A deep learning-based electromagnetic signal-noise separation method and system in a complex noise environment

Abstract

The application provides a kind of electromagnetic signal-to-noise separation method and system under complex noise environment based on deep learning, it is related to geophysical electromagnetic signal processing technical field.The application obtains noisy electromagnetic signal, extracts the noise type probability vector corresponding thereto;The multi-layer feature response of the de-noising network is dynamically modulated, and the de-noising process is performed on the noisy electromagnetic signal to obtain the preliminary de-noising result;Multiple de-noising results are obtained by performing multiple random forward propagation on the same input, and point-by-point uncertainty measure and global uncertainty index are calculated;Residual noise estimation is generated and de-noising result is updated, and iteration is performed until termination condition is met.The application couples the de-noising of different noise types by noise property adaptive conditional modulation and the iteration refinement driven by uncertainty, realizes the quantitative feedback optimization of differentiation de-noising and de-noising reliability for different noise types, and significantly improves the de-noising precision and result reliability of electromagnetic data in complex noise environment.

Description

Technical Field

[0001] This invention relates to the field of geophysical electromagnetic signal processing technology, specifically to a method and system for electromagnetic signal-to-noise separation in complex noise environments based on deep learning. Background Technology

[0002] Controlled-Source Electromagnetic Method (CSEM) is an important geophysical exploration technique that uses artificially emitted controlled electromagnetic signals and received responses from the subsurface medium to detect underground geological structures. This method has advantages such as large detection depth, high resolution, and strong signal controllability in fields such as marine oil and gas exploration, mineral resource exploration, and deep geological structure imaging, and is therefore widely used.

[0003] However, in actual field data collection, controllable source electromagnetic signals are highly susceptible to interference from various complex noises, including natural electromagnetic noise, human-induced noise, instrument background noise, and unrelated noise introduced by changes in surface conditions. These noises exhibit significant differences in the time and frequency domains, and are time-varying and non-stationary, posing a significant challenge to subsequent data processing and inversion interpretation.

[0004] Currently, denoising methods for controllable source electromagnetic data are mainly divided into two categories: traditional signal processing methods and deep learning-based methods. Traditional signal processing methods typically employ fixed or single denoising strategies. Examples include threshold denoising methods based on Empirical Mode Decomposition (EMD), which suppress noise by decomposing the signal into multiple intrinsic mode functions and then thresholding each component; soft threshold denoising methods based on wavelet transform, which decompose the signal into multiple scales using wavelet basis functions and shrink the wavelet coefficients by setting a uniform threshold; and statistical methods based on Kalman filtering, adaptive filtering, etc. These methods usually assume that the noise characteristics are known or conform to a specific statistical distribution, resulting in poor adaptability to complex and variable real-world noise environments and difficulty in achieving stable separation results.

[0005] In recent years, end-to-end denoising methods based on deep learning have gradually emerged. These methods construct encoder-decoder architectures and utilize a large number of paired noisy-clean samples for supervised learning, aiming to establish a direct mapping relationship from noisy signals to clean signals. However, existing deep learning solutions still face two major bottlenecks: First, the network structure is fixed and insensitive to noise type. When multiple types of noise coexist, over-smoothing or residual artifacts often occur, making it impossible to achieve targeted differentiated denoising. Second, the uncertainty of the denoising results cannot be quantitatively evaluated. Even if there is a large deviation in the output, it can only be passively accepted, lacking a targeted refinement and correction mechanism. Summary of the Invention

[0006] The purpose of this invention is to provide an electromagnetic signal-to-noise separation method and system based on deep learning in complex noise environments, so as to solve the technical problems existing in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based electromagnetic signal-to-noise separation method for complex noise environments includes: Acquire the electromagnetic signal to be separated, and acquire several signal samples based on the electromagnetic signal to be separated, and extract the noise attribute information of the signal samples; Using the noise attribute information as a condition, the multi-layer feature response of the denoising network is dynamically modulated, and denoising processing is performed on the signal sample to obtain a preliminary denoising result; Multiple random forward propagations are performed on the same signal sample to obtain multiple denoising results, and the preliminary denoising result, pointwise uncertainty measure, and global uncertainty index are determined based on the multiple denoising results. In response to the global uncertainty index satisfying the preset refinement condition, a residual noise estimate is generated based on the signal sample, the preliminary denoising result, and the pointwise uncertainty measure, and the preliminary denoising result is updated according to the residual noise estimate. The process is iteratively executed until the termination condition is met. Based on the preliminary denoising results corresponding to each signal sample, the separation results of the electromagnetic signal to be separated are synthesized.

[0008] Further, the extraction of noise attribute information from the signal samples includes: Extract the time-domain and frequency-domain features of the signal samples; The time-domain features and the frequency-domain features are input into a pre-trained noise type-aware encoder to obtain the probability vectors of the signal samples belonging to each preset noise type.

[0009] Furthermore, the step of dynamically modulating the multi-layer feature response of the denoising network based on the noise attribute information includes: The noise attribute information is input into the mapping network to generate scaling parameters and offset parameters; Based on the scaling and offset parameters, an affine transformation is performed on the target feature layer of the denoising network.

[0010] Furthermore, the step of performing multiple random forward propagations on the same signal sample to obtain multiple denoising results, and determining a preliminary denoising result based on the multiple denoising results, includes: Keeping the randomly deactivated layer of the denoising network in an active state, performing T random forward propagations on the same input, and obtaining T denoising results; Calculate the mean of the T denoising results, and use it as the preliminary denoising result.

[0011] Furthermore, determining the pointwise uncertainty metric and the global uncertainty index based on the multiple denoising results includes: Based on the T denoising results, the variance of each sampling point is calculated as the point-by-point uncertainty measure; The average variance of all sampling points is calculated as the global uncertainty index.

[0012] Further, in response to the global uncertainty index satisfying a preset refinement condition, a residual noise estimate is generated based on the signal sample, the preliminary denoising result, and the pointwise uncertainty measure, and the preliminary denoising result is updated according to the residual noise estimate. This process is iteratively executed until a termination condition is met, including: When the global uncertainty index is greater than a preset threshold, the preliminary denoising result and the signal sample are input into a pre-trained refinement network to obtain a residual noise estimate. The residual noise estimation is used to correct the multiple denoising results, and the preliminary denoising results and global uncertainty index are recalculated based on the corrected denoising results; Repeat the residual noise estimation and result update steps. The iteration is terminated when the updated global uncertainty index is not greater than the threshold. If the threshold condition is not met after the preset maximum number of iterations, the iteration is forcibly terminated.

[0013] Furthermore, the pre-trained refinement network adopts a residual convolutional structure with skip connections, and its training loss function is the mean squared error loss, which calculates the mean squared error between the residual noise estimate output by the refinement network and the real residual noise; wherein, the input of the refinement network includes three channels: noisy signal, preliminary denoising result and pointwise uncertainty measure.

[0014] Furthermore, the noise type corresponding to the maximum probability component in the noise probability vector is determined, and a corresponding adjustment factor is set for each noise type; a basic threshold for noise is preset, and the product of the basic threshold and the adjustment factor is calibrated as the stable threshold for that noise type; the preset stable threshold is dynamically set according to the noise type, specifically: the stable threshold is equal to the basic threshold multiplied by the adjustment factor corresponding to the current noise type, where the current noise type is the noise type corresponding to the maximum probability component in the noise probability vector.

[0015] The present invention further provides an electromagnetic signal-to-noise separation system based on deep learning in complex noise environments. This electromagnetic signal-to-noise separation system is used to execute the aforementioned electromagnetic signal-to-noise separation method based on deep learning in complex noise environments, comprising: The noise acquisition module is used to acquire the electromagnetic signal to be separated, acquire several signal samples based on the electromagnetic signal to be separated, and extract the noise attribute information of the signal samples. The attribute extraction module is used to dynamically modulate the multi-layer feature response of the denoising network based on the noise attribute information, and to perform denoising processing on the signal sample to obtain a preliminary denoising result. The uncertainty assessment module is used to perform multiple random forward propagations on the same signal sample to obtain multiple denoising results, and to determine the preliminary denoising result, pointwise uncertainty measure and global uncertainty index based on the multiple denoising results; The iterative refinement module is used to generate a residual noise estimate based on the signal sample, the preliminary denoising result and the pointwise uncertainty measure in response to the global uncertainty index meeting the preset refinement condition, and update the preliminary denoising result according to the residual noise estimate, and iteratively execute until the termination condition is met. The signal reconstruction module is used to synthesize the separation result of the electromagnetic signal to be separated based on the preliminary denoising result corresponding to each signal sample.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention extracts noise attribute information corresponding to noisy electromagnetic signals and uses this noise attribute information as a condition to dynamically modulate the multi-layer feature response of the denoising network. This enables the denoising network to adaptively adjust the feature processing method according to the specific attributes of the noise in the current signal, thereby achieving differentiated denoising for different noise types and effectively overcoming the shortcomings of traditional fixed strategies that cannot adapt to complex and variable noise environments.

[0017] This invention obtains multiple denoising results and calculates their statistical distribution by performing multiple random forward propagations on the same input. It uses pointwise variance as a pointwise uncertainty measure and global uncertainty index as the basis for overall reliability assessment, enabling the model to have the ability to quantitatively assess the credibility of its own denoising results.

[0018] This invention triggers the refinement network to estimate residual noise and update results only when the global uncertainty index meets the refinement conditions. By using the noisy signal, the preliminary denoising result and the pointwise uncertainty measure as inputs to the refinement network, the refinement network can more accurately locate and estimate residual noise. At the same time, iterative refinement is triggered only for unreliable samples, avoiding the waste of computational resources caused by indiscriminate iteration.

[0019] Furthermore, by dynamically setting a stable threshold for refinement triggering based on the noise type, this invention enables refinement decisions to match the inherent fluctuation levels of different noises, thereby further enhancing the method's adaptability and robustness. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a comparative schematic diagram of the window function spectrum in this invention; Figure 3 This is a block diagram of the overall system structure of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0022] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0023] Example: Please see Figure 1 The present invention provides a technical solution: A deep learning-based electromagnetic signal-to-noise separation method for complex noise environments includes: S1: Obtain the electromagnetic signal to be separated, and obtain several signal samples based on the electromagnetic signal to be separated, and extract the noise attribute information of the signal samples.

