Transient electromagnetic data fast denoising method based on tensor mapping and multi-task joint

Abstract

The application discloses a transient electromagnetic data fast denoising method based on tensor mapping and multi-task joint, relates to the technical field of geophysical exploration and intelligent signal processing, and comprises the following steps: collecting one-dimensional discrete time sequence signals, mapping the one-dimensional discrete time sequence signals into high-dimensional feature tensors through pseudo-two-dimensional tensorization folding and normalizing; constructing a denoising network containing a deep residual cascade module and a feedforward jump connection; training the network by using normalized tensors, constructing a joint loss function weighted by mean square error and mean absolute error, and combining dynamic learning rate annealing optimization; and in the inference stage, inputting data to be denoised into the network to predict background noise, subtracting the background noise, and obtaining pure signals through inverse normalization. The application solves the problems of traditional methods, such as easy erasing of late effective signals, large algorithmic cost and poor robustness to sferics pulses, and realizes high-fidelity and high-efficiency fast denoising.

Description

Technical Field

[0001] This invention relates to the fields of geophysical exploration and intelligent signal processing technology, specifically to a fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration. Background Technology

[0002] Transient electromagnetic methods, with their advantages of small volume effect and sensitivity to low-resistivity anomalies, are widely used in underground exploration fields such as deep mineral prospecting and coal mine water hazard investigation. However, in actual exploration, the secondary induced field signal decays sharply with time in a power function order. The late-stage signal reflecting deep geological information is extremely weak and easily submerged by complex environmental noise such as power frequency harmonics, atmospheric pulses, and Gaussian white noise, which severely limits the accuracy of deep exploration.

[0003] In related technologies, traditional filtering methods such as wavelet transform and empirical mode decomposition are not only cumbersome in parameter selection when processing such non-stationary weak signals, but also often erase high-frequency inflection points containing geological features, resulting in distorted attenuation curves and late-stage phase lag. Deep learning denoising models introduced in recent years still have significant drawbacks: First, transient electromagnetic single-sample sequences are extremely long, and if a global attention mechanism is used, the computational complexity increases quadratically, making it difficult to meet the needs of real-time processing in the field; second, forcibly splicing a one-dimensional sequence into a two-dimensional image and inputting it into a traditional CNN will destroy temporal continuity; third, the training phase often uses a single mean square error as the loss function, imposing an excessively large quadratic penalty on high-amplitude random atmospheric pulses, causing the network weights to be "biased" by extreme noise, severely reducing the reconstruction accuracy of late-stage weak signals.

[0004] Therefore, there is an urgent need for a fast noise reduction method for transient electromagnetic data that can reduce computational overhead while preserving effective late signals and is robust to atmospheric pulses, in order to solve the problems of easy erasure of late signals, high computational overhead and poor robustness in existing technologies. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the above-mentioned deficiencies in the prior art by providing a fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration achieves fast and robust noise reduction of complex environmental noise while preserving the effective signal in the later stage.

[0006] To solve the above problems, the present invention adopts the following technical solution: A fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration includes the following steps: S1. Acquire transient electromagnetic one-dimensional discrete time-series signals containing environmental noise, reconstruct and map them into high-dimensional feature tensors using a pseudo-two-dimensional tensor folding strategy, and perform global extremum adaptive normalization processing on the high-dimensional feature tensors to obtain zero-mean symmetric standardized input tensors. S2, Construct a deep residual denoising network, which contains multiple cascaded residual blocks, and establish feedforward jump connections between the network input and output; S3, the deep residual denoising network is trained using the standardized input tensor, a multi-task joint loss function including mean square error and mean absolute error is constructed, and the network weights are optimized by combining dynamic learning rate annealing strategy so that the network fits the target close to the real noise tensor. S4. In the inference phase, the transient electromagnetic data to be denoised is preprocessed in the same way as in S1 and then input into the trained deep residual denoising network to predict the background noise tensor. The predicted background noise tensor is subtracted from the input tensor and inverse normalization is performed to obtain the denoised transient electromagnetic clean signal.

