A Multi-Source Electromagnetic Noise Suppression Method Based on Noise Classification and Deep Learning

Abstract

This invention discloses a multi-source electromagnetic noise suppression method based on noise classification and deep learning, belonging to the field of geophysical electromagnetic exploration technology. The method includes: segmenting and preprocessing the original electromagnetic observation sequence; using an improved U-Net network with an encoder embedded in a Mamba time-series modeling module for low-frequency noise suppression; identifying strong noise types in the data segments using a ROCKET classifier; based on the classification results, calling a second improved U-Net network trained for the corresponding noise type for class-based targeted denoising; and finally, splicing the data segments to obtain complete, high-quality data. This invention, through a phased processing framework of "low-frequency pre-suppression—noise classification—class-based denoising," combined with the strong time-series modeling capabilities of the Mamba module and the efficient classification performance of ROCKET, significantly improves the suppression accuracy and signal fidelity for complex, multi-source electromagnetic noise, and is particularly suitable for processing ground and airborne electromagnetic exploration data.

Description

Technical Field

[0001] This invention belongs to the field of geophysical electromagnetic detection data processing technology, specifically relating to a multi-source electromagnetic noise suppression method based on noise classification and deep learning. Background Technology

[0002] Electromagnetic exploration methods acquire electrical structure information by observing the response of subsurface media to electromagnetic fields, and have been widely used in resource exploration and subsurface structure detection. However, actual electromagnetic observation data is susceptible to various types of noise interference. On the one hand, environmental disturbances (such as wind-induced vibration and platform movement) during the operation of the observation system can introduce low-frequency noise with significant time-series correlation, causing slow drift in the observation data. On the other hand, strong noises such as power frequency interference, triangular wave interference, and pulse interference are common in complex electromagnetic environments. These noises have high amplitudes, are diverse in type, and often superimpose, resulting in a low signal-to-noise ratio, which seriously affects the accuracy and reliability of subsequent data processing and geological interpretation.

[0003] Existing denoising techniques are mainly divided into two categories: traditional signal processing methods and deep learning-based methods.

[0004] Traditional denoising methods, such as wavelet transform, Kalman filtering, and empirical mode decomposition, usually rely on manual parameter setting and are mainly designed for single types of noise, with poor adaptability to complex superimposed noise of multiple types.

[0005] Deep learning-based denoising methods, such as recurrent neural networks, convolutional neural networks, and encoder-decoder structures (e.g., U-Net), can achieve end-to-end denoising. However, most of them use a single model to process all noise types uniformly, lacking mechanisms for identifying and differentiating different noise characteristics. In complex environments with low-frequency noise and multiple strong interferences, existing technologies struggle to simultaneously achieve both noise suppression and signal fidelity, resulting in incomplete denoising, trend distortion, or loss of detail.

[0006] Compared to ground-based observations, aerial electromagnetic data is more susceptible to low-frequency noise and various types of strong interference noise due to platform motion, attitude changes, and environmental disturbances. The noise types are more complex and their superposition characteristics are more pronounced, thus placing higher demands on the adaptability and robustness of denoising methods. This invention aims to provide a multi-source electromagnetic noise suppression method applicable to various electromagnetic observation data, including ground-based and aerial data, capable of handling complex noise environments under different acquisition methods. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-source electromagnetic noise suppression method based on noise classification and deep learning. By identifying and classifying different noises and combining them with an improved deep learning network for differentiated processing, this method achieves efficient and high-fidelity suppression of multi-source noise in complex electromagnetic observation data.

[0008] To achieve the above objectives, this invention provides a multi-source electromagnetic noise suppression method based on noise classification and deep learning, comprising the following steps:

[0009] Step 1, Data Segmentation and Preprocessing: Obtain the original electromagnetic observation time series, segment it according to a fixed time window length to obtain data segments, and preprocess each data segment;

[0010] Step 2, First stage denoising - low frequency noise suppression: The preprocessed data segment is input into the pre-trained first improved U-Net network to suppress low frequency noise in the data segment; the first improved U-Net network embeds the Mamba time series modeling module into each downsampling stage of the U-Net encoder and works alternately with the convolutional layer to enhance the ability to capture long-term dependencies and output the data segment after low frequency noise suppression;

[0011] Step 3, Noise Identification and Classification: Input the data segment after low-frequency noise suppression into the pre-trained ROCKET classifier to identify whether the data segment contains strong noise and the specific type of strong noise; the strong noise types include power frequency interference, triangular wave interference and pulse interference;

[0012] Step 4, Second Stage Denoising—Class-Based Directional Denoising: Based on the classification results of the ROCKET classifier, data segments without strong noise are directly retained; for data segments containing specific types of strong noise, they are input into a pre-trained second improved U-Net network corresponding to that type of strong noise for directional denoising processing, resulting in the final denoised data segments.