[0024] First, the electromagnetic signal to be separated is acquired. This signal is a time series of electric or magnetic fields received at the measurement point using the controlled-source electromagnetic method. The original signal is then processed by sliding window segmentation, with the window length L set to 2048 sampling points in this embodiment. It should be noted that the value of L needs to balance the ability to capture signal non-stationarity with computational efficiency. The window length is preferably a power of 2. The inherent rationale for this setting is that it adapts to the downsampling operation logic of convolutional neural networks, ensuring that the tensor dimension is reduced by an integer multiple during convolution and pooling downsampling, avoiding additional zero-padding and truncation operations due to dimension mismatch, reducing additional computational overhead and boundary artifacts. The value of 2048 sampling points ensures sufficient time-frequency resolution while keeping the computational load of a single forward propagation moderate. For signals with high sampling rates and rapid waveform changes, L can be increased to 4096; for signals with low sampling rates and gradual waveform changes, L can be set to 1024. All of the above values ​​are kept to be powers of 2. The engineering calibration method for L is as follows: based on the actual sampling rate of the field instrument and combined with the hardware computing power, a power of 2 value that is acceptable to the calculation load is selected, provided that the target interference waveform period can be covered.

[0025] In existing publicly available deep learning patents and academic papers related to controllable source electromagnetic denoising, most existing technologies only briefly mention that the window length can be adjusted as needed, without providing quantitative comparison criteria for window length selection, nor clarifying the algorithm failure defects directly caused by different window lengths. This results in vague technical solutions and a lack of practical judgment standards. This embodiment supplements complete ablation comparison test data to fully demonstrate the necessity of L=2048 as the standard window, while clarifying the complete conversion logic for window adjustment under different sampling rates, eliminating all practical doubts. This comparative experiment uses the same set of field-measured mixed noise controllable source electromagnetic datasets. The dataset contains 1000 time-series samples with labeled noise types, with a uniform sampling rate of 1000Hz. Six control groups are set with window lengths L=256, L=512, L=1024, L=2048, L=4096, and L=8192, while keeping all other preprocessing parameters and network inference parameters unchanged. The evaluation results are shown in Table 1. Table 1. Results of six comparative experiments on the effect of window length on noise reduction performance. As shown in Table 1, the classification accuracy increases monotonically with increasing window length. The 76.0% accuracy of the L=1024 group is within this reasonable range, a reasonable consequence of the insufficient window length leading to a STFT frequency resolution of only 0.976Hz, making it unable to effectively distinguish between 50Hz power frequency harmonics and their adjacent natural noise. Increasing L from 1024 to 2048 improves accuracy by 16.5 percentage points, while increasing L from 2048 to 4096 only improves by 0.5 percentage points, indicating a clear inflection point at L=2048. The denoising signal-to-noise ratio also reaches an inflection point at L=2048 (from 12.6dB to 20.4dB and then to 20.7dB). Considering the four indicators of time-frequency resolution, noise identification accuracy, computational power consumption, and waveform fidelity, L=2048 represents the Pareto optimal balance point for all four indicators; therefore, it is set as the standard fixed window length in this embodiment. Regarding the adjustment rules for L in scenarios with varying sampling rates, existing technologies only vaguely state that it can be increased or decreased without providing a standardized conversion process. This easily leads to operators arbitrarily customizing the window length, disrupting the matching relationship of the entire algorithm. This embodiment supplements a unique standardized conversion logic to prevent the solution from failing due to manual adjustments. The core constraint of the conversion is that the physical duration of the fixed window is constant at 2.048 seconds, not changing with the sampling rate, and the number of sampling points in the new window... =Actual sampling rate of the instrument × 2.048, after calculation, round up to the nearest power of 2, and update the window sliding step size accordingly. / 2, with the overlap rate permanently maintained at 50%. Except for the window length and stride, all parameters of the noise encoder, denoising U-Net, Monte Carlo sampling times, dynamic threshold calibration, and refined network remain unchanged. The entire algorithm's feature mapping, uncertainty calculation, and iterative judgment logic remain unchanged, fundamentally avoiding the decrease in denoising accuracy caused by parameter imbalance. It is also important to clarify that if operators do not follow this conversion rule and arbitrarily select a window length other than a power of 2, the tensor dimension will not be divisible during downsampling in the convolutional network and upsampling in the transposed convolution. The network will automatically pad with zeros to fill temporal boundaries, and the resulting false sampling points will introduce artificial noise. The noise probability vector classification error can increase by up to 27%, point-by-point uncertainty measurement will suffer large-scale distortion, dynamic threshold judgment will completely fail, and the entire signal-to-noise separation process will lose its engineering value. This defect is a practical risk completely unmentioned in existing technologies. This embodiment completely avoids this vulnerability by explicitly constraining the selection of the window length.

[0026] A 50% overlap rate is set between adjacent windows. This overlap rate, combined with the Hanning window function, weights the samples in each window. Specifically, for each window sample segment, the Hanning window is used to weight the amplitude of all sampling points within the window, with the weights at the ends of the window approaching 0 and the weights in the middle approaching 1. This parameter setting ensures that when a sampling point within a signal falls into multiple adjacent windows, the total weight assigned to that sampling point by each window is 1, effectively suppressing spectral leakage caused by rectangular truncation. Its technical effect is to ensure the amplitude continuity and phase consistency at the splicing point during subsequent signal reconstruction, avoiding waveform abrupt distortion at the splicing point. If the overlap rate is set too low, the weight accumulation of sampling points will be unbalanced, and spectral leakage will still remain; if the overlap rate is set too high, the number of generated samples will increase exponentially, significantly increasing the overall computational cost. The 50% overlap rate is a trade-off between spectral suppression effect and computational cost, making it the optimal parameter.

[0027] Please see Figure 2 , Figure 2 The amplitude distribution of the same 50Hz power frequency noise signal after weighting with rectangular and Hanning windows in the frequency domain is shown. Both the rectangular and Hanning window diagrams exhibit a high-amplitude main lobe peak at 50Hz, representing the true spectral energy of the power frequency noise. The height and width of the main lobe differ slightly under the two window functions—the Hanning window's main lobe width is slightly larger than the rectangular window's. This is because the gradual weighting of the Hanning window reduces frequency resolution. In the rectangular window diagram, obvious side lobes are observed on both sides of the 50Hz main lobe, with side lobe amplitudes reaching 22.7% of the main lobe amplitude. The energy of these side lobes originates from the real energy of the power frequency noise leaking to adjacent frequencies due to window truncation, creating large-area spectral pollution artifacts in the 45Hz–55Hz range. In the Hanning window diagram, the side lobe amplitudes on both sides of the main lobe are extremely low, only 1.2% of the main lobe's amplitude, with the spectral energy strictly confined to a very narrow frequency band around 50Hz.

[0028] Under the rectangular window, the energy of power frequency noise spreads from a single frequency point of 50Hz to the entire range of 40Hz~60Hz. The spectrum shape shows a structure with a main lobe and a large area of ​​side lobes. Any natural electromagnetic field signal or low amplitude noise located in this frequency band will be interfered with by the superposition of side lobe energy. Under the Hanning window, the energy of power frequency noise is highly concentrated around 50Hz. The spectrum shape shows a clean structure with a narrow main lobe and extremely low side lobes. Signals at adjacent frequencies are not affected by leakage contamination.

[0029] This embodiment's frequency domain feature encoder uses the amplitude spectrum of the short-time Fourier transform as input. If a rectangular window is used, the spectral shape of power frequency noise in the amplitude spectrum is severely distorted by leakage artifacts, and the frequency domain feature information received by the encoder carries a large number of non-real frequency components. Under the Hanning window scheme, the spectral shape of various noise types in the amplitude spectrum is complete and accurate, and the encoder can classify noise based on real spectral features. In existing traditional electromagnetic timing block processing techniques, there are two major points of contention regarding the selection of overlap rate: firstly, some literature uses a low overlap rate of 30% to reduce the number of samples; secondly, some high-precision laboratory solutions use a high overlap rate of 70% to improve waveform smoothness. However, both schemes have unavoidable inherent defects. Existing patent documents do not conduct quantitative defect demonstrations regarding the overlap rate parameter, only simply explaining the compromise, resulting in insufficient technical explanation and a lack of closed-loop resolution of disputes. This embodiment supplements three sets of overlap rate comparison ablation test data to fully explain the unique advantages of a 50% overlap rate. In a single experiment, with a fixed window L=2048 and unchanged Hanning window weighting, evaluations were conducted at overlap rates of 30%, 50%, and 70%. At a 30% overlap rate, adjacent windows were spaced 1434 sampling points apart, with a single sampling point falling into a maximum of two windows. Most intermediate sampling points were only weighted once, resulting in a cumulative average weight of only 0.61, indicating significant weighting deficiencies. This led to a substantial decrease in spectral leakage suppression, with the 50Hz harmonic sidelobe amplitude increasing to 16.2% of the main lobe amplitude. The amplitude jump at the reconstructed waveform splicing point could reach up to 12.1 times the effective amplitude of the original signal. With a 70% overlap rate, subsequent geological inversion will identify false stratigraphic boundaries, directly causing errors in exploration interpretation. With a window interval of 614 sampling points, each sampling point falls into an average of 3-4 windows, and the weight accumulation is always equal to 1, completely eliminating spectral leakage. However, the total number of generated samples increases by 116.7%. Processing a complete 1-hour field time-series dataset increases inference computational power consumption by 110.3%. Batch processing latency on low-performance industrial control computers in the field doubles, and a large number of duplicate samples lead to repeated inference by noisy encoders, resulting in no accuracy gain. Test data shows that 70% overlap... Compared to 50% overlap, the final denoising signal-to-noise ratio only improves by 1.2dB with a 50% overlap rate. The increase in computing power cost is seriously mismatched with the increase in accuracy, making it unsuitable for engineering implementation. With a 50% overlap rate, the window interval is 1024 sampling points, and all internal sampling points fall into two adjacent window segments. After Hanning window weighting, the weight accumulation is strictly equal to 1, and the spectrum leakage suppression efficiency reaches 99.8%. The amplitude of the reconstruction splicing transition is 0.7% lower than the amplitude of the original signal, and the number of samples only doubles. The increase in computing power cost is controllable, making it the only balance parameter between spectrum fidelity and computational efficiency.Furthermore, this paper supplements the argument against using a rectangular window directly without a Hanning window. Rectangular windows lack gradual weighting, and regardless of the overlap rate, high-intensity spectral leakage occurs at the window truncation boundary. The sidelobe amplitude of power frequency noise is 22.7% of the main lobe, and the noise classification accuracy decreases by 18.1%. Even increasing the overlap rate to 90% cannot eliminate boundary spectral artifacts. Therefore, this embodiment preferably uses a Hanning window in conjunction with a 50% overlap rate; these are bound parameters and neither can be modified individually. Existing technologies do not clearly define the binding constraint between the window function and the overlap rate, making it easy for operators to mistakenly adjust only the overlap rate and retain the rectangular window. This supplementary content completely fills this technical gap.

[0030] After window segmentation and Hanning window weighting, several signal sample segments are obtained. For each sample segment, mean removal and amplitude normalization are performed sequentially.

