] Based on the noise intensity probability, assign corresponding weights to the first denoised data and the second denoised data, and generate third denoised data based on the first denoised data, the second denoised data and their corresponding weights;

[0027] The data corresponding to the first segment is used as the regression target of the data corresponding to the noise segment, and the denoising model is backpropagated to finally obtain the trained denoising model.

[0028] The target signal denoising module is used to respond to the denoising command of the target AMT signal, and input the target AMT signal into the trained denoising model to obtain the denoised target AMT signal.

[0029] A third aspect of this application provides an electronic device including at least one controller and a memory for communicatively connecting to the controller; the memory stores instructions executable by the at least one controller, the instructions being executed by the at least one controller to cause the at least one controller to perform an AMT signal denoising method based on an adaptive multi-level U-Net as described above.

[0030] A fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform an AMT signal denoising method based on an adaptive multi-level U-Net as described above.

[0031] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating an AMT signal denoising method based on an adaptive multi-level U-Net according to an embodiment of this application;

[0034] Figure 2This is a flowchart illustrating an AMT signal denoising method based on an adaptive multi-level U-Net, provided in another embodiment of this application.

[0035] Figure 3 This is a schematic diagram of the structure of an AMT signal denoising device based on an adaptive multi-level U-Net according to an embodiment of this application;

[0036] Figure 4 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0038] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0039] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed or function in a specific orientation, and therefore should not be construed as a limitation of this application.

[0040] One embodiment of this application provides an AMT signal denoising method based on adaptive multi-level U-Net, the method comprising:

[0041] Step S100: In response to the training instructions of the denoising model, the noisy AMT signal is classified to determine the noise segments classified in the noisy AMT signal.

[0042] Step S200: Select the clean AMT signal corresponding to the noisy AMT signal and mark the first segment in the clean AMT signal; wherein the position of the first segment in the clean AMT signal corresponds to the position of the noise segment in the noisy AMT signal;

[0043] Step S300: Train a denoising model based on the data corresponding to the noisy segments of the noisy AMT signal and the data corresponding to the first segment of the clean AMT signal; the denoising model includes a first U-Net network and a second U-Net network; wherein, the process of training the denoising model includes:

[0044] The first denoised data is obtained by performing the first denoised data on the data corresponding to the noise segment using the first U-Net network.

[0045] The first denoised data is then denoised a second time using the second U-Net network to obtain the second denoised data;

[0046] Calculate the residual between the data corresponding to the noise segment and the first denoised data, and determine the noise intensity probability based on the residual;

[0047] Based on the noise intensity probability, assign corresponding weights to the first and second denoised data, and generate the third denoised data based on the first denoised data, the second denoised data and their corresponding weights.

[0048] The data corresponding to the first segment is used as the regression target for the data corresponding to the noise segment, and the denoising model is backpropagated to finally obtain the trained denoising model.

[0049] Step S400: In response to the denoising instruction of the target AMT signal, the target AMT signal is input into the trained denoising model to obtain the denoised target AMT signal.

[0050] In this embodiment, steps S100 and S200 are data processing steps before training the denoising model, and their contents are described below:

[0051] Noisy AMT signals can be generated under experimental conditions (with noise) and the noise data is labeled, or they can be measured AMT signals. For example, by empirically analyzing the approximate difference between completely noiseless AMT signals and simulated noisy AMT signals, a threshold is set for AMT signals. After slicing the AMT signal data, if the slice contains data greater than the threshold, the slice is considered a noise sample; otherwise, it is a non-noise sample. In this way, training samples are obtained, and classification training of the classification model is performed. After reasonable parameter tuning, the optimal classification model is output.

[0052] In step S100, a classification model can be preset, and then the noisy AMT signal can be classified according to the model to determine the noisy segments (i.e., classified as noise segments) and the noiseless segments (i.e., classified as non-noise segments). For example, in 10 segments, if segments 1, 3, and 5 are identified as noise segments containing noise, then the other segments are non-noise segments. It should be noted that if the noisy signal is generated based on the simulated AMT signal under experimental conditions, then the classified non-noise segments are completely noise-free. However, if it is a measured noisy AMT signal, then due to environmental limitations (such as the accuracy of the classification model), its non-noise segments are not completely noise-free.