[0007] Furthermore, the pseudo-two-dimensional tensor folding strategy in step S1 includes: folding a one-dimensional discrete sequence of length L into a high-dimensional tensor of size W×H×C according to a preset spatial dimension matrix deformation function, where W×H=L and C is the number of channels.

[0008] Furthermore, the global extremum adaptive normalization process in step S1 is used to map the input tensor to a symmetric interval, eliminating the amplitude differences between different samples.

[0009] Furthermore, the network structure in step S2 sequentially includes a feature initialization layer, a deep residual concatenation module, and a feature reconstruction output layer; The deep residual concatenation module consists of multiple residual blocks connected in series. The bias terms of the convolutional layer are removed from each residual block, and the calculation process follows "convolution-regularization-activation-convolution-regularization-skip addition-activation". By introducing a feedforward skip connection, the model's fitting target is transformed from a clean signal to a noise residual mapping, and the training label is set to the real noise, i.e., the noisy signal minus the clean signal. In the inference phase, the clean signal is obtained by subtracting the noise tensor predicted by the network from the noisy signal.

[0010] Furthermore, in step S2, the feedforward skip connection directly passes the network input to the output, so that the intermediate nonlinear layer only fits the noise residual mapping; the supervision label in the training phase is set as the noisy signal minus the clean signal; in the inference phase, the clean signal is obtained by subtracting the noise tensor predicted by the network from the noisy signal.

[0011] Furthermore, the multi-task joint loss function in step S3 is composed of a weighted average of mean squared error and mean absolute error, wherein the mean squared error is used to accelerate network convergence, and the mean absolute error is used to enhance robustness to high-amplitude noise.

[0012] Furthermore, the dynamic learning rate annealing strategy in step S3 includes: monitoring the loss value on the validation set in real time during training; and reducing the current learning rate by a preset decay coefficient when the validation set loss does not decrease within multiple consecutive training cycles until the lower limit of the learning rate is reached.

[0013] Furthermore, in step S4, during the inference phase, data loading and forward inference of the deep residual denoising network are deployed in an asynchronous data pipeline, and the asynchronous preloading of background data and the overlapping parallelism of matrix calculation are achieved through the underlying scheduling protocol.

[0014] Furthermore, the inverse normalization process in step S4 is the inverse operation of the normalization process in step S1, restoring the normalized tensor output by the network to its original physical dimensions.

[0015] Furthermore, the feature initialization layer of the deep residual denoising network in step S2 includes convolutional layers, batch normalization layers, and activation functions, which are used to expand the input tensor to a high-dimensional feature space; the feature reconstruction output layer is used to map the high-dimensional feature space back to the output tensor.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention employs a pseudo-two-dimensional tensor folding strategy to reduce the dimensionality of one-dimensional long sequences and reconstruct them into high-dimensional matrices. It cleverly transforms the temporal coherence of signals into spatial adjacency, enabling standard lightweight two-dimensional convolutional kernels to accurately capture local high-frequency noise. This fundamentally overcomes the computational bottleneck of traditional algorithms and global attention models when processing long sequences. Combined with the engineering deployment of asynchronous data pipeline technology, this system achieves a total processing and rendering time of millimeter-level compression (average measured time between 40ms and 50ms) when performing forward computation on a single folded high-dimensional tensor with a size of 30×30×1. Compared to traditional serial algorithms such as mode decomposition, which require several minutes of iterative looping, this invention achieves a leap in both operational efficiency and massive data throughput by several orders of magnitude.

[0017] This invention introduces a noise residual mapping mechanism based on skip connections into deep networks, changing the traditional approach of "fitting the signal" to "fitting only the noise." This design allows weak low-frequency baseband signals containing geological features to penetrate the network without loss, effectively solving the problems of late signal distortion and phase lag that traditional filtering easily leads to, and significantly improving the high-fidelity restoration capability of extremely weak effective signals. Quantitative measurements on a large-scale noisy test set show that after reconstruction using the method of this invention, the denoised waveform not only eliminates phase lag and late distortion, but also strictly controls the mean absolute error (MAE) and root mean square error (RMSE) between the predicted clean signal and the theoretical reference signal to below 0.1, and the coefficient of determination (R²) is also significantly improved. 2 The value remains consistently above 0.9. This indicates that the present invention can retain more than 90% of the effective physical response characterizing deep geological features, possessing extremely high theoretical signal fidelity.