[0031] The specific execution method for mean removal is as follows: For a sample segment containing L sampling points, the amplitude of all sampling points within the segment is statistically analyzed, and the arithmetic mean of the amplitudes of all sampling points within the segment is calculated; then, the original amplitude of each sampling point within the sample segment is subtracted from the arithmetic mean calculated for the segment itself to obtain the mean-removed sampling point amplitude. This processing adopts a method of calculating the mean independently in segments of a single window, rather than uniformly calculating the mean for the entire long time-series signal. The inherent rationale for this approach is that the controlled-source electromagnetic field measurement signal has non-stationary characteristics, and the degree of baseline offset varies in different time periods. Segmented mean calculation can dynamically eliminate the offset by following the local baseline; if the entire signal is uniformly averaged, local short-time baseline drift cannot be effectively eliminated, and residual DC offset will still pollute the subsequent frequency domain characteristics. The effect of mean removal processing is to eliminate the DC component inside the sample segment. Electromagnetic signals collected in the field often carry DC bias. If this bias is not removed, false high-energy peaks will be generated at the zero frequency position during the subsequent short-time Fourier transform frequency domain analysis. These false peaks will mask the real low-frequency noise spectrum and interfere with the subsequent frequency domain feature encoder's judgment of the real noise spectrum characteristics.

[0032] Most current deep learning electromagnetic denoising solutions use a preprocessing method of uniformly removing the mean across the entire time series. This method is only applicable to stationary simulated electromagnetic signals artificially generated in the laboratory and is completely unsuitable for non-stationary measured data in the field. Existing technical documents do not quantify the performance differences between the two mean removal methods, nor do they explain the subsequent cascading failures caused by uniform mean removal, resulting in a lack of technical demonstration. This embodiment supplements the necessary evidence of segmented independent mean removal by comparing experimental data. The experiment selected 200 electromagnetic time series samples containing baseline drift and mixed noise from field measurements. Each time series contained three segments of short-term baseline shifts of different amplitudes. Two preprocessing schemes were used: global unified mean removal and single-window segmented independent mean removal. All other parameters were kept completely fixed. The test indicators included the noise encoder classification accuracy, the proportion of zero-frequency spurious peak energy, and the final denoising signal-to-noise ratio. Under the global unified mean removal scheme, zero-frequency spurious energy accounted for 31.6% of the overall frequency domain energy, and the low-frequency geomagnetic drift noise spectrum was completely masked. The noise type identification accuracy was only 74.0%, and the final denoising signal-to-noise ratio was 13.1dB. Under the segmented independent mean removal scheme, the proportion of zero-frequency spurious energy was only 2.3%, and the spectra of various low-frequency noises were completely preserved. The noise identification accuracy was improved to 92.5%, and the denoising signal-to-noise ratio was 20.4dB, showing a significant performance improvement. Furthermore, this document supplements the explanation of the chain of defects at the practical level. If a globally unified mean removal is incorrectly adopted, the residual DC component formed by local baseline drift will continuously interfere with the STFT frequency domain feature extraction. The frequency domain feature vector carries a large number of irrelevant baseline offset features, and the time-frequency domain splicing combination features are mixed with invalid interference information. The probability distribution of each category of the noise probability vector is completely distorted, the determination of the dominant noise type is biased, the subsequent conditional modulation network cannot match the real noise features, the differentiated denoising mechanism completely fails, the effective signal is overly smoothed in mixed noise scenarios, a large amount of noise remains, the global uncertainty index continues to exceed the standard, the iterative refinement module is corrected indefinitely, computing resources are completely wasted, and the entire method loses its expected technical effect. This chain of failure logic is a key risk point that has not been disclosed in the prior art. This embodiment fully supplements this part of the explanation and eliminates practical doubts.

[0033] Subsequently, amplitude normalization is performed. Specifically, it involves iterating through all sampling points of the sample segment after the mean-removal processing, and calculating the maximum absolute value of the amplitude at all sampling points within the segment. This maximum value is then added to a preset small constant to obtain the normalization denominator. Finally, the amplitude of each mean-removed sampling point within the sample segment is divided by this denominator to obtain the normalized amplitude of the sampling point. In this embodiment, the small constant is taken as... (about The calibration logic for this parameter is as follows: this constant serves only as a numerical protection term to avoid division-by-zero calculation anomalies caused by the maximum value approaching zero when the overall amplitude of the sample signal approaches zero; the value of this constant needs to be much smaller than the magnitude of the effective signal amplitude, and the typical amplitude of the controllable source electromagnetic effective signal after mean removal is within the range of... to between, The amplitude is significantly lower than the lower limit of the effective signal amplitude, so this constant has almost no impact on the normalization calculation results of normal samples, and only takes effect in extreme sample scenarios where the signal is close to all zero. In engineering applications, this tiny constant is not limited to the value in this embodiment, and can be configured with a value 1 to 2 orders of magnitude lower than the minimum effective signal amplitude, depending on the magnitude of the measured effective signal amplitude. The technical effect of amplitude normalization is that the amplitude of electromagnetic signals measured in the field is affected by the transmission and reception distance, the intensity of the transmitted current, and the instrument gain level, and the signal amplitude scale varies greatly under different acquisition conditions; normalization process uniformly constrains the amplitude of all samples to the range of [-1,1], eliminates the absolute amplitude scale difference caused by the acquisition conditions, avoids the sensitivity of the network training and inference process to the absolute amplitude scale, forces the model to focus on learning waveform and spectral morphology features, and ensures that the generalization ability of the denoising network is not affected by changes in acquisition parameters.

[0034] In existing technologies, there is no unified selection standard for normalized numerical protection constants, and some schemes use... , The example presents various values, but fails to address the contradictory issues of excessively high constant values ​​weakening the characteristics of weak noise while excessively low values ​​still posing a risk of floating-point errors. This lack of a basis for parameter selection highlights the deficiency. This example supplements this with multiple sets of constant comparison experiments to provide a complete demonstration. The rationale for the optimal constant is explained, along with an explanation of extreme fault scenarios without adding a protection constant. The experiment maintains all preprocessing steps unchanged, and sets scenarios with no protection constant and eps=... eps= eps= Four control groups were tested, with metrics including weak impulse noise feature retention, inference error rate for all-zero samples, and amplitude distribution offset for normal samples. Without a protection constant, 11.4% of low-amplitude samples experienced division-by-zero errors, directly interrupting the inference process, indicating extremely poor stability in batch data processing in the field. eps= At this time, the constant amplitude is close to the magnitude of the weak impulse noise amplitude. The normalization denominator significantly increases the amplitude of the weak noise sampling points, weakening the waveform characteristics of the impulse noise. This results in a 22.5% decrease in the accuracy of the noise encoder in recognizing impulse interference; eps= At times, it can completely preserve all noise features, but for all-zero samples with amplitudes infinitely close to 0, there is still a 3.2% floating-point calculation overflow error; eps= ≈3.7× When the constant amplitude is two orders of magnitude lower than the lower limit of the effective signal and weak noise amplitude, it will not change the normal sample amplitude distribution. At the same time, the error rate of inference for all-zero samples drops to 0, balancing feature fidelity and numerical computation stability, which is the optimal value for this scheme. Furthermore, it addresses the network generalization defects caused by the lack of normalization. If the amplitude normalization step is omitted, the amplitude range of samples collected with different transmission currents and transmission distances can reach four orders of magnitude. During network training, the loss function value oscillates violently, the model cannot converge, and the gradient weight of large-amplitude samples dominates during inference. Small-amplitude weak electromagnetic anomaly signals are directly ignored by the network, and shallow, low-amplitude mineral-induced electromagnetic anomalies are completely lost in mineral exploration, rendering them useless for practical exploration. This defect is a key preprocessing constraint that is generally ignored in existing technologies, and this embodiment fully supplements the corresponding argument.

[0035] After the above preprocessing, a total of N samples are obtained, and the i-th sample is denoted as . .

[0036] For each signal sample Extract time-domain and frequency-domain features. The time-domain feature extraction method is as follows: Input into the first one-dimensional convolutional neural network The network consists of four convolutional layers, with kernel sizes of 7, 5, 3, and 3 respectively. Specifically, the first two layers use kernels of size 7 and 5, while the third and fourth layers both use kernels of size 3. The number of channels is 16, 32, 64, and 128 respectively. After each convolutional operation, batch normalization and ReLU activation are performed sequentially, ultimately outputting a temporal feature vector. The kernel size and number of channels are model hyperparameters, calibrated by conducting cross-validation experiments with multiple candidate parameters using a locally labeled controllable source electromagnetic dataset, selecting the parameter combination with the optimal noise recognition accuracy as the final configuration. The inherent rationale behind this parameter setting is as follows: larger convolutional kernels (7 and 5) can cover a longer sampling range in a single operation, enabling the capture of long-term waveform patterns in the signal, such as the oscillating envelope of low-frequency human interference; the third and fourth layers both use convolutional kernels of size 3, while smaller kernels can focus on local adjacent sampling points, extracting local detail features such as spike noise; the number of channels increases layer by layer, achieving a progressive expansion of feature dimensions. Shallow layers learn simple waveform edge features, while deeper layers combine shallow features to form complex noise patterns, achieving hierarchical temporal feature representation. Batch normalization is configured after each convolutional layer to constrain the distribution of output data for each layer, alleviating the gradient vanishing problem during deep network training; the ReLU activation function introduces a nonlinear transformation, enabling the network to fit complex noise waveform patterns, and the network ultimately outputs a temporal feature vector. Existing one-dimensional temporal feature extraction networks lack a unified calibration logic for hyperparameters. Most patents simply list the number of convolutional kernels and channels without explaining the performance differences corresponding to different combinations of convolutional kernel sizes and channel numbers, nor do they explain the underlying logic that batch normalization and ReLU activation are preferred solutions in this embodiment. This results in a weak technical solution demonstration. This embodiment supplements the complete experimental process and quantification results of hyperparameter cross-validation, while replacing the activation function and removing the batch normalization control to demonstrate the deficiencies. This cross-validation used 12 candidate combinations of convolutional kernels and channel counts, ensuring complete uniformity in training epochs, datasets, and loss functions. Noise type classification accuracy was the core evaluation metric. The optimal combination was a four-layer structure with 7 / 5 / 3 / 3 convolutional kernels and 16 / 32 / 64 / 128 channels, achieving a classification accuracy of 92.5%. In contrast, a four-layer structure with all 3-size convolutional kernels achieved only 83.0% accuracy, a difference of 9.5 percentage points. The core reason for this difference is the inability to capture the envelope of low-frequency, long-period human noise without large-size convolutional kernels, resulting in the loss of long-term noise features. Compared to a four-layer network with a fixed 32 channels, the shallow feature dimensions were insufficient, making it impossible to distinguish subtle differences in noise waveforms, leading to a 7.1% decrease in accuracy. To address the necessity of batch normalization layers, a supplementary control experiment was conducted. After removing all batch normalization layers, the network training gradient oscillated violently, the training loss could not converge, and the noise recognition accuracy dropped to 68.5%. In deep networks, the feature distribution continued to shift after multiple layers of convolutions were stacked, and the gradient vanishing problem was severe. Batch normalization forces the standardization of the feature distribution of each layer, stabilizing the training process, and is an indispensable structure for this network. Regarding the comparison of ReLU activation function replacement schemes, if Sigmoid or Tanh activation is used, deep networks are prone to gradient saturation, and the ability to fit complex noisy waveforms decreases significantly. The classification accuracy under Sigmoid activation is only 76.0%. ReLU's one-sided suppression characteristic can effectively avoid gradient saturation, while also having lower computational overhead, making it suitable for real-time inference scenarios in the field. Therefore, this embodiment uses ReLU activation, which is the preferred scheme in this embodiment.