[0053] In step S200, a clean AMT signal is first selected. This clean AMT signal serves as the regression target corresponding to the input data (i.e., the noisy segment of the noisy AMT signal) in the denoising model. The clean AMT signal corresponds to the noisy AMT signal. For example, if the sequence length of the noisy AMT signal is divided into 10 segments, then the sequence of the clean AMT signal can also be divided into 10 segments to find two corresponding segments. The clean AMT signal can be a completely noise-free simulated AMT signal generated under experimental conditions, or it can be a relatively clean AMT signal after denoising using a certain denoising method.

[0054] In this embodiment, step S300 is the denoising training process of the denoising model, which mainly includes:

[0055] First, the network structure of the denoising model is determined, which includes two U-Net networks. The U-Net structure consists of two main parts: an encoder and a decoder. The encoder gradually reduces the spatial resolution of the image through a series of convolutional and pooling layers, while increasing the number of feature channels, thereby extracting high-level features of the image. Each encoding stage can be represented as (1):

[0056] (1);

[0057] in, This indicates the encoder output data. This indicates a max pooling operation. Indicates the activation layer. This represents convolution.

[0058] The decoder gradually restores the spatial resolution of the image through upsampling and convolutional layers, while reducing the number of feature channels to generate an accurate segmentation map. Each decoding stage can be represented as (2):

[0059] (2);

[0060] in, Output data for the decoder. This indicates a splicing operation. Indicates upsampling, This indicates a skip connection to the data.

[0061] One of the key features of U-Net is the use of skip connections in the decoder, which connect feature maps in the encoder to corresponding upsampled feature maps in the decoder. These connections help the network utilize contextual information from the encoder when recovering image details.

[0062] Data is obtained after using the decoding module in U-Net. It directly maps the number of channels to the number of target categories through a 1×1 convolutional layer, and uses sigmoid (binary classification) or softmax (multi-class classification) on the output.

[0063] This embodiment uses two networks to form a U-Net cascade structure, building a framework of first-level coarse denoising and second-level fine denoising. The first U-Net network receives the noise segments classified as containing noise from the original noisy AMT signal and performs preliminary noise suppression. The second U-Net network focuses on processing the segment data output by the first U-Net network. Then, it uses the residual estimated after the first-level denoising to generate noise intensity probabilities. The noise intensity probabilities are used as driving signals and corresponding weight values ​​are assigned to achieve dynamic control of the selective absorption and fusion of the segment data output by the second U-Net network on the first U-Net network. This mechanism enables the two U-Net networks to adaptively focus on regions with significant residual noise, achieving precise complementarity and optimization at the feature level.

[0064] It's important to note that since the denoising model only denoises noisy segments, the resulting segments replace the original noisy AMT signal, resulting in a relatively clean AMT signal. For example, if segments 1, 3, and 5 are identified as noise in 10 segments, then the denoised segments 1, 3, and 5 output by the denoising model replace the original noise segments 1, 3, and 5. This achieves accurate identification and prevents damage to valid data caused by denoising segments not identified as noise in the noisy data.

[0065] It should also be noted that the denoising model is used for regression tasks, and the optimization of the model (i.e., the backpropagation process) is a common technique in the field, so the process will not be described in detail here.

[0066] In this embodiment, step S400 is the application stage of the model. In this step, the target AMT signal is input into the trained denoising model to obtain the denoised target AMT signal.

[0067] The AMT signal denoising method based on adaptive multi-level U-Net provided in this embodiment has the following beneficial effects:

[0068] (1) This method performs binary classification (noise segment and non-noise segment) on each noisy AMT signal and records the corresponding position label mask of the corresponding clean AMT signal sample, which provides accurate noise mask guidance for the subsequent training process of the denoising model and provides corresponding accurate noise localization information to improve the accuracy of the subsequent denoising model training.