[0013] Step 5, Data Reconstruction: All data fragments processed in Steps 2 and 4 are sequentially spliced ​​together according to their original time order to restore complete electromagnetic observation data with the same length as the original input time series, thus obtaining high-quality electromagnetic observation results.

[0014] Preferably, in step 1, each data segment is 1 second long, corresponding to 2400 sampling points; the preprocessing includes mean removal and normalization of the data segments; first, mean removal is performed on each data segment to eliminate DC bias; then, normalization is performed to reduce the impact of amplitude scale differences between different samples on model training and classification results; the normalization process uses the maximum absolute value normalization method, and the process is as follows:

[0015] set up For samples to be normalized, Indicates the first in the sample One element, N is the number of samples; the elements of normalization Defined as:

[0016] ;

[0017] The normalized sample is ,and This ensures that the normalized sample values ​​fall within the [-1,1] interval, thereby improving the stability of network training and the generalization ability of the model while maintaining the relative shape of the original waveform.

[0018] Preferably, the first improved U-Net network described in step 2 includes an encoder, a bottleneck layer, and a decoder; the encoder extracts features progressively through three levels of downsampling, the encoder input being a single-channel one-dimensional time series, with the number of channels increasing sequentially from 1 to 64, 128, and 256; each level consists of two residual blocks, a max-pooling layer, and a Mamba temporal modeling module; the Mamba temporal modeling module is placed after each level of downsampling in the encoder and alternates with the convolutional layers; the bottleneck layer contains two residual blocks and a Mamba temporal modeling module; the bottleneck layer maps features to 512 channels. The decoder is used to extract higher-level global temporal features. It restores spatial resolution through three levels of upsampling, each level containing a deconvolutional layer and a residual block, and fuses the encoder features of the corresponding level through skip connections. The Mamba temporal modeling module is based on a selective state-space model, which includes a state-space branch and a convolutional branch. The state-space branch is used to capture long-term dependencies and global trends, while the convolutional branch is used to extract local short-term patterns. The two branches are fused at the output through a gating mechanism, enabling the model to adaptively adjust feature weights at different time scales to capture long-term dependencies and global trends in the sequence.

[0019] Preferably, in step 2, the first improved U-Net network is obtained by supervised training on a sample set containing low-frequency noise; the output result has the same length as the input data segment and is used to represent the reconstructed time series after low-frequency noise suppression; the first improved U-Net network is trained using supervised learning, with mean squared error as the loss function, Adam as the optimizer, an initial learning rate of 0.001, a batch size of 16, and 100 training rounds.

[0020] Preferably, the ROCKET classifier in step 3 is trained using a labeled sample set, wherein the sample labels include four categories: "no strong noise", "power frequency interference", "triangular wave interference", and "pulse interference". The training and classification process of the ROCKET classifier includes: generating several random one-dimensional convolution kernels, with the kernel length uniformly distributed logarithmically and the dilation rate randomly sampled to cover different time scales; performing convolution operations on the data segments to be classified with each convolution kernel to perform multi-scale feature mapping on the data segments and obtain response sequences; extracting the maximum value (MAX) and positive percentage value (PPV) from the response sequence of each convolution kernel as statistical features to construct a feature vector; the feature vector is used to characterize the multi-scale temporal morphological features of the input data segments and serves as the input to the subsequent linear classifier; inputting the feature vector into a ridge regression classifier with L2 regularization for class discrimination and outputting the class label: "no strong noise", "power frequency interference", "triangular wave interference", or "pulse interference".

[0021] Preferably, in step 4, the structure of the second improved U-Net network corresponding to the noise type is the same as that of the first improved U-Net network. However, each second improved U-Net network is obtained through independent supervised training for power frequency interference, triangular wave interference, and pulse interference sample sets, and the parameters of each network are independent of each other and do not share weights, so as to enhance the targeted modeling ability of the corresponding strong noise type features. The training parameter settings of each second improved U-Net network are the same as those of the first improved U-Net network, all of which use the Adam optimizer, mean square error loss function, initial learning rate of 0.001, batch size of 16, and training rounds of 100.

[0022] The beneficial effects of this invention are as follows:

[0023] High precision and high fidelity: Through a phased processing framework of "low-frequency noise pre-suppression - strong noise identification and classification - class-based directional denoising", different types of noise are processed separately, which significantly improves the overall denoising accuracy in complex noise environments and effectively preserves the detailed features in the signal, avoiding the over-smoothing problem that is easily caused by traditional single-model methods.

[0024] Highly targeted and robust: The introduction of the ROCKET classifier enables efficient and accurate identification of strong noise types, allowing the system to adaptively select the optimal denoising model based on the specific form of the noise (power frequency, triangular wave, pulse), overcoming the shortcomings of existing technologies that treat different noises in a "one-size-fits-all" manner, and exhibiting stronger robustness in the case of multiple noise superposition.