[0018] This invention constructs a multi-task joint loss function that fuses mean squared error (MSE) and mean absolute error (MAE), supplemented by a dynamic learning rate annealing strategy. This ensures rapid convergence of the model to broadband Gaussian white noise while effectively suppressing the extreme interference of high-amplitude atmospheric pulses on the weights by utilizing the absolute error characteristics of MAE. This synergistic optimization of "algorithm architecture + loss constraint" significantly improves the model's engineering robustness and real-time inference efficiency in complex geophysical exploration sites. Independent repeated impact experiments with multiple modified random seeds demonstrate that, when facing non-stationary interference from extremely complex geophysical environments and the superposition of thermal noise from different equipment, the method of this invention can stably improve the signal-to-noise ratio (SNR) by an average of over 20 dB. Simultaneously, the standard deviation of its SNR gain is strictly controlled within 1 dB, and the standard deviation of the normalized mean squared error (NMSE) is extremely low (less than 0.01), proving that the model has extremely low sensitivity to network initialization weights and can adapt remarkably stably to various highly interference-prone sites such as mining areas and cities, providing a solid data quality guarantee for high-precision geological inversion of transient electromagnetic data. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration of the present invention. Figure 2 This is a schematic diagram illustrating the principle of a pseudo-two-dimensional tensor quantization folding strategy for one-dimensional time-series signals. Figure 3 This is a schematic diagram of the topology of a deep residual denoising network (ResNet); Figure 4 This is a comparison diagram of the original noisy signal and the noise-reconstructed signal of this invention. Detailed Implementation

[0020] The technical solutions of the invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without creative effort are within the scope of the invention.