[0069] (2) This method uses two U-Net networks to form a U-Net cascade structure. The first U-Net network receives the noise segments classified as containing noise in the noisy AMT signal and then performs preliminary noise suppression. The second U-Net network processes the segment data output by the first U-Net network and then uses the residual estimated by the first U-Net network after denoising to generate the noise intensity probability. The noise intensity probability is used as the driving signal and the corresponding weight value is assigned to achieve dynamic control of the selective absorption and fusion of the segment data output by the first U-Net network by the second U-Net network. It can adaptively focus on the region with significant residual noise, achieve precise complementarity and optimization at the feature level, and finally improve the denoising effect of the denoising model on the AMT signal.

[0070] In some embodiments of this application, step S300, which determines the noise intensity probability based on the residual, includes:

[0071] Step S311: Input the residual into the first convolutional layer to obtain the first convolutional feature output by the first convolutional layer;

[0072] Step S312: Input the first convolutional feature into the first activation layer based on the LeakyReLU function to obtain the first activation feature;

[0073] Step S313: Input the first activation feature into the second convolutional layer to obtain the second convolutional feature output by the second convolutional layer;

[0074] Step S314: Input the second convolutional features into the second activation layer based on the sigmoid function to obtain the noise intensity probability.

[0075] This embodiment provides a gated network, including a first convolutional layer, a first activation layer with a LeakyReLU function, a second convolutional layer, and a second activation layer with a sigmoid function. It can generate corresponding noise intensity probabilities from the residual estimated after denoising using the first U-Net network, and then assign different weights to different denoised segments based on the noise intensity probabilities to achieve the purpose of adaptive noise intensity matching for secondary denoising.

[0076] In some embodiments of this application, step S300 assigns corresponding weights to the first denoised data and the second denoised data according to the noise intensity probability, and generates third denoised data based on the first denoised data, the second denoised data and their corresponding weights, including:

[0077] Step S321: The difference between 1 and the noise intensity probability is used as the first weight of the first denoised data;

[0078] Step S322: Use the noise intensity probability as the second weight of the second denoised data;

[0079] Step S323: The third denoised data is obtained by weighted summation of the first denoised data and the first weight, the second denoised data and the second weight.

[0080] In this embodiment, for noisy AMT signals, if the data point has a high noise intensity, then the intensity coefficient is close to 1, and the final result "third denoised data" is biased towards the second denoised data. If it is close to 0, then it is biased towards the first denoised data, indicating low noise intensity. This method uses a gated adaptive cascaded two-level U-Net structure, which can achieve the purpose of adaptive noise intensity matching for secondary denoising.

[0081] In some embodiments of this application, before classifying the noisy AMT signal in step S100 to determine the noise segments in the noisy AMT signal that are classified as noise, the method further includes:

[0082] Step S101: Construct a classification network;

[0083] Step S102: Generate AMT signal;

[0084] Step S103: Slice the AMT signal and identify the noisy segments containing noise and the non-noisy segments without noise to obtain labeled training samples.

[0085] Step S104: The classification network is trained based on the training samples to obtain a trained classification model.

[0086] Step S105: Use the training samples participating in the classification model training as noisy AMT signals;

[0087] Step S100, classifying the noisy AMT signal to determine the noise segments classified within the noisy AMT signal, includes:

[0088] Step S110: Input the noisy AMT signal into the classification model to classify the noise segments in the noisy AMT signal according to the classification model.

[0089] The classification network is a third U-Net network; wherein, the third U-Net network is a one-dimensional U-Net network, and the third U-Net network outputs binary classification.

[0090] The U-Net structure consists of two main parts: an encoder and a decoder. This embodiment uses U-Net to classify noisy segments. While preserving skip connections, the feature maps in the encoder are connected to the corresponding upsampled feature maps in the decoder, which can improve the classification accuracy.

[0091] The specific implementation process includes: modifying the encoding-decoding stage to a one-dimensional input while keeping the structure unchanged, and outputting the pixel-level output head. The input data is passed through a 1x1 convolution layer and then through a Sigmoid or other separation probability output function. For example, the input data passes through a convolutional layer to compress the channels, then through non-linear activation, and the value of each channel is calculated through adaptive average pooling. The spatial scale is then compressed through Flatten, and the classification is output through a fully connected classification function.