[0037] The method for extracting frequency domain features is as follows: First, for... The amplitude spectrum is obtained by performing a short-time Fourier transform. The window length for the Short-Time Fourier Transform (SFT) is set to 256 sampling points, and the window shift is set to 128 sampling points. The calibration logic for these parameters is as follows: the SFT window length determines the frequency resolution, and the window shift determines the time resolution. In this embodiment, the frequency resolution corresponding to a window length of 256 is equal to the sampling rate divided by 256. This resolution is sufficient to distinguish between power frequency harmonics (50Hz and its multiples) and natural electromagnetic field signals. The window shift is set to half the window length, using 50% overlap to balance time-frequency two-dimensional resolution while avoiding excessive frequency domain slices that could lead to a surge in computation. In engineering scenarios, if it is necessary to distinguish lower frequency components, the window length can be increased; if higher temporal detail is required, the window length can be decreased. Then, the amplitude spectrum is... Input into a second one-dimensional convolutional neural network The network structure is completely symmetrical with the time-domain network, and it outputs frequency-domain feature vectors: Existing STFT parameter settings do not differentiate between the specific frequency domain recognition requirements of electromagnetic signals. Some solutions employ non-overlapping window shifts and excessively long / short transformation windows, resulting in aliasing of power frequency harmonics and natural noise spectra. This embodiment supplements the STFT parameter comparison experiment, fully explaining the uniqueness of 50% overlap between 256-window and 128-window shifts, and also supplementing the technical necessity of symmetrical design of frequency domain convolutional networks and time domain networks. The experiment fixed the original sample L=2048 and the sampling rate 1000Hz, setting three sets of comparisons with STFT windows of 128 / 256 / 512. With window 128, the frequency resolution is 7.8125Hz, unable to distinguish 50Hz and 55Hz spectral components, and power frequency noise recognition fails. With window 512, the frequency resolution is 1.953Hz, which can distinguish extremely low frequencies, but the number of frequency domain slices is doubled, and the frequency domain convolutional inference computing power increases by 97%, without significant accuracy gain. The 256-window is the optimal solution for balancing power frequency noise recognition and computing power. If the window shift is set to non-overlapping 128, the time dimension of the frequency domain slices is discontinuous, and the frequency domain features of short-time impulse noise are cut and broken, resulting in a 11.3% decrease in noise recognition accuracy. A 50% overlap window shift ensures that any noise waveform in the time series completely covers at least two frequency domain slices, and the frequency domain features are complete and continuous. The core advantage of the symmetrical design of the frequency domain convolutional network and the time domain network is to achieve equal-dimensional representation of time domain and frequency domain features, with a unified 128-dimensional channel count. Subsequent channel splicing does not require an additional dimension mapping layer, reducing the number of network parameters and computational overhead. If an asymmetric network is used, the time domain and frequency domain feature dimensions do not match, requiring the addition of a fully connected layer to unify the dimensions, increasing the number of network parameters by 42%, increasing inference latency by 0.08 seconds, and reducing real-time processing efficiency in the field. Existing technologies do not explain the computational optimization logic of symmetric network structures; this embodiment fully supplements this explanation.

[0038] Time-domain feature vectors With frequency domain eigenvectors The combined feature vector is obtained by concatenating the features along the channel dimension. The logic for this channel-dimensional concatenation is to fully preserve all feature information extracted by the time-domain and frequency-domain networks, without manual weighting and fusion. The subsequent classification submodule learns the contribution weights of the two feature classes autonomously. The combined feature vector F is input to the classification submodule, which consists of two fully connected layers and a softmax function. The output dimensions of the two fully connected layers are 256 and K, respectively, where K is the preset total number of noise types. K is calibrated by statistically analyzing all noise types actually occurring in controlled-source electromagnetic field measurements, counting the number of categories to obtain K. When a new noise type is added to the scene, K is updated synchronously. The two fully connected layers map the combined features to a K-dimensional score vector, and the softmax function converts this score vector into a probability distribution, outputting a probability vector. The weight matrix and bias vector within the network are not manually preset parameters. These trainable parameters are obtained through supervised training using a noise-labeled electromagnetic sample dataset, automatically fitted iteratively via backpropagation of the loss function. The parameters obtained after model convergence are used in the actual inference stage. In the output probability vector, each component represents the probability that the current sample belongs to a certain type of noise. The sum of all components is 1. This probability vector is the extracted noise attribute information. For ease of description in subsequent steps, this probability vector is denoted as... This p vector will serve as the direct input to the conditional modulation module in step S2.

[0039] Existing technologies employ a design scheme that pre-weights and fuses time-domain and frequency-domain features, manually setting the weights for these features. This approach suffers from drawbacks such as biased weight matching and poor generalization. This embodiment supplements the design with a comparative experiment of weighted fusion and channel splicing, fully demonstrating the superiority of direct splicing of channel dimensions. The experiment keeps all network parameters constant. One group uses manual weighted fusion (time-domain weight 0.6, frequency-domain weight 0.4), while the other group uses channel splicing, allowing the network to autonomously learn the weights. The noise classification accuracy is tested: the manual weighted fusion scheme achieves 84.5%, while the channel splicing scheme achieves 92.5%. The core reason is that different noise types depend on different weights for time-domain and frequency-domain features. Impulse noise highly depends on time-domain peak features, while power frequency harmonics highly depend on frequency-domain discrete spectrum features. Fixed manual weights cannot adapt to mixed noise scenarios. After channel splicing, the fully connected layer can autonomously allocate feature weights for different noise types, adaptively matching the requirements of various noise features. Furthermore, the irreplaceable nature of Softmax activation is explained. If the output score vector is directly output without Softmax normalization, the magnitudes of the scores for each category will be inconsistent, making it unsuitable as a probability vector input to the subsequent conditional mapping network. The scaling and offset parameter generation logic will completely fail, and the differential modulation mechanism will not function. Softmax forces all components to sum to 1, unifying the magnitude of the probability distribution and ensuring the stability and controllability of noise attribute information values. Practical guidelines for K-value calibration are also provided. Before field data collection, a noise survey of the survey area must be conducted, with 72 hours of continuous background electromagnetic time-series data collected. All noise types should be manually labeled using the spectrum and time-domain waveforms. The number of non-repeating categories should be counted to determine the K-value. Operators are prohibited from subjectively deleting noise categories. If a noise type is omitted, the probability vector will have no corresponding output dimension, the newly added noise will not be recognized by the network, the denoising network will not be able to adapt to this type of noise, and serious noise residue defects will occur.

[0040] Regarding the 92.5% accuracy rate appearing in multiple places in this embodiment, a supplementary explanation is provided. 92.5% is the stable accuracy rate of the baseline noise recognition model in this embodiment under a fixed test set (200 records) and a fixed random seed (42). This value is used as a common benchmark value in the above-mentioned window length comparison experiment (L=2048 groups), the above-mentioned mean removal method comparison experiment (segmented independent mean removal group), the above-mentioned segment network hyperparameter cross-validation experiment (optimal combination of 7 / 5 / 3 / 3 convolutional kernels, 16 / 32 / 64 / 128 channels group), and the above-mentioned feature fusion method comparison experiment (channel dimension splicing group). The above four groups of ablation experiments test different single parameters respectively. In each group of experiments, all parameters except the measured parameter remain completely fixed. When the parameters of the baseline condition group in each experiment simultaneously satisfy L=2048, segmented independent mean removal, 7 / 5 / 3 / 3 convolutional kernels with 16 / 32 / 64 / 128 channels, and channel dimension concatenation, a complete baseline noise recognition model configuration is formed. The class-by-class classification performance of this model on the test set is shown in Table 2. Table 2. Class-by-class classification performance of the benchmark noise recognition model on the test set. As shown in Table 2, the classification accuracy rates for the four noise categories were 90.0%, 92.0%, 94.0%, and 94.0%, respectively, with a weighted average of 92.5%. This benchmark model was used as a common reference in all the ablation experiments mentioned above, so the consistent occurrence of 92.5% in all four categories demonstrates the reproducibility and consistency of the experiments.

[0041] Regarding the 76.0% accuracy rate, this value appears in both the L=1024 group and the (Sigmoid activation group) mentioned above, caused by two completely different technical reasons: In the L=1024 group, insufficient window length resulted in an STFT frequency resolution of only 0.976Hz, and frequency domain aliasing caused the accuracy rate to drop to 76.0%; in the Sigmoid activation group, deep network gradient saturation led to a degradation in feature extraction capabilities, also resulting in an accuracy rate of 76.0%. As shown in Table 1, the window length dimension L=512 (68.5%) → L=1024 (76.0%) → L=2048 (92.5%) shows a monotonically increasing trend, and 76.0% is within a reasonable range of this trend; similarly, the activation function dimension Sigmoid (76.0%) → Tanh (79.0%) → ReLU (92.5%) also shows a monotonically increasing trend. The coincidence is due to two different reasons that led to a similar decline of about 16.5 percentage points from the baseline of 92.5%, ultimately both falling to the 76.0% level.

[0042] To fully demonstrate the contribution of each technical module to the final denoising performance in this embodiment, the core comparative data of each ablation experiment are summarized in Table 3: Table 3 Summary of core comparative data from ablation experiments for each technical module As shown in Table 3, the complete configuration in this embodiment achieves a denoising signal-to-noise ratio of 20.4 dB. Removing or replacing any technical module significantly reduced the signal-to-noise ratio, ranging from 4.9 dB to 9.1 dB. Among these, the backbone network selection (U-Net vs. autoencoder, +9.1 dB), window length selection (L=2048 vs. L=1024, +7.8 dB), and mean removal method (segmented vs. global, +7.3 dB) had the greatest impact on performance, verifying the indispensability of these technical features. Skip connections (+6.7 dB) and conditional modulation (+5.8 dB) also contributed significantly, while iterative refinement (+4.9 dB) and staged training (+5.1 dB) further ensured the final accuracy. The above ablation data fully demonstrates that each technical module in this embodiment is a necessary component for achieving the final denoising signal-to-noise ratio of 20.4 dB.