[0021] In the description of this invention, it should be noted that the terms "above" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience and simplification of the description and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connection," "setting," "installation," "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] like Figure 1 As shown, this invention discloses a fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration, comprising the following steps: S1, Data Acquisition and Tensor Quantization Preprocessing: In electromagnetic inversion and denoising, high-quality data input is fundamental for feature extraction. Since it is difficult to obtain clean labels from field measurements, this embodiment uses forward numerical simulation to generate a theoretically clean signal sequence S(t) with gradual decay and steep anomalous abrupt changes. Subsequently, mathematical modeling is used to generate 50Hz / 150Hz periodic power frequency noise, high-amplitude random atmospheric pulses, and broadband Gaussian white noise. These mixed noises are then superimposed onto the clean signal at different signal-to-noise ratio gradients to form a noisy observation sequence X(t) = S(t) + N(t), constructing the dataset required for model training. Specifically, the detection principle of the transient electromagnetic method is based on the secondary induction excited after the step current is turned off. The vortex field (i.e., the smoke ring effect) is introduced because the signal decreases sharply in the late stage of detection according to the power function of time (5 / 2), resulting in an extremely weak effective response reflecting deep geological information, which is easily submerged by environmental noise in the same frequency band. The introduced 50Hz / 150Hz power frequency interference mainly originates from urban power grids or high-power electromechanical equipment. High-amplitude random atmospheric pulses mostly originate from lightning or electrostatic discharge, and exhibit non-stationary outlier mutations in the time domain. Gaussian white noise simulates the internal thermal noise of the instrument. This method, which integrates physical forward modeling with real complex noise modeling, provides a reliable training basis for subsequent deep learning models covering various extreme geological environments and signal-to-noise ratio gradients. The original transient electromagnetic observation sequence is quite long, and directly using one-dimensional convolutions can easily lead to a surge in network parameters and a decrease in the ability to detect local high-frequency spikes; therefore, such as Figure 2 As shown, this step implements a pseudo-two-dimensional tensor folding strategy: reconstructing a one-dimensional long sequence into a high-dimensional feature tensor, transforming the temporal coherence of the original signal into spatial adjacency in the pseudo-two-dimensional topology. In the pseudo-two-dimensional topological space, random atmospheric pulses and Gaussian white noise appear as high-frequency isolated texture bright spots, which are easily captured and accurately stripped by the two-dimensional receptive field. Meanwhile, the low-frequency decay trend in the deep regions is presented as a smooth pixel gradient transition, which is preserved to the maximum extent during downsampling. This transformation allows the subsequent network to capture both local high-frequency noise and global low-frequency decay trend simultaneously using only a lightweight two-dimensional convolutional kernel, without the need for a large number of parameters in the one-dimensional long convolutional kernel. This fundamentally breaks through the computational bottleneck of traditional deep learning models when processing long sequences. Finally, global extremum adaptive normalization is performed on the high-dimensional feature tensor. Since the amplitude of the early and late transient electromagnetic signals may span several orders of magnitude, without normalization, it is easy to cause saturation of the network activation function or gradient explosion. This step maps the data to the zero-mean symmetric interval, eliminates the amplitude difference between different samples, improves the topology of the error hyperplane, and provides a stable input distribution for subsequent network training. S2, Construct a deep residual denoising network: An end-to-end deep residual denoising network is constructed, which consists of a feature initialization layer, a deep residual cascade module, and a feature reconstruction output layer, with a feedforward jump connection established between the network input and output. The feature initialization layer expands the input tensor to a high-dimensional feature space. The deep residual cascade module is composed of multiple residual blocks connected in series, used to extract deep features step by step. The feature reconstruction output layer maps the high-dimensional features back to the output tensor. The feedforward jump connection directly passes the input to the output, so that the intermediate layers of the network only need to fit the noise residual mapping instead of the complete clean signal, thereby greatly reducing the difficulty of feature extraction and ensuring the lossless penetration of weak effective signals. S3, Model Training and Optimization: The normalized input tensor obtained in step S1 is used to train the deep residual denoising network. A multi-task joint loss function including mean squared error (MSE) and mean absolute error (MAE) is constructed: MSE accelerates network convergence by utilizing its smooth and differentiable properties, while MAE enhances robustness to high-amplitude atmospheric pulses by utilizing its insensitivity to outliers. At the same time, combined with a dynamic learning rate annealing strategy, the validation set loss is monitored in real time during training. When the loss no longer decreases for several consecutive training cycles, the learning rate is automatically reduced proportionally to achieve optimization of "large step size exploration in the early stage and small step size refinement in the later stage", ensuring that the network converges stably to the global optimum. S4, Inference and Signal Reconstruction: In the inference phase, the transient electromagnetic data to be denoised is preprocessed in the same way as in S1 and then input into the trained deep residual denoising network. The network predicts the background noise tensor through forward inference, and then uses the input tensor to subtract the predicted background noise tensor to separate the high-fidelity clean feature tensor. Finally, inverse normalization is performed (i.e., the inverse operation of the normalization in S1) to restore the data to its original physical dimensions, obtaining the denoised transient electromagnetic clean signal. The denoising effect can be referenced. Figure 4 The waveforms of the original noisy signal and the denoised reconstructed signal are shown in the comparison. Actual measurements show that this method can compress the average processing time per sample to 40-50 milliseconds, improve the signal-to-noise ratio by more than 20 dB on average, and increase the coefficient of determination R0. 2 It remains stable above 0.9.

[0025] Through the above steps, this method effectively solves the technical problems of traditional filtering algorithms easily erasing late effective signals, the large computational cost of deep learning models, and the poor robustness of a single loss function to atmospheric pulses, and achieves high-fidelity and high-efficiency noise reduction of transient electromagnetic data.

[0026] Furthermore, the one-dimensional discrete sequence X=[x1,x2,...,xL] of length L is folded and mapped into a high-dimensional tensor X of size W×H×C according to a preset spatial dimension matrix transformation function (reshape). tensorWhere W×H=L, and the number of channels C is usually 1 (or greater than 1 if it is multi-channel data). For example, when L=900, it can be folded into a 30×30×1 tensor. This strategy makes the temporal adjacency of signals appear as spatial adjacency in the two-dimensional topology, so that the two-dimensional convolution kernel can be used to simultaneously capture local high-frequency noise (such as atmospheric pulses) and global low-frequency attenuation trends (such as late signals), avoiding the parameter inflation problem caused by directly processing one-dimensional long sequences.