[0092] In some embodiments of this application, step S200, selecting a clean AMT signal corresponding to the noisy AMT signal, includes:

[0093] Step S210: Acquire the actual AMT signal that corresponds to the noisy AMT signal in the sequence;

[0094] Step S220: Perform data-driven tight-frame DDTF denoising on the actual AMT signal to generate a pseudo-clean signal of the actual AMT signal, and use the pseudo-clean signal as the clean AMT signal corresponding to the noisy AMT signal.

[0095] Step S230: Generate a completely noise-free analog AMT signal, and use both the analog AMT signal and the pseudo-clean signal as the clean AMT signal corresponding to the noisy AMT signal.

[0096] Step S300 trains a denoising model based on the data corresponding to the noisy segments of the noisy AMT signal and the data corresponding to the first segment of the clean AMT signal, including:

[0097] Step S331: Take the data corresponding to the first segment of the simulated AMT signal and the data corresponding to the noise segment as a first segment data pair, and take the data corresponding to the first segment of the pseudo-clean signal and the data corresponding to the noise segment as a second segment data pair; wherein, the data corresponding to the first segment in the first segment data pair is used as the regression target of the data corresponding to the noise segment, and the data corresponding to the first segment in the second segment data pair is used as the regression target of the data corresponding to the noise segment.

[0098] Step S332: Train a denoising model based on the first data segment pair and the second data segment pair.

[0099] When denoising signals, noise-containing segments are usually denoised while relatively clean, noise-free segments are left unprocessed. However, this presents a problem: if noise is not thoroughly identified, the accuracy of denoising will be greatly affected. The goal here is to identify a segment of a signal that is suspected of being noise as noise, but not to identify the entire segment of a noisy signal as noise.

[0100] Therefore, this embodiment adopts a labeling strategy: for simulated AMT signals (i.e., with corresponding clean data as a reference), the simulated AMT signals are labeled: noisy AMT signals are classified and identified as noise segments, and segments at the same position in clean AMT signals are labeled; for actual AMT signals, since there is no corresponding clean reference data, the DDTF denoising method is used to denoise the actual AMT signals to obtain a relatively clean signal after denoising (i.e., a pseudo-clean signal), which conforms to the statistical law of the clean part of noisy signals. DDTF is Data-Driven Tight Frames, and its principle is: to establish an optimization model to find a dictionary that can provide the sparsest representation of the training data and satisfy the unit tight frame constraint. By alternately optimizing the sparse coefficients and the dictionary itself, and using singular value decomposition to efficiently solve the constrained dictionary update problem, a tight frame system with powerful representation capabilities learned from the data is finally obtained; details are not elaborated here.

[0101] like Figure 2 As shown in one embodiment of this application, an AMT signal denoising method based on adaptive multi-level U-Net is provided, the method comprising:

[0102] Step S910: Select the first U-Net network and the second U-Net network to build a denoising model, and select the third U-Net network to build a classification model.

[0103] like Figure 2 As shown, the process is divided into Step 1 and Step 2. Step 1 is the process of training the classification model, identifying noise in the classification model, and determining the regression target data.

[0104] Step S920: Select the dataset and generate a clean AMT signal.

[0105] Obtain the AMT signal (which can be a noisy AMT signal generated under experimental conditions or a measured AMT signal). Then, for example, set a threshold. After slicing the AMT signal, if the slice contains data greater than the threshold, the AMT signal is considered a noisy AMT signal and used as a training sample. Noisy segments are marked as 0, and otherwise marked as 1. This is how the training samples are obtained.

[0106] A portion of the training samples is used to train a third U-Net network to obtain a classification model. The other portion serves as input to the classification model for classification, and after classification, it participates in the process of generating regression target data, as well as in the training process of the subsequent denoising model.

[0107] A clean AMT signal can come from an analog AMT signal or an actual AMT signal. The actual AMT signal is denoised using the DDTF denoising method to obtain a relatively clean pseudo-clean signal after denoising. This pseudo-clean signal is also used as a clean AMT signal.

[0108] Step S930: Train a third U-Net network based on a portion of the training samples to obtain a classification model.