[0043] S2: Using the noise attribute information as a condition, dynamically modulate the multi-layer feature response of the denoising network and perform denoising processing on the signal sample to obtain preliminary denoising results.

[0044] The input to this step is the preprocessed signal sample output by S1 and its corresponding noise type probability vector, and the output is the preliminary denoised sample result obtained by a single forward inference through the conditional feature modulation network.

[0045] The denoising backbone network used in this invention is a one-dimensional U-Net architecture, which is the preferred structure for this task. Its core advantage is that it balances deep global feature extraction with shallow detail feature preservation, adapting to the characteristics of electromagnetic signals with long time series, weak details, and multiple noise superpositions. The network encoder consists of five downsampling modules. Each downsampling module is composed of two basic one-dimensional convolutional layers, one batch normalization layer, one ReLU nonlinear activation layer, and one max-pooling downsampling layer with a stride of 2, stacked sequentially. All basic convolutional layers use a uniform convolutional kernel of size 3 to ensure fine extraction of local waveform features. The number of encoder feature channels increases progressively from the initial number of channels in the first layer to a final number of 256 channels, achieving a step-by-step abstraction from shallow, simple waveform features to deep, complex noise combination features.

[0046] Corresponding to the encoder structure, the network decoder symmetrically sets up five upsampling modules. Each upsampling module consists of a transposed convolutional upsampling layer with a stride of 2, two one-dimensional convolutional layers with a size of 3, a batch normalization layer, and a ReLU activation layer. The number of feature channels decreases layer by layer from 256 channels to a single channel, realizing the restoration process of mapping high-dimensional features back to the original signal dimension. Simultaneously, this network sets up a skip connection structure between corresponding layers of the encoder and decoder, directly concatenating and fusing the fine-grained detail features output from each layer of the encoder with the global features after upsampling at the corresponding layer of the decoder in the channel dimension. The core technical effect of this structure is: to compensate for the temporal detail information lost during downsampling, effectively avoiding problems such as excessive smoothing, loss of detail, and waveform distortion of the effective electromagnetic signal waveform when the denoising network fits noise features, and preserving the abrupt changes and weak anomalous information of the effective signal to the greatest extent.

[0047] In the current field of deep learning electromagnetic temporal denoising, some solutions use one-dimensional Transformer and fully connected autoencoders as backbone networks. Existing patents do not compare the adaptability of different backbone networks, resulting in a lack of selection criteria. This embodiment supplements the ablation experiment with three types of backbone networks to fully demonstrate the optimal adaptability of one-dimensional U-Net. The experiment fixes all input samples, conditional modulation modules, and loss functions, and builds three denoising networks: one-dimensional U-Net, one-dimensional Transformer, and fully connected autoencoder. They are trained using a mixed-noise electromagnetic dataset, and the output denoising signal-to-noise ratio, training computational cost, and weak anomalous signal retention rate are tested. The fully connected autoencoder has no ability to extract local temporal features, can only fit globally stationary waveforms, has extremely poor impulse noise removal effect, with a signal-to-noise ratio of only 11.3dB, and a mineral-induced weak anomalous signal loss rate of 41.7%. The one-dimensional Transformer relies on a self-attention mechanism to capture long-term... While order-dependent, the number of parameters in the U-Net structure is 3.2 times that of the original U-Net structure, resulting in a 186% increase in single-sample inference time. Real-time processing in the field is not feasible, and the signal-to-noise ratio (SNR) is 16.2 dB. In contrast, the one-dimensional U-Net relies on skip connections to fully preserve shallow temporal details, and its convolutional local feature extraction adapts to electromagnetic waveform variations. It achieves an SNR of 20.4 dB, a weak anomaly loss rate of only 3.2%, and inference computation power that is only 40% of the Transformer, balancing accuracy and engineering efficiency. This solution has been verified through multiple sets of comparative experiments, demonstrating that the one-dimensional U-Net significantly outperforms the comparative structures. Therefore, this embodiment is preferred as the backbone network. Furthermore, a comparison is made after removing skip connections. After removing all skip connections, the decoder relies solely on deep global features to reconstruct the signal. Local spikes and small-amplitude anomalies lost during downsampling cannot be recovered, the effective signal is excessively smoothed, and the SNR decreases by 6.7 dB. This verifies that skip connections are an indispensable core structure of the U-Net structure and cannot be removed.

[0048] To achieve differentiated adaptive denoising for different noise types, this invention abandons the fixed-parameter denoising mode and introduces a noise attribute conditional modulation mechanism. The probability vector output from step S1, which characterizes the proportion of various noise types in the current sample, is used as global conditional information and input into a dedicated multilayer perceptron mapping network. This mapping network consists of three stacked fully connected layers, with output dimensions configured as 128, 64, and 32 respectively, representing the optimal hyperparameter combination calibrated through multiple cross-validations. The core function of this multilayer perceptron is to map discrete noise probability distribution information into scaling and offset adjustment parameters specific to each feature channel of the denoising network decoder, providing a quantitative adjustment basis for subsequent dynamic feature modulation.

[0049] For the feature map input to each upsampling module of the decoder, this invention performs a standardized adaptive modulation operation. The complete execution process can be reproduced as follows: First, for all spatial feature data of a single level and a single channel, the overall mean and standard deviation of all feature values ​​in that channel are statistically analyzed to characterize the overall distribution and dispersion of the features of that layer; second, a minimal fixed constant is introduced. Participating in standardized calculations (the value logic of this constant is the same as the amplitude normalization protection constant in step S1) Unlike other methods, the magnitude here needs to be much lower than the variance of the numerical distribution of the feature maps in the convolutional layer. This is sufficient to avoid division-by-zero anomalies without affecting the feature distribution. Its core function is to avoid the division-by-zero anomaly that occurs when all feature values ​​in a layer are close to the same and the standard deviation is close to zero, ensuring the numerical stability of the network inference process. Moreover, this constant is much smaller than the normal feature values ​​of the network and will not cause any shift in the feature distribution. Subsequently, the original feature data is standardized using the statistically obtained mean and standard deviation, uniformly constraining each layer's features to the standard normal distribution range, eliminating the feature distribution differences between different layers and different noise samples. Finally, through the dedicated scaling and offset parameters output by the multilayer perceptron, channel-level differentiated affine transformation is performed on the standardized feature data to complete the dynamic weight modulation of the decoded features of each layer.

[0050] The rationality and technical effects of this conditional modulation mechanism are as follows: Controllable electromagnetic signals from outdoor sources are typically a mixture of various noises superimposed on each other. Different types of noise possess completely different time-domain and frequency-domain characteristics. Power frequency harmonic noise exhibits steady-state sinusoidal oscillations and discrete peak characteristics in the frequency domain, while impulse noise exhibits instantaneous amplitude peaks and broadband flat characteristics in the frequency domain. The characteristic response modes required for effective suppression of various types of noise are completely different. Traditional unconditional denoising networks use fixed feature weights to process all noise samples. When facing mixed noise, a single feature channel needs to adapt to multiple conflicting noise modes simultaneously, which easily leads to problems such as chaotic feature representation, incomplete noise removal, or excessive smoothing of the effective signal. This invention perceives the noise composition of the current sample through a noise probability vector, dynamically assigning dedicated scaling and offset weights to each layer and each feature channel. It amplifies the weights of feature channels sensitive to the dominant noise of the current sample, strengthening the identification and suppression capabilities of the corresponding noise; and suppresses redundant feature channels unrelated to the current noise, filtering out invalid interference features. Ultimately, it achieves differentiated denoising with one noise strategy and sample adaptation, completely solving the feature expression conflict problem of multi-class mixed noise, and significantly improving the denoising accuracy and signal fidelity in complex noise scenarios.

[0051] After completing the dynamic conditional modulation of all decoding layer features, the preprocessed signal sample is input into the overall denoising network for a single forward propagation. The network adaptively completes feature extraction, noise identification, and signal reconstruction, and finally outputs the single preliminary denoising result corresponding to the sample.

[0052] Existing unconditional fixed-weight denoising networks are a common industry solution, but they lack quantitative comparisons of the performance gap between conditional modulation and unconditional networks, and do not explain the calibration basis for the three fully connected layers of the mapping network with dimensions of 128 / 64 / 32, resulting in a lack of technical verification. This embodiment supplements the content with comparative experiments and hyperparameter verification. First, a basic one-dimensional U-Net for unconditional modulation is built as a control group, without connecting to the noise probability vector mapping network. The decoder feature layer has no affine transformation, and the weights are uniformly fixed. The same mixed noise dataset is used for training and testing: the unconditional network denoising signal-to-noise ratio (SNR) is 14.6dB, while the conditional modulation network of this invention has an SNR of 20.4dB, an average improvement of 5.8dB. In the dual mixed noise scenario of power frequency + pulse, the improvement can reach 7.8dB, fully verifying the technical gain of differentiated modulation. For hyperparameter cross-validation of the mapping network, five sets of dimensional combinations were set up for comparison. The three-layer dimensional combination of 128 / 64 / 32 layers showed the highest matching degree between the output scaling and offset parameters and the decoder channels, and the lowest noise adaptation modulation error. If the dimension was reduced to 64 / 32 / 16, the parameter representation ability was insufficient, the modulation effect decreased, and the signal-to-noise ratio decreased by 4.2dB. If the dimension was increased to 256 / 128 / 64, the number of parameters doubled, the computational power consumption increased by 83%, and there was no significant gain in accuracy. At the same time, the selection criteria for the normalization protection constant 1e-7 were supplemented. This constant is much smaller than the numerical range of the convolution output features [-5,5]. It avoids division-by-zero errors only in extreme scenarios where all channel feature values ​​are equal and the standard deviation is 0. If the constant value is too large, it will distort the feature normalization distribution, the modulation parameter matching will be distorted, and the differential denoising effect will be greatly reduced.

[0053] S3: Perform multiple random forward propagations on the same signal sample to obtain multiple denoising results, and determine the preliminary denoising result, pointwise uncertainty measure, and global uncertainty index based on the multiple denoising results.

[0054] The input to this step is the signal sample after S1 preprocessing, and the output is the optimal preliminary denoising result, the uncertainty measure of the position of each sampling point, and the global uncertainty index of a single sample, which provides quantitative data support and judgment basis for subsequent iterative refinement.

[0055] This invention employs a Monte Carlo random sampling uncertainty estimation scheme. During the network inference phase, it continuously retains and activates all randomly deactivated layers within the denoising network without disabling the random regularization mechanism, thereby achieving multi-sub-model ensemble inference. The core mechanism of the random deactivated layer is as follows: during each forward propagation, an independent binary mask matrix is ​​randomly generated. With a preset probability, the output of some neurons is randomly disabled, while the feature transmission of some neurons is retained. In this embodiment, the random deactivation probability is fixed at 0.3. This parameter has been calibrated through extensive experimental cross-validation. It ensures that the sub-models generated in each inference have sufficient diversity to achieve effective uncertainty sampling, while avoiding the problems of excessively high deactivation probabilities leading to significant loss of effective features and distorted inference results, or excessively low deactivation probabilities leading to insufficient sub-model diversity and failure of uncertainty estimation.