[0027] Furthermore, the global extremum adaptive normalization process in step S1 is used to map the input tensor to the symmetric interval [−1,1], eliminating the network training instability caused by amplitude differences between different samples (early and late signals may span several orders of magnitude). Due to the "smoke ring effect" of electromagnetic induction, the early and late amplitudes of transient electromagnetic signals typically span 10... 4 Up to 10 6 The magnitude is several orders of magnitude. Without rigorous algebraic preprocessing, directly inputting such a large dynamic range of raw physical dimension data into a deep neural network will directly cause the nonlinear activation function (such as ReLU) of the convolutional layer to fall into the hard saturation region and fail, or cause gradient explosion, resulting in the network failing to converge.

[0028] The formula for adaptive normalization of global extrema is: ; in, X norm The normalized input tensor; X tensor This is the high-dimensional feature tensor after pseudo-two-dimensional tensor quantization; g max and g min These are the global maximum and minimum values ​​in the dataset or the current batch of data, respectively. ϵ is a very small positive number (e.g., 10) to prevent division by zero anomalies. -8 ); Coefficient 2 is a scaling factor that maps the data from the original range to the interval [-1, 1]. This mapping mechanism effectively stretches the extremely unevenly distributed electromagnetic attenuation data to a uniform scale, improves the topological structure of the high-dimensional error hyperplane, and ensures the smoothness of gradient descent optimization. After inference is complete, the system performs a rigorous reversible inverse normalization algebraic operation based on the extreme values ​​of the corresponding batch, reconstructing the output back to the true physical dimensions for accurate calculation of apparent resistivity across the entire region.

[0029] Furthermore, the network structure in step S2 includes, in sequence: Feature initialization layer: It consists of a convolutional layer with 64 3×3 filters, a batch normalization (BN) layer and a ReLU activation function, which is used to expand the single-channel input tensor to a high-dimensional feature space (64 channels) and initially extract shallow local texture and temporal correlation features; Deep residual cascade module: composed of multiple (e.g., 7) residual blocks connected in series; such as Figure 3 As shown, the bias terms of the convolutional layers are removed from each residual block (because the subsequent BN layer will perform a mean-removal operation, removing the bias can save computation and avoid overfitting), and strictly follows the calculation process of "convolution-BN-ReLU-convolution-BN-skip addition-ReLU"; this module enables the signal to undergo multiple nonlinear transformations in the deep network, gradually separating the geological response from the background noise. Feature reconstruction output layer: used to map the high-dimensional feature space back to the output tensor, typically composed of convolutional layers and activation functions; Feedforward Skip Connection: Introduces an identity mapping (Skip Connection) directly from the network input to the output, transforming the model's fitting target from the pure signal S. clean Convert to a noise residual mapping F(X). The training labels are set to real noise. ; Among them, Y label Labels for actual noise residuals; X noisy Input is a noisy signal; S clean For a pure signal; During the inference phase, the clean signal is obtained by subtracting the noise tensor predicted by the network from the noisy signal: ; in, S denoised The noise-reconstructed signal from the network; N pred The noise tensor predicted by the network; This residual mapping mechanism allows the network to learn only high-frequency noise with a mean approaching zero, while weak low-frequency baseband signals containing geological features can penetrate the network without loss, thereby significantly improving the reconstruction accuracy of weak signals in the later stages.

[0030] Furthermore, the feedforward skip connection in step S2 directly passes the network input to the output, so that the intermediate nonlinear layer only fits the noise residual mapping, rather than directly fitting the clean signal; the supervision label in the training phase is the noisy signal minus the clean signal; in the inference phase, the clean signal is obtained by subtracting the noise tensor predicted by the network from the noisy signal; this feedforward skip connection effectively avoids the problem of weak effective signal attenuation and distortion in the layer-by-layer transmission process of traditional autoencoders.

[0031] Furthermore, the feature initialization layer of the deep residual denoising network in step S2 includes convolutional layers, batch normalization layers, and activation functions, which are used to expand the input tensor to a high-dimensional feature space, for example, from a single-channel input to 64 channels, thereby providing a rich feature base for the deep network and helping to extract shallow local texture and temporal correlation features. The feature reconstruction output layer is used to map the high-dimensional feature space back to the output tensor, usually compressing the multi-channel feature map back to a single channel, and outputting the predicted noise tensor or the reconstructed clean signal. Through the above settings, the network can complete the efficient mapping from noisy input to noisy residuals end-to-end, reducing model complexity while ensuring denoising effect.