[0109] A portion of the training samples is divided into a training set and a validation set. The training set and validation set are then input into a third U-Net network for training and testing, while retaining the weight parameters. After training is complete, a classification model is obtained.

[0110] Step S940: Input the training samples that did not participate in the classification model training into the classification model to obtain the noise segments and non-noise segments in the classified training samples.

[0111] Step S950: Based on the noise and non-noise segments in the classified training samples, the clean AMT signal is processed to obtain the target data for the regression task.

[0112] When denoising noisy AMT signals, the approach is to denoise only the noisy segments and leave the relatively clean segments unprocessed. However, this presents a problem: incomplete noise identification significantly impacts denoising accuracy. The goal is to identify potentially noisy segments as noise, but not the entire noisy signal segment. Therefore, this embodiment employs a labeling strategy:

[0113] Labeling the clean AMT signal: For noisy segments in the noisy AMT signal that the classification model identifies as noise, the corresponding segments in the clean AMT signal are labeled. This prepares the corresponding target data for subsequent denoising by the first and second U-Net networks. For example, if segments 1, 3, and 5 in the 10 sequences in the training samples are noisy segments, then segments 1, 3, and 5 in the clean AMT signal are labeled, and the data corresponding to segments 1, 3, and 5 in the clean AMT signal are used as the target data for the regression task of segments 1, 3, and 5 in the training samples.

[0114] Step S960: Train the denoising model.

[0115] like Figure 2 Step 2, input data for training the denoising model: data corresponding to the noise segments of the training samples + target data, that is, data corresponding to the noise segments and the first segment of the simulated AMT signal, as well as data corresponding to the noise segments and the first segment of the pseudo-clean AMT signal.

[0116] The goal of training a denoising model is to fit and regress the noise fragments onto the target data.

[0117] This embodiment provides a gated adaptive cascaded two-level U-Net structure for denoising. The input data is first subjected to coarse denoising by the first U-Net network to obtain the data. For example (3):

[0118] (3);

[0119] Then calculate the residual, which is the data corresponding to the noise segment. Subtract data Obtain the residual , formula (4):

[0120] (4);

[0121] This method employs an innovative gating mechanism: residual noise absolute value-driven gating. This gating uses the absolute value of the residual noise estimated after denoising by the first U-Net network as the driving signal to dynamically control the selective absorption and fusion of features from the first U-Net network by the second U-Net network. The specific operation is as follows:

[0122] residual The data is fed into a convolutional layer, an activation layer, another convolutional layer, and then the convolutional data is fed into a sigmoid function to output the noise intensity probability. The formula is as follows (5):

[0123] (5);

[0124] The second U-Net network input is noise. The output is noise. For example (6):

[0125] (6);

[0126] Based on the calculated noise intensity probability, for noisy AMT signals, if the data points with high noise intensity have an intensity coefficient close to 1, the final result will be noise. Preferring noise If the output is close to 0, it is biased towards noise. This indicates that the noise intensity is low, and only one denoising is needed. Apart from these two extreme cases, other data are calculated using the same strategy. This achieves the purpose of adaptive noise intensity matching for secondary denoising, as shown in formula (7):

[0127] (7);

[0128] Step S970: Use a denoising model to denoise the target AMT signal.

[0129] In summary, the overall process includes: first, using a threshold to coarsely divide the noisy AMT signal into noisy segments and non-noisy segments; then, dividing the signal into training and validation sets and inputting them into a third U-Net network for training, while retaining the weight parameters. Secondly, noisy AMT signals not trained in the third U-Net network are predicted. These are processed using the same slice size as during training and input into the trained classification model for classification. Noisy segments identified as noise and non-noise segments identified as non-noise are recorded. Then, for the target data used for regression, a simulated (noise-free) AMT signal is used as a reference. The positions where the noisy AMT signal is classified as noise or non-noise segments are used as a reference. Corresponding positions in the simulated AMT signal are then marked as noise or non-noise segments. Since there is no clean reference for the actual AMT signal, the actual AMT signal denoised using DDTF is chosen as a "pseudo-clean signal," and the same labeling is applied at the corresponding positions. This process serves as the target data for subsequent denoising. Finally, the data corresponding to the noise segments classified from the noisy AMT signal and the target data (i.e., the data in the labeled corresponding segments of the simulated AMT signal and the labeled corresponding segments of the pseudo-clean signal) are imported into a gated adaptive noise intensity cascaded U-Net denoising model for training, and the optimal denoising weight parameters are saved. When using the model, the noisy target AMT signal is input to obtain the final denoising result.