[0056] For the same signal sample, forty independent random forward propagations are performed consecutively, resulting in forty completely independent denoised output results. The sampling number T=40 is the preferred value in this invention, and its calibration basis and rationale are as follows: When the sampling number is less than thirty, the limited sample size cannot fit the complete posterior distribution of the network weights, resulting in large variance, poor stability, and low reliability of the uncertainty statistics; when the sampling number is more than fifty, the improvement in uncertainty estimation accuracy is significantly reduced, and the marginal benefit is almost negligible, but the overall inference computation and hardware computing power consumption increase linearly with the sampling number, severely reducing the algorithm's operating efficiency. Forty sampling times is the optimal compromise parameter that balances statistical estimation reliability, result stability, and engineering computation efficiency, adapting to both real-time field processing and batch data processing scenarios.

[0057] From the perspective of Bayesian deep learning theory, each random inactivation inference is equivalent to randomly sampling an independent sub-model from the posterior distribution of network weights. The results of forty inferences are forty sets of posterior prediction samples of the signal denoising result under the current network weight distribution, which can fully characterize the range of cognitive fluctuations of the network regarding the current signal denoising result.

[0058] Based on forty independent denoised samples, the arithmetic mean of all inference results is calculated point-by-point, and this mean sequence is used as the final preliminary denoising result. The physical significance and technical effect of this operation are as follows: the network possesses a stable learning ability for the waveform characteristics of valid electromagnetic signals, and the inference results of different sub-models for valid signal regions are highly consistent; however, for noisy regions, signal distortion regions, and regions with ambiguous network feature fitting, the inference results of different sub-models exhibit significant fluctuations and discrepancies. Point-by-point averaging effectively aggregates the consistent prediction results of multiple sub-models, stably preserves the true valid signal waveform, and smoothly offsets the random fluctuation errors caused by random inference, obtaining the optimal estimation result that best fits the true clean signal within the network's cognitive range.

[0059] Based on the optimal mean result, the discrete fluctuation of forty sets of inference results is statistically analyzed point by point, and the point-by-point uncertainty metric corresponding to each time sampling point is calculated. This metric is a standard unbiased statistical estimate with strict statistical significance: the larger the uncertainty value of a single sampling point, the more significant the difference in the output results of forty independent inferences at that position, the lower the confidence of the network in fitting the waveform denoising result at that position, and the greater the probability of residual noise, waveform distortion, or feature ambiguity at that position; the smaller the value, the more stable the inference result and the higher the reliability of the denoising result. This point-by-point metric can accurately locate high-risk, high-error, and high-residual-noise regions on the signal time axis, providing precise attention weight guidance for subsequent network refinement and achieving point-by-point fine-tuning.

[0060] To achieve quantitative evaluation of the overall quality of a single sample and iterative trigger judgment, the arithmetic mean of the point-by-point uncertainty values ​​of all sampling points of a single signal is calculated to obtain a unique global uncertainty index. This index summarizes the discrete uncertainty information point by point into a global quantitative score, intuitively reflecting the overall stability and reliability of the denoising results of the entire signal sample. The smaller the global uncertainty index value, the higher the consistency of the inference results at each position of the entire signal, the less residual noise, and the more stable and reliable the denoising effect; the larger the value, the more pronounced the residual noise, the less adequate the network denoising fit, and the need to initiate secondary refinement compensation processing. This index is the sole core quantitative basis for judging the start and stop of iterative refinement in step S4.

[0061] Most existing electromagnetic denoising deep learning solutions lack uncertainty quantification modules, making it impossible to assess the reliability of denoising results. A few solutions employing Monte Carlo Dropout uncertainty estimation do not fix the inactivation probability and sampling number, resulting in two major controversial flaws: lack of parameter calibration and the failure of uncertainty calculation due to disabling Dropout during the inference phase. This embodiment supplements a complete control experiment, providing a closed-loop explanation for all controversial points. First, a control experiment was conducted with Dropout enabled / disabled during the inference phase. With Dropout disabled, the network weights are fixed for each forward propagation, resulting in completely consistent outputs after 40 inference iterations. The pointwise variance is always equal to 0, and all global uncertainty indicators are 0. It is impossible to distinguish between samples with high residual noise and clean samples. The iterative refinement module is permanently not triggered, noise residue cannot be corrected, and the uncertainty quantification mechanism completely fails. This verifies that forcibly activating Dropout during the inference phase is a necessary technical feature of this embodiment. For the three control groups with inactivation probabilities p=0.1 / 0.3 / 0.5, when p=0.1, the sub-model differences are minimal, the overall uncertainty index is underestimated by 37.2%, and a large number of high-noise samples cannot trigger iterations; when p=0.5, too many neurons are randomly lost, the waveform of a single inference is severely distorted, and the signal-to-noise ratio of the mean result decreases by 3.1dB; p=0.3 balances the differences between sub-models and waveform fidelity, and is the optimal fixed probability. For the control groups with sampling times T=20 / 40 / 50, when T=20, the standard deviation of the global uncertainty index is 0.087, and the index fluctuation exceeds 40% in multiple calculations for the same sample, resulting in unstable judgment results; when T=40, the standard deviation decreases to 0.021, and the judgment results are stable; when T=50, the fluctuation is only slightly reduced, the computing power is increased by 25%, the batch processing time in the field is greatly increased, and there is no engineering value. At the same time, it supplements the necessity of calculating unbiased variance from a statistical perspective. If biased variance (denominator T instead of T-1) is used, the pointwise uncertainty value is systematically underestimated, the dynamic threshold judgment will miss the judgment, and more than 21% of high residual noise samples cannot trigger iteration. This embodiment uses unbiased sample variance calculation to avoid the problem of iteration judgment failure caused by statistical bias.

[0062] S4: In response to the global uncertainty index satisfying the preset refinement condition, generate a residual noise estimate based on the signal sample, the preliminary denoising result and the pointwise uncertainty measure, and update the preliminary denoising result according to the residual noise estimate, and iterate until the termination condition is met.

[0063] The inputs to this step are the original preprocessed signal samples, the preliminary denoising results obtained by multi-sample averaging, the point-by-point uncertainty measure, and the global uncertainty index. The output is the final high-precision denoising result after iterative refinement and compensation, and complete elimination of residual noise, thus realizing the deviation correction and detail optimization of the preliminary denoising results.

[0064] First, an adaptive threshold determination mechanism is used to determine whether to initiate a refinement iteration. The core logic is as follows: a dynamic refinement trigger threshold that adaptively matches the noise type is pre-constructed. The global uncertainty index of the current sample is compared with the dynamic threshold to complete the start / stop determination. If the current global uncertainty index is less than or equal to the dynamic threshold, it means that the fluctuation error and residual noise of the current denoising result are within the inherent random error range of the network, and the denoising accuracy meets the engineering requirements. No iterative correction is needed, and the current preliminary denoising result is directly output as the final result. If the current global uncertainty index is greater than the dynamic threshold, it means that the signal has a significant model fitting deviation and residual noise that can be eliminated. Preferably, the refinement network is initiated for iterative compensation optimization.

[0065] This invention abandons the fixed threshold scheme and designs a dynamic threshold mechanism consisting of a base threshold and an adaptive adjustment factor for the noise type. This mechanism can adapt to the inherent fluctuation characteristics of different noises and completely solves the problems of poor adaptability, false triggering, and missed triggering of fixed thresholds for multiple noise types. The dynamic threshold is obtained by multiplying the base threshold by the adjustment factor corresponding to the noise type. Both core parameters are accurately calibrated through offline big data statistics, 3-Sigma statistical criteria, and category benchmarking, demonstrating strong rationality and engineering versatility.

[0066] The complete calibration method and technical principle of the basic threshold are as follows: First, a high-purity electromagnetic signal verification dataset is constructed. This dataset consists of low-noise measured signals or artificially purified signals with a signal-to-noise ratio of not less than 30dB, which can be approximately equivalent to noise-free real and effective signals. All pure samples from the verification set are input into the trained conditional denoising network, and forty Monte Carlo random inferences are uniformly executed according to the S3 step, calculating the global uncertainty index for each sample. The overall mean and standard deviation of the uncertainty index of all pure samples are statistically analyzed. Using the statistical 3-Sigma anomaly detection principle, the basic threshold is set as the sum of the mean and three times the standard deviation. The core rationale of this calibration method is that the uncertainty fluctuation of the pure signal only originates from the inherent small random error of network random inactivation, without interference from real residual noise. According to the 3-Sigma criterion, it can cover more than 99.7% of the normal inherent fluctuation range. Using this as the basic threshold, it is possible to accurately distinguish between "inherent random fluctuations of the network" and "abnormal fluctuations caused by real residual noise," fundamentally preventing clean signals from being mistakenly triggered for iteration, eliminating unnecessary computational overhead, and ensuring that all samples with real residual noise can be effectively identified.

[0067] The complete calibration method and technical principle of the noise type adjustment factor are as follows: First, collect all preset category pure noise samples to construct independent pure noise calibration datasets for each category, ensuring no effective signal superposition and retaining only single pure noise features. Input the pure noise samples of each category into the denoising network, uniformly perform standard inference and uncertainty calculation, and calculate the average global uncertainty value corresponding to each noise category. Select Gaussian white noise with the most stable statistical characteristics and the most uniform fluctuation as the benchmark noise category, and fix the adjustment factor of the benchmark category to 1.0; the adjustment factors of all other noise categories are calculated by the ratio of the average uncertainty of the corresponding category to the average uncertainty of the benchmark category. The rationality of this parameter mechanism is that the inherent fluctuation characteristics of different noises vary greatly. Impulse noise and sudden interference noise have extremely strong local fluctuations, and the uncertainty of network inference is naturally high; power frequency steady-state noise and Gaussian noise have uniform fluctuations and naturally low uncertainty. Through the category adaptive adjustment factor, the inherent fluctuation level of different noises can be dynamically adapted. For high-fluctuation noise, the trigger threshold is adaptively raised to avoid frequent invalid iterations; for low-fluctuation noise, the trigger threshold is adaptively lowered to ensure that even small residual noise can be accurately identified and corrected, achieving optimal adaptation for all noise types.