[0032] In this embodiment, the multi-task joint loss function in step S3 is composed of a weighted average of mean squared error (MSE) and mean absolute error (MAE), where the MSE is used to accelerate network convergence, and the MAE is used to enhance robustness to high-amplitude noise; its formula is: ; in, 𝓛 tatal This represents the total loss value calculated in the current batch for a single forward pass of the model. Y i The true noise residual label for the i-th sample; Let be the noise residual of the i-th sample predicted by the network; α and β are hyperparameters that adjust the weight distribution of mean squared error (MSE) and mean absolute error (MAE), respectively. N is the batch size of samples loaded into the model memory for gradient iteration in a single operation (e.g., set to 128). The multi-task joint loss function can also be simplified to: ; The smooth differentiability of MSE helps to quickly probe the error plane in the early stages of training and capture the main Gaussian white noise; while MAE is insensitive to outliers such as high-amplitude atmospheric pulses, effectively resisting the interference of outliers on the gradient direction, and finely adjusting parameters in the later stages of training to avoid being "biased" by extreme noise. The combination of the two achieves a balance between fast convergence and strong robustness.

[0033] Furthermore, the dynamic learning rate annealing strategy in step S3 includes: monitoring the loss value on the validation set in real time during training; if the validation set loss does not show a significant decrease within K consecutive training epochs (i.e., entering a plateau), then the annealing mechanism is triggered. Reduce the current learning rate LR by the set decay factor γ: ; in, LR new The updated learning rate; LR old This is the current learning rate; γ is the attenuation coefficient;

[0034] When optimizing low-level parameters, the model is typically paired with an adaptive optimizer (such as the Adam algorithm) that has a momentum estimation mechanism. At the beginning of training, a relatively large initial learning rate (e.g., 10) is set. -4 This gives the model a strong global exploration drive, allowing it to quickly overcome and escape the misleading effects of local saddle points in the multidimensional error space. However, in the later stages of training when the model is approaching the optimal solution, an excessively large learning rate can easily cause severe oscillations and loss value rebounds near the extreme points. The dynamic learning rate annealing monitoring mechanism introduced in this invention obtains the performance evaluation status of the validation set in real time. Once it is determined that the network convergence has entered a plateau period (i.e., it cannot advance further for K consecutive epochs), it automatically triggers a step-by-step reduction (e.g., halving by γ=0.5). This optimization dynamic mechanism of "large step size escape in the early stage and small step size refinement in the later stage" enables the model to achieve absolutely stable, rebound-free convergence within a limited number of epochs (e.g., 50 to 100 epochs), statistically eliminating the occurrence of model overfitting.

[0035] Furthermore, in the inference phase of step S4, data loading and forward inference of the deep residual denoising network are deployed in an asynchronous data pipeline, and the asynchronous preloading of background data and the overlapping parallelism of GPU matrix calculation are achieved through the underlying scheduling protocol (such as TensorFlow's AUTOTUNE).

[0036] Specifically, while the GPU executes forward inference of the high-dimensional tensor matrix for the current batch, the CPU asynchronously loads, normalizes, and shuffles the tensor data for the next batch from disk in parallel in the background. This architecture completely breaks through the I / O transmission bottleneck in processing massive geophysical data, reducing the average processing time per sample to 40–50 milliseconds, meeting the real-time processing requirements in the field. Simultaneously, the model should switch to deterministic inference mode at this stage, freezing the statistical parameters within the batch normalization layer to ensure absolute consistency of the output results.

[0037] Furthermore, the inverse normalization process in step S4 is the inverse operation of the normalization process in step S1, restoring the normalized tensor output by the network to its original physical dimensions.

[0038] Specifically, after the inference phase is completed, the system then performs a strictly reversible inverse normalization algebraic operation based on the global extrema and minimal positive constants of the corresponding batch saved in step S1, reconstructing the output back to the true physical dimensions. This operation ensures that the denoised transient electromagnetic signals can be directly used for the accurate calculation and geological interpretation of the apparent resistivity of the entire area, and is a key link connecting the deep learning denoising results with engineering geophysical applications.