[0130] The advantage of this embodiment is that:

[0131] A one-dimensional variant of the original two-dimensional U-Net network is proposed to adapt to AMT signals, forming a noise identifier for noisy AMT signals. It retains the advantages of the U-shaped encoder-decoder structure and skip connections. Its goal is to perform binary classification (noise / non-noise) on the input noisy AMT signals and record the label mask at the corresponding position of the clean AMT signal, providing accurate noise mask guidance for the regression target in the subsequent denoising stage.

[0132] This paper presents a noise extraction method based on empirical knowledge and DDTF. During separation and identification, the method empirically analyzes the approximate difference between the clean signal and the simulated noisy signal. For the AMT signal, a threshold is set. After slicing the data, if the slice contains data exceeding the threshold, the sample is considered a noise sample. This yields the input data training samples for classification training. After reasonable parameter tuning, the optimal classification model is output. Furthermore, the target data for regression includes simulated AMT signals and pseudo-clean signals based on actual AMT signals. For simulated AMT signals, the simulated AMT signals are labeled: that is, noisy segments identified as noise in the noisy AMT signals are labeled, and segments at the same position in the simulated AMT signals are also labeled. For actual AMT signals, since there is no corresponding clean reference data, the DDTF denoising method is used to denoise the actual AMT signals to obtain pseudo-clean signals. The pseudo-clean signals conform to the statistical laws of the clean part of the noisy AMT signals and can also be used as the target data for subsequent U-Net denoising.

[0133] A cascaded two-stage U-Net structure is provided, which is then combined with a specific gating method. This gating method uses the absolute value of the residual noise estimated after the first-stage denoising as the driving signal to dynamically control the selective absorption and fusion of features from the first-stage U-Net network in the second stage. This mechanism enables the second-stage network to adaptively focus on regions with significant residual noise, achieving precise complementarity and optimization at the feature level.

[0134] like Figure 3 One embodiment of this application provides an AMT signal denoising device based on an adaptive multi-level U-Net, the device comprising:

[0135] The signal classification module 1001 is used to respond to the training instructions of the denoising model and classify the noisy AMT signal to determine the noise segments classified in the noisy AMT signal.

[0136] The regression target acquisition module 1002 is used to select a clean AMT signal corresponding to the noisy AMT signal and mark a first segment in the clean AMT signal; wherein the position of the first segment in the clean AMT signal corresponds to the position of the noise segment in the noisy AMT signal.

[0137] The denoising model training module 1003 is used to train a denoising model based on data corresponding to noise segments of the noisy AMT signal and data corresponding to the first segment of the clean AMT signal; the denoising model includes a first U-Net network and a second U-Net network; wherein, the process of training the denoising model includes:

[0138] The first denoised data is obtained by performing the first denoised data on the data corresponding to the noise segment using the first U-Net network.

[0139] The first denoised data is then denoised a second time using the second U-Net network to obtain the second denoised data;

[0140] Calculate the residual between the data corresponding to the noise segment and the first denoised data, and determine the noise intensity probability based on the residual;

[0141] Based on the noise intensity probability, assign corresponding weights to the first and second denoised data, and generate the third denoised data based on the first denoised data, the second denoised data and their corresponding weights.

[0142] The data corresponding to the first segment is used as the regression target for the data corresponding to the noise segment, and the denoising model is backpropagated to finally obtain the trained denoising model.

[0143] The target signal denoising module 1004 is used to respond to the denoising command of the target AMT signal, input the target AMT signal into the trained denoising model, and obtain the denoised target AMT signal.