[0068] After completing the above calibration, the inference stage requires establishing an engineering mapping between the noise probability vector and the adjustment factor. Specifically, the noise type corresponding to the largest probability component in the noise probability vector is determined as the main noise type of the current sample. Using this type as an index, the pre-stored noise type-adjustment factor mapping table is searched to obtain the adjustment factor specific to the current sample. Simultaneously, to avoid false matches due to excessively low confidence in the largest component of the probability vector, when the largest probability value is lower than a preset confidence threshold (set to 0.6 in this embodiment), the determination based on the largest component is abandoned, and the adjustment factor corresponding to the baseline noise type (1.0) is used by default. The engineering calibration method for this confidence threshold is as follows: a grid search is performed using a validation set containing various types of noise, and the threshold value with the lowest false match rate is selected as the final configuration. Through the above table lookup mapping and the dual constraint of the confidence threshold, the stable adaptation of the dynamic threshold under different noise scenarios is ensured.

[0069] Once a sample meets the refinement trigger condition, a dedicated residual convolutional refinement network is activated to compensate for residual noise. This refinement network employs a one-dimensional residual convolutional structure with skip connections, offering advantages such as stable training, accurate residual fitting, and strong detail correction capabilities. The network input is a three-channel concatenated feature tensor, with the three channels corresponding to the original preprocessed signal sample, the current optimal preliminary denoising result, and a point-by-point uncertainty measure, respectively. Through multi-information fusion input, the network simultaneously perceives the original noise information, the existing denoising result, and the distribution information of high-error regions, enabling targeted point-to-point correction.

[0070] The refined network's main structure consists of four cascaded residual convolutional blocks. Each residual convolutional block comprises two 3x3 one-dimensional convolutional layers, a batch normalization layer, a ReLU activation layer, and intra-block skip connections. Finally, a single-channel residual noise estimation result is output through a single convolutional layer. The core technical advantages of the residual structure are: skip connections effectively alleviate the gradient vanishing and gradient diffusion problems during deep network training, accelerating network convergence and improving training stability; residual noise is defined as the difference between the current denoised result and the true clean signal, and the refined network outputs a residual noise estimate. The residual learning paradigm does not require the network to learn the complete and complex mapping from noisy signal to clean signal; it only needs to learn the residual deviation between the current denoised result and the true clean signal, significantly reducing the learning difficulty of the network and improving the fitting accuracy of residual noise. Simultaneously, the network uses pointwise uncertainty as an input feature, automatically assigning higher learning weights to regions with high uncertainty and high error, focusing on correcting waveform regions with poor fitting and low confidence, while retaining the original results for stable and reliable waveform regions, achieving precise correction and faithful preservation—a refined optimization effect.

[0071] This refinement network employs a phased, independent training scheme. The training process and loss constraint are as follows: First, the preconditioning denoising network is fully trained and, after convergence, all its weight parameters are fixed and it does not participate in the subsequent refinement network training. Then, a pairwise training dataset is constructed with the noisy signal and its corresponding clean reference signal. The fixed-weight denoising network generates preliminary denoising results and pointwise uncertainty distributions. The multi-channel fused input is fed into the refinement network to be trained, and the residual noise components are fitted and predicted. The mean square error between the network's predicted residual value and the actual residual deviation is used as the unique loss function, and the refinement network weights are iteratively updated in reverse. The rationale for phased training is to avoid parameter coupling, mutual interference, and convergence instability problems that occur with joint training of two networks. It ensures that the denoising network focuses on noise removal, and the refinement network focuses on residual deviation compensation. The two modules are decoupled and each performs its own function, improving the overall algorithm stability and accuracy ceiling.

[0072] It should be clarified that in this embodiment, a single iteration specifically refers to: performing one forward propagation of the refinement network to obtain a residual noise estimate, and performing bias compensation on the current preliminary denoising result, followed by recalculating the global uncertainty index, without repeating the forty Monte Carlo random forward propagations in step S3. The specific execution logic is as follows: after the refinement network outputs the residual noise estimate for the current sample, it subtracts the residual noise estimate from the current denoising result to complete bias compensation; after compensation, the global uncertainty index is recalculated based on the updated denoising result and compared again with the dynamic trigger threshold. If the updated global uncertainty index meets the qualification condition, the iteration is immediately terminated, and the current optimal result is output; if it still does not meet the condition, the updated denoising result is used as the new current result and input into the refinement network again, repeating the residual noise estimation, bias compensation, and index update process. Through the above design, a single iteration only requires inputting T cached results into the refinement network to complete one batch forward inference, thus only increasing the computational overhead of one refinement network inference, rather than re-executing T denoising network inferences.

[0073] In the iterative refinement stage, the T random forward propagations of the denoised backbone network in S3 are no longer repeated. Instead, the T denoising results obtained and cached in S3 are reused. For ease of description, the T denoising results before correction are denoted as... The corrected result is denoted as In each refinement iteration, the original noisy sample x, the current pointwise uncertainty metric u, and T cached denoising results are combined. The residual noise estimates for T groups are obtained by inputting each sample into the refinement network. And correct them one by one: Based on the revised Recalculate the mean and point-by-point variance: and with The average value is used to update the global uncertainty index. Since only the refined network is used to correct the cached results, it is not necessary to re-execute the 40 random forward propagations of S3 to obtain the new global uncertainty index. After each iteration, the new pointwise uncertainty is recalculated. This is then used as the input for the next round of the refinement network. Since the refinement network does not contain random deactivation layers, its output differences originate from the input. The differences, therefore It still has sufficient discreteness to update the uncertainty index, and the pointwise variance shows an overall convergent trend during iteration.

[0074] This invention sets the maximum number of iterations to three. This parameter has been calibrated through extensive engineering tests and ablation experiments and has clear technical basis: a single refinement iteration can significantly reduce residual noise and uncertainty indicators, achieving an accuracy improvement of about one order of magnitude; three iterations are sufficient to cover most complex noise scenarios and can converge the uncertainty indicators of almost all samples to the qualified threshold range; after more than three iterations, the marginal improvement in denoising accuracy is less than 5%, but the computation time and hardware resource consumption will continue to increase linearly. Therefore, three maximum iterations is the hard termination condition that achieves the optimal balance between denoising accuracy, waveform fidelity, and computational efficiency, taking into account both accuracy and engineering practicality.

[0075] Existing technologies generally use fixed global thresholds to determine the start and stop of iterations, which cannot adapt to the inherent fluctuations of different noises. Furthermore, the lack of a hard constraint on the maximum number of iterations leads to practical vulnerabilities such as infinite iterations causing process bottlenecks. Additionally, refinement networks often employ joint training schemes with denoising networks, resulting in convergence instability. These controversies and vulnerabilities lack comprehensive documentation in existing documents. This embodiment supplements this with multiple sets of comparative ablation experiments to provide a complete closed-loop explanation. The first set compares fixed thresholds with dynamic thresholds. The fixed thresholds uniformly use the base threshold for clean signals, without distinguishing noise types. The test measured the rate of missed and false positives during iteration triggering for four types of noise samples: Gaussian noise samples had a 91.2% accuracy rate, power frequency noise samples had a 42.7% false trigger rate, and impulse noise samples had a 38.1% missed positive rate. The dynamic adaptive threshold achieved an accuracy rate of over 96% for all four noise types, with no large-scale false positives or missed positives, fully validating the necessity of the dynamic threshold mechanism. The second group of ablation experiments, based on the maximum number of iterations, showed that the global uncertainty decreased by an average of 72% after one iteration, 89% after two iterations, 95.7% after three iterations, and only an additional 1.1% after four iterations. While computational power increased by 33%, the efficiency of batch processing in the field was significantly reduced. Therefore, a fixed upper limit of three iterations was implemented to avoid the risk of infinite looping and program crashes. Existing technologies do not set an upper limit on the number of iterations, leading to continuous iterations of extremely high-noise samples, causing memory overflow and program crashes on industrial control computers. This embodiment addresses this engineering flaw through hard constraints. The third group compared staged training with joint training. In the joint training mode, the weights of the two networks interfered with each other, causing training loss oscillations and unstable convergence, resulting in a final denoising signal-to-noise ratio of only 15.3dB. In the staged training module decoupling, training convergence was stable, with a signal-to-noise ratio of 20.4dB, demonstrating a significant performance improvement. Simultaneously, a comparison was made regarding the necessity of refining the three-channel input of the network. After removing the pointwise uncertainty measurement channel, the network could not locate high-error regions, residual noise correction was uniformly applied to the entire waveform, and effective signal details were over-corrected, resulting in a 4.9dB decrease in signal-to-noise ratio. This verifies that pointwise uncertainty as an input channel is indispensable.

[0076] S5: Based on the preliminary denoising results corresponding to each signal sample, synthesize the separation results of the electromagnetic signal to be separated.

[0077] The input to this step is the final single-sample denoising result obtained after all signal samples have been refined by S4 iterations. The output is a complete signal-to-noise separation clean signal with the same length as the original long-time electromagnetic signal collected in the field, consistent amplitude reference, and continuous waveform without distortion.

[0078] First, following the temporal order of the original signal's sliding window segmentation, the final denoising results of all independent samples are sequentially arranged and aligned to reconstruct the overall temporal structure. Since step S1 employs a 50% window overlap segmentation scheme, the intermediate non-boundary sampling points of the original signal are simultaneously included by the two adjacent windows, resulting in two independent denoising outputs. To eliminate boundary truncation artifacts, splicing jumps, and phase misalignment issues caused by independent window segmentation, this invention employs an overlapping region mean fusion strategy to complete signal reconstruction.

[0079] The specific execution method is as follows: for the overlapping time interval of two adjacent samples, the arithmetic mean of the two independent denoised values ​​corresponding to the same time position is calculated, and the mean result is used as the final output amplitude of the sampling point. The technical effect of this fusion method is: smoothly connecting the output differences of adjacent windows, completely eliminating the waveform jump, amplitude change and phase discontinuity problems caused by block processing at the splicing point, ensuring that the reconstructed time series signal has a smooth waveform, continuous phase and natural amplitude transition in the global range, and fully adapts to the accuracy requirements of subsequent spectrum analysis, geological inversion and imaging interpretation.

[0080] After completing the fusion of all overlapping window regions and splicing of the entire time series signal, a reverse preprocessing operation is performed to restore the original signal amplitude baseline. Since the mean-removing DC offset elimination process was performed on all samples in step S1, the amplitude baseline of the denoised result is inconsistent with the original acquired signal. Therefore, it is necessary to re-add the local mean subtracted in each window preprocessing stage to the reconstructed signal to restore the amplitude scale and baseline position of the original observation. This reverse recovery operation can ensure that the final clean output signal is completely consistent with the amplitude baseline and offset baseline of the original input signal, without changing the true physical amplitude characteristics of the signal, ensuring the accuracy of subsequent geological interpretation, data inversion, and anomaly analysis, and avoiding systematic amplitude deviation caused by the preprocessing operation.

[0081] The final output is a clean, complete time-series signal whose duration, sampling rate, amplitude reference, and other parameters are completely consistent with the original field-acquired signal. This is the final result of separating noise from effective signal.