[0039] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration, characterized in that, Includes the following steps: S1. Acquire transient electromagnetic one-dimensional discrete time-series signals containing environmental noise, reconstruct and map them into high-dimensional feature tensors using a pseudo-two-dimensional tensor folding strategy, and perform global extremum adaptive normalization processing on the high-dimensional feature tensors to obtain zero-mean symmetric standardized input tensors. S2, Construct a deep residual denoising network, which contains multiple cascaded residual blocks, and establish feedforward jump connections between the network input and output; S3, the deep residual denoising network is trained using the standardized input tensor, a multi-task joint loss function including mean square error and mean absolute error is constructed, and the network weights are optimized by combining dynamic learning rate annealing strategy so that the network fits the target close to the real noise tensor. S4. In the inference phase, the transient electromagnetic data to be denoised is preprocessed in the same way as in S1 and then input into the trained deep residual denoising network to predict the background noise tensor. The predicted background noise tensor is subtracted from the input tensor and inverse normalization is performed to obtain the denoised transient electromagnetic clean signal.

2. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration according to claim 1, characterized in that, The pseudo-two-dimensional tensor folding strategy in step S1 includes: folding a one-dimensional discrete sequence of length L into a high-dimensional tensor of size W×H×C according to a preset spatial dimension matrix deformation function, where W×H=L and C is the number of channels.

3. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration according to claim 1, characterized in that, The global extremum adaptive normalization process in step S1 is used to map the input tensor to a symmetric interval, eliminating the amplitude differences between different samples.

4. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration according to claim 1, characterized in that, The network structure in step S2 includes, in sequence, a feature initialization layer, a deep residual concatenation module, and a feature reconstruction output layer; The deep residual concatenation module consists of multiple residual blocks connected in series. The bias terms of the convolutional layer are removed from each residual block, and the calculation process follows "convolution-regularization-activation-convolution-regularization-skip addition-activation". By introducing a feedforward skip connection, the model's fitting target is transformed from a clean signal to a noise residual mapping, and the training label is set to the real noise, i.e., the noisy signal minus the clean signal. In the inference phase, the clean signal is obtained by subtracting the noise tensor predicted by the network from the noisy signal.

5. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration according to claim 4, characterized in that, In step S2, the feedforward skip connection directly passes the network input to the output, so that the intermediate nonlinear layer only fits the noise residual mapping; the supervision label in the training phase is set as the noisy signal minus the clean signal; in the inference phase, the clean signal is obtained by subtracting the noise tensor predicted by the network from the noisy signal.

6. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration according to claim 4, characterized in that, The feature initialization layer of the deep residual denoising network in step S2 includes convolutional layers, batch normalization layers, and activation functions, which are used to expand the input tensor to a high-dimensional feature space; the feature reconstruction output layer is used to map the high-dimensional feature space back to the output tensor.

7. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration according to claim 1, characterized in that, The multi-task joint loss function in step S3 is composed of a weighted average of mean squared error and mean absolute error, where mean squared error is used to accelerate network convergence and mean absolute error is used to enhance robustness to high-amplitude noise.

8. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration according to claim 1, characterized in that, The dynamic learning rate annealing strategy in step S3 includes: real-time monitoring of the loss value on the validation set during training; when the validation set loss does not decrease within multiple consecutive training cycles, reducing the current learning rate by a preset decay coefficient until the lower limit of the learning rate is reached.

9. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration according to claim 1, characterized in that, In step S4, during the inference phase, data loading and forward inference of the deep residual denoising network are deployed in an asynchronous data pipeline, and the asynchronous preloading of background data and the overlapping parallelism of matrix calculation are achieved through the underlying scheduling protocol.

10. The fast noise reduction method for transient electromagnetic data based on tensor mapping and multi-task collaboration according to claim 1, characterized in that, The inverse normalization process in step S4 is the inverse operation of the normalization process in step S1, which restores the normalized tensor output by the network to its original physical dimensions.

Images