[0144] It should be noted that the AMT signal denoising device based on adaptive multi-level U-Net provided in this embodiment is based on the same inventive concept as the AMT signal denoising method based on adaptive multi-level U-Net described above. Therefore, the content of the AMT signal denoising method based on adaptive multi-level U-Net described above is also applicable to the content of the AMT signal denoising device based on adaptive multi-level U-Net in this embodiment, and will not be repeated here.

[0145] like Figure 4 One embodiment of this application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described AMT signal denoising method based on adaptive multi-level U-Net. The electronic device includes:

[0146] At least one battery;

[0147] At least one memory;

[0148] At least one processor;

[0149] At least one program;

[0150] The program is stored in memory, and the processor executes at least one program to implement the above-described AMT signal denoising method based on adaptive multi-level U-Net as described in this disclosure.

[0151] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0152] The electronic devices according to embodiments of this application will now be described in detail.

[0153] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure.

[0154] The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to implement an AMT signal denoising method based on an adaptive multi-level U-Net according to an embodiment of this disclosure.

[0155] The input / output interface 1800 is used to implement information input and output.

[0156] The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0157] Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900);

[0158] The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0159] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the above-described AMT signal denoising method based on adaptive multi-level U-Net.

[0160] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0161] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0162] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0163] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0164] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0165] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0166] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0167] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0168] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0169] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0170] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0171] The above is a detailed description of the preferred embodiments of this application. However, the embodiments of this application are not limited to the above-described implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the embodiments of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of the embodiments of this application.

Claims

1. A method for AMT signal denoising based on adaptive multi-level U-Net, characterized in that, The method includes: In response to the training instructions of the denoising model, the noisy AMT signal is classified to determine the noise segments classified in the noisy AMT signal; Select a clean AMT signal corresponding to the noisy AMT signal, and mark a first segment in the clean AMT signal; wherein the position of the first segment in the clean AMT signal corresponds to the position of the noise segment in the noisy AMT signal; The denoising model is trained based on the data corresponding to the noise segment of the noisy AMT signal and the data corresponding to the first segment of the clean AMT signal; the denoising model includes a first U-Net network and a second U-Net network; wherein, the process of training the denoising model includes: The first U-Net network is used to perform a first denoising process on the data corresponding to the noise segment to obtain the first denoised data. The first denoised data is denoised a second time using the second U-Net network to obtain the second denoised data; Calculate the residual between the data corresponding to the noise segment and the first denoised data, and determine the noise intensity probability based on the residual; Based on the noise intensity probability, assign corresponding weights to the first denoised data and the second denoised data, and generate third denoised data based on the first denoised data, the second denoised data and their corresponding weights; The data corresponding to the first segment is used as the regression target of the data corresponding to the noise segment, and the denoising model is backpropagated to finally obtain the trained denoising model. In response to the denoising instruction of the target AMT signal, the target AMT signal is input into the trained denoising model to obtain the denoised target AMT signal.

2. The AMT signal denoising method based on adaptive multi-level U-Net according to claim 1, characterized in that, The step of determining the noise intensity probability based on the residual includes: The residual is input into the first convolutional layer to obtain the first convolutional feature output by the first convolutional layer. The first convolutional feature is input into the first activation layer based on the LeakyReLU function to obtain the first activation feature; The first activation feature is input into the second convolutional layer to obtain the second convolutional feature output by the second convolutional layer; The second convolutional feature is input into the second activation layer based on the sigmoid function to obtain the noise intensity probability.

3. The AMT signal denoising method based on adaptive multi-level U-Net according to claim 2, characterized in that, Based on the noise intensity probability, corresponding weights are assigned to the first denoised data and the second denoised data, and third denoised data is generated based on the first denoised data, the second denoised data, and their corresponding weights, including: The difference between 1 and the noise intensity probability is used as the first weight of the first denoised data; The noise intensity probability is used as the second weight of the second denoised data; The third denoised data is obtained by weighted summation of the first denoised data and the first weight, and the second denoised data and the second weight.