[0082] Existing segmented time-series reconstruction techniques generally employ direct stitching without mean fusion schemes, which suffer from boundary waveform jump defects and lack quantitative verification. Furthermore, most schemes omit the preprocessing inverse mean recovery step, resulting in the output signal losing its physical amplitude reference, rendering it unusable for quantitative geological exploration. This embodiment supplements two sets of comparative experiments for complete explanation. The first set compares direct stitching with overlapping mean fusion. In the direct stitching scheme, the maximum amplitude jump at the window boundary reaches 10.1% of the original signal amplitude, leading to false high / low resistivity anomalies in subsequent two-dimensional electromagnetic inversion and incorrect mineralization identification in geological interpretation. In contrast, the mean fusion scheme exhibits a stitching jump amplitude below 0.7%, with no false anomaly generation, fully meeting the exploration interpretation accuracy standards. The second group, omitting the inverse mean recovery comparison with the complete inverse operation comparison, shows that omitting the inverse mean results in an overall shift in the signal baseline, with the relative deviation of the ore-induced electromagnetic anomaly amplitude reaching up to 42.3%. Quantitative calculations of underground medium resistivity and ore layer thickness are completely distorted, losing their value for quantitative analysis in mineral exploration. After complete overlay of the local mean window, the output signal amplitude perfectly matches the original observation physical scale, with a quantitative inversion error of less than 3.5%, meeting the data accuracy standards of the geophysical exploration industry. Simultaneously, a time-series reconstruction fallback verification standard is added. After reconstruction, the total number of sampling points of the original signal is automatically compared. If the difference exceeds one sampling point, an error is determined in the window truncation and fusion process, and the program automatically throws an exception and terminates the output. Based on this, the mathematical relationship between the reconstructed signal length and the number of windows is further verified: Let the number of windows be N, the window length be L, and the overlap rate be r (50% in this embodiment), then the theoretical total number of sampling points after reconstruction should be... If the actual reconstruction length does not match this theoretical value, the process is also judged to have an error and the output is terminated, thus avoiding timing misalignment caused by window sliding step calculation errors from the root.

[0083] All comparative experiments in this embodiment used the same field-controlled source electromagnetic (CSEM) measurement dataset. This dataset contains 1000 time-series samples labeled with noise types, with a uniform sampling rate of 1000Hz and a single sample length of 2048 sampling points. The noise types include four categories: power frequency harmonics, impulse interference, natural electromagnetic noise, and instrument background noise, with 250 samples in each category. The dataset is divided into a training set (700 samples), a validation set (100 samples), and a test set (200 samples) in a 7:1:2 ratio. Stratified random sampling is used to ensure that the distribution ratio of each type of noise is consistent in the three subsets. The random seed is fixed at 42 during partitioning to ensure that all ablation experiments use the exact same training, validation, and test set samples.

[0084] During the testing phase, noise classification accuracy (defined as the percentage of correctly classified samples in the test set out of the total number of samples) and denoising signal-to-noise ratio (defined as 10×log) were used as the benchmarks. 10The core evaluation metrics are (signal power / noise power, in dB) and waveform distortion rate (defined as the percentage of the root mean square error between the denoised output and the true clean signal to the signal amplitude). The experimental hardware environment consists of an NVIDIA RTX 3090 GPU (24GB VRAM), an Intel i7-12700 CPU, and 32GB DDR4 memory.

[0085] The software environment consisted of Ubuntu 20.04 LTS operating system, PyTorch 2.0 deep learning framework, CUDA 11.8, and Python 3.10. All experiments used fixed PyTorch, NumPy, and CUDA random seeds to ensure reproducibility. The experimental conditions were kept completely consistent across all comparative experiments; only the target parameter was changed. The controlled variable method was used to ensure that differences in experimental results across groups stemmed solely from variations in the measured parameter, rather than from other random factors.

[0086] Please see Figure 3 The present invention further provides an electromagnetic signal-to-noise separation system based on deep learning in complex noise environments. This electromagnetic signal-to-noise separation system is used to execute the aforementioned electromagnetic signal-to-noise separation method based on deep learning in complex noise environments, comprising: The noise acquisition module is used to acquire the electromagnetic signal to be separated, acquire several signal samples based on the electromagnetic signal to be separated, and extract the noise attribute information of the signal samples. The attribute extraction module is used to dynamically modulate the multi-layer feature response of the denoising network based on the noise attribute information, and to perform denoising processing on the signal sample to obtain a preliminary denoising result. The uncertainty assessment module is used to perform multiple random forward propagations on the same signal sample to obtain multiple denoising results, and to determine the preliminary denoising result, pointwise uncertainty measure and global uncertainty index based on the multiple denoising results; The iterative refinement module is used to generate a residual noise estimate based on the signal sample, the preliminary denoising result and the pointwise uncertainty measure in response to the global uncertainty index meeting the preset refinement condition, and update the preliminary denoising result according to the residual noise estimate, and iteratively execute until the termination condition is met. The signal reconstruction module is used to synthesize the separation result of the electromagnetic signal to be separated based on the preliminary denoising result corresponding to each signal sample.

[0087] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0088] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0089] 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; 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, depending on actual needs.

[0090] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for electromagnetic signal-to-noise separation in complex noise environments based on deep learning, characterized in that, include: Acquire the electromagnetic signal to be separated, and acquire several signal samples based on the electromagnetic signal to be separated, and extract the noise attribute information of the signal samples; Using the noise attribute information as a condition, the multi-layer feature response of the denoising network is dynamically modulated, and denoising processing is performed on the signal sample to obtain a preliminary denoising result; Multiple random forward propagations are performed on the same signal sample to obtain multiple denoising results, and the preliminary denoising result, pointwise uncertainty measure, and global uncertainty index are determined based on the multiple denoising results. In response to the global uncertainty index satisfying the preset refinement condition, a residual noise estimate is generated based on the signal sample, the preliminary denoising result, and the pointwise uncertainty measure, and the preliminary denoising result is updated according to the residual noise estimate. The process is iteratively executed until the termination condition is met. Based on the preliminary denoising results corresponding to each signal sample, the separation results of the electromagnetic signal to be separated are synthesized.

2. The electromagnetic signal-to-noise separation method based on deep learning in complex noise environments according to claim 1, characterized in that, The extraction of noise attribute information from the signal samples includes: Extract the time-domain and frequency-domain features of the signal samples; The time-domain features and the frequency-domain features are input into a pre-trained noise type-aware encoder to obtain the probability vectors of the signal samples belonging to each preset noise type.

3. The electromagnetic signal-to-noise separation method based on deep learning in complex noise environments according to claim 1, characterized in that, The step of dynamically modulating the multi-layer feature response of the denoising network based on the noise attribute information includes: The noise attribute information is input into the mapping network to generate scaling parameters and offset parameters; Based on the scaling and offset parameters, an affine transformation is performed on the target feature layer of the denoising network.

4. The electromagnetic signal-to-noise separation method based on deep learning in complex noise environments according to claim 1, characterized in that, The process of performing multiple random forward propagations on the same signal sample to obtain multiple denoising results, and determining a preliminary denoising result based on the multiple denoising results, includes: Keeping the randomly deactivated layer of the denoising network in an active state, performing T random forward propagations on the same input, and obtaining T denoising results; Calculate the mean of the T denoising results, and use it as the preliminary denoising result.

5. The electromagnetic signal-to-noise separation method based on deep learning in a complex noise environment according to claim 4, characterized in that, The determination of pointwise uncertainty measures and global uncertainty indices based on the multiple denoising results includes: Based on the T denoising results, the variance of each sampling point is calculated as the point-by-point uncertainty measure; The average variance of all sampling points is calculated as the global uncertainty index.

6. The electromagnetic signal-to-noise separation method based on deep learning in a complex noise environment according to claim 5, characterized in that, In response to the global uncertainty index satisfying a preset refinement condition, a residual noise estimate is generated based on the signal sample, the preliminary denoising result, and the pointwise uncertainty measure. The preliminary denoising result is then updated based on the residual noise estimate, and this process is iteratively executed until a termination condition is met, including: When the global uncertainty index is greater than a preset threshold, the preliminary denoising result and the signal sample are input into a pre-trained refinement network to obtain a residual noise estimate. The residual noise estimation is used to correct the multiple denoising results, and the preliminary denoising results and global uncertainty index are recalculated based on the corrected denoising results; Repeat the residual noise estimation and result update steps. The iteration is terminated when the updated global uncertainty index is not greater than the threshold. If the threshold condition is not met after the preset maximum number of iterations, the iteration is forcibly terminated.

7. The electromagnetic signal-to-noise separation method based on deep learning in a complex noise environment according to claim 6, characterized in that, The pre-trained refinement network adopts a residual convolutional structure with skip connections, and its training loss function is the mean squared error loss, which calculates the mean squared error between the residual noise estimate output by the refinement network and the true residual noise. The input of the refinement network includes three channels: noisy signal, preliminary denoising result, and pointwise uncertainty metric.

8. The electromagnetic signal-to-noise separation method based on deep learning in a complex noise environment according to claim 1, characterized in that, Determine the noise type corresponding to the highest probability component in the noise probability vector, and set the corresponding adjustment factor for each noise type; A preset base threshold for noise is established, and the product of the base threshold and the adjustment factor is calibrated as the stable threshold for that noise type. The preset stable threshold is dynamically set according to the noise type, specifically: the stable threshold is equal to the base threshold multiplied by the adjustment factor corresponding to the current noise type, where the current noise type is the noise type corresponding to the maximum probability component in the noise probability vector.

9. A deep learning-based electromagnetic signal-to-noise separation system for complex noise environments, the system being used to execute the deep learning-based electromagnetic signal-to-noise separation method for complex noise environments according to any one of claims 1-8, characterized in that, include: The noise acquisition module is used to acquire the electromagnetic signal to be separated, acquire several signal samples based on the electromagnetic signal to be separated, and extract the noise attribute information of the signal samples. The attribute extraction module is used to dynamically modulate the multi-layer feature response of the denoising network based on the noise attribute information, and to perform denoising processing on the signal sample to obtain a preliminary denoising result. The uncertainty assessment module is used to perform multiple random forward propagations on the same signal sample to obtain multiple denoising results, and to determine the preliminary denoising result, pointwise uncertainty measure and global uncertainty index based on the multiple denoising results; The iterative refinement module is used to generate a residual noise estimate based on the signal sample, the preliminary denoising result and the pointwise uncertainty measure in response to the global uncertainty index meeting the preset refinement condition, and update the preliminary denoising result according to the residual noise estimate, and iteratively execute until the termination condition is met. The signal reconstruction module is used to synthesize the separation result of the electromagnetic signal to be separated based on the preliminary denoising result corresponding to each signal sample.

Images