4. The AMT signal denoising method based on adaptive multi-level U-Net according to claim 1, characterized in that, Before classifying the noisy AMT signal to determine the noise segments classified as noise in the noisy AMT signal, the method further includes: Construct a classification network; Generate AMT signal; The AMT signal is sliced, and noisy segments containing noise and non-noisy segments without noise are identified to obtain labeled training samples; The classification network is trained based on the training samples to obtain a trained classification model; The training samples used in the classification model training are used as the noisy AMT signal; The classification of noisy AMT signals to determine the noise segments identified in the noisy AMT signals includes: The noisy AMT signal is input into the classification model to classify the noise segments in the noisy AMT signal according to the classification model.

5. The AMT signal denoising method based on adaptive multi-level U-Net according to claim 4, characterized in that, The classification network is a third U-Net network; wherein, the third U-Net network is a one-dimensional U-Net network, and the third U-Net network outputs binary classification.

6. The AMT signal denoising method based on adaptive multi-level U-Net according to claim 1, characterized in that, The selection of a clean AMT signal corresponding to the noisy AMT signal includes: Acquire the actual AMT signal that corresponds to the noisy AMT signal in sequence; The actual AMT signal is subjected to data-driven tight-frame DDTF denoising to generate a pseudo-clean signal of the actual AMT signal, and the pseudo-clean signal is used as the clean AMT signal corresponding to the noisy AMT signal.

7. The AMT signal denoising method based on adaptive multi-level U-Net according to claim 6, characterized in that, After performing data-driven tight-frame DDTF denoising on the actual AMT signal to generate a pseudo-clean signal of the actual AMT signal, and using the pseudo-clean signal as the clean AMT signal corresponding to the noisy AMT signal, the method further includes: Generate a completely noise-free analog AMT signal, and use both the analog AMT signal and the pseudo-clean signal as the clean AMT signal corresponding to the noisy AMT signal; Training the denoising model based on the noise segment of the noisy AMT signal and the first segment of the clean AMT signal includes: The data corresponding to the first segment of the simulated AMT signal and the data corresponding to the noise segment are used as a first segment data pair, and the data corresponding to the first segment of the pseudo-clean signal and the data corresponding to the noise segment are used as a second segment data pair; wherein, the data corresponding to the first segment in the first segment data pair is used as the regression target corresponding to the data corresponding to the noise segment, and the data corresponding to the first segment in the second segment data pair is used as the regression target corresponding to the data corresponding to the noise segment. The denoising model is trained based on the first data segment pair and the second data segment pair.

8. An AMT signal denoising device based on adaptive multi-level U-Net, characterized in that, The device includes: The signal classification module is used to classify the noisy AMT signal in response to the training instructions of the denoising model, so as to determine the noise segments classified in the noisy AMT signal. The regression target acquisition module is used to select a clean AMT signal corresponding to the noisy AMT signal and mark a first segment in the clean AMT signal; wherein the position of the first segment in the clean AMT signal corresponds to the position of the noise segment in the noisy AMT signal; A denoising model training module is used to train the denoising model based on data corresponding to the noise segment of the noisy AMT signal and data corresponding to the first segment of the clean AMT signal; the denoising model includes a first U-Net network and a second U-Net network; wherein, the process of training the denoising model includes: The first U-Net network is used to perform a first denoising process on the data corresponding to the noise segment to obtain the first denoised data. The first denoised data is denoised a second time using the second U-Net network to obtain the second denoised data; Calculate the residual between the data corresponding to the noise segment and the first denoised data, and determine the noise intensity probability based on the residual; Based on the noise intensity probability, assign corresponding weights to the first denoised data and the second denoised data, and generate third denoised data based on the first denoised data, the second denoised data and their corresponding weights; The data corresponding to the first segment is used as the regression target of the data corresponding to the noise segment, and the denoising model is backpropagated to finally obtain the trained denoising model. The target signal denoising module is used to respond to the denoising command of the target AMT signal, and input the target AMT signal into the trained denoising model to obtain the denoised target AMT signal.

9. An electronic device, characterized in that, It includes at least one controller and a memory for communicatively connecting with the controller; the memory stores instructions executable by the at least one controller, which, when executed by the at least one controller, causes the at least one controller to perform an AMT signal denoising method based on an adaptive multi-level U-Net as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions for causing a computer to perform an AMT signal denoising method based on an adaptive multi-level U-Net as described in any one of claims 1 to 7.

Images