[0025] Superior Low-Frequency Noise and Long-Range Dependency Modeling Capabilities: By embedding the Mamba temporal modeling module into the U-Net encoder, the network's ability to capture long-term sequence dependencies is creatively enhanced. This enables the method to effectively suppress low-frequency drift and coherent interference caused by environmental disturbances or platform motion, solving the problems of limited receptive field and poor low-frequency trend recovery in existing convolutional networks.

[0026] Significant physical consistency and application effects: Data processed by this method exhibits a more continuous and stable distribution of tilt parameters, with a clearer and more explicit correspondence to geological structures. Field imaging results demonstrate that this method effectively suppresses background noise, significantly enhances the identification ability of electrically abrupt boundary changes (such as ore deposit boundaries), and improves the geological interpretability of electromagnetic exploration data. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0028] Figure 2 This is a schematic diagram of the ROCKET classifier process of the present invention;

[0029] Figure 3 This is a schematic diagram of the Mamba module process of the present invention;

[0030] Figure 4 This is a schematic diagram of the overall architecture of the improved U-Net of this invention;

[0031] Figure 5 Here are typical sample examples from the sample library: (a) low-frequency noise and power frequency interference, (b) low-frequency noise and triangular wave interference, (c) low-frequency noise and pulse interference, and (d) low-frequency noise.

[0032] Figure 6 This is a heatmap of the confusion matrix according to an embodiment of the present invention;

[0033] Figure 7 This is a bar chart showing the accuracy / recall / F1 score of an embodiment of the present invention.

[0034] Figure 8 The following are time-domain waveform comparison diagrams of synthesized data and clean signal in an embodiment of the present invention: (a) Time-domain waveform comparison diagram of the first 60 seconds, (b) Time-domain waveform comparison diagram of the first 4 seconds.

[0035] Figure 9 The following are time-domain waveform comparisons of different methods used to denoise the synthesized data in the embodiments of the present invention: (a) low-frequency noise and power frequency interference, (b) low-frequency noise and triangular wave interference, (c) low-frequency noise and pulse interference, and (d) low-frequency noise. Detailed Implementation

[0036] To construct a high-quality sample library suitable for model training, this invention comprehensively uses three types of data: measured clean signals, simulated strong noise, and measured low-frequency noise, to generate training samples covering typical interference types, and divides them into three datasets according to different task requirements.

[0037] First, this invention selects electromagnetic observation magnetic field data obtained from a ground-based acquisition station in Jixi City, Heilongjiang Province, as the source of the clean signal. This acquisition station is far from residential areas and major sources of electromagnetic interference such as high-voltage transmission lines, resulting in a low overall noise level. Manual inspection confirmed that the signal exhibits high stability and a high signal-to-noise ratio, making it suitable as an effective signal reference. The sampling rate of this data is 2400Hz. To maintain a consistent sample structure, the complete sequence is divided into segments of 2400 points each (corresponding to 1 second), and 2000 high-quality records with no significant noise are selected as clean signal samples.

[0038] To construct the various types of strong noise samples required for model training, this invention simulates and models power frequency interference, triangular wave interference, and pulse interference based on the morphological characteristics of common electromagnetic interference. The amplitude, frequency, period, and pulse width of the noise waveform are randomly sampled within a reasonable range based on field observation experience to cover the possible variations of different strong noises in the actual observation environment. The parameter settings are shown in Table 1, where A... s This represents the root mean square amplitude of the effective electromagnetic signal within the corresponding time window.

[0039] Table 1. Simulation parameters for environmental noise

[0040] Noise type parameter Range of values unit Power frequency interference baseband 49–51 Hz Power frequency interference Harmonic order 1–5 – Power frequency interference Amplitude <![CDATA[3–10×A s ]]> – Power frequency interference initial phase 0–2π rad Power frequency interference Duration 0.5–5 s Triangular wave interference baseband 1–20 Hz Triangular wave interference Amplitude <![CDATA[2–8×A s ]]> – Triangular wave interference Duty cycle 0.3–0.7 – Triangular wave interference slope random – Triangular wave interference Duration 1–10 s Pulse interference Amplitude <![CDATA[5–20×A s ]]> – Pulse interference width 1–20 ms Pulse interference interval 0.1–2 s Pulse interference Pulse polarity ±1 – Pulse interference probability of occurrence 0.1–0.5 –

[0041] For each type of noise, several different parameter combinations are randomly generated and superimposed with 2000 clean signals respectively to construct three strong noise sample sets of the same size (100,000 samples each, for a total of 300,000 samples).

[0042] Furthermore, this invention utilizes aerial observation magnetic field data collected by an aerial platform-mounted magnetic field device in Baishan City, Jilin Province, and combines it with ground magnetic field data collected from stationary ground points in the same area and at the same time. Through correlation analysis between the aerial magnetic field and the ground magnetic field, low-frequency noise introduced by platform motion and attitude changes is separated. This noise sampling rate is also 2400Hz, and after processing, a low-frequency noise sequence of approximately 2000 seconds is obtained. To maintain consistency with all previously mentioned samples, this sequence is sliced ​​into 2400-point segments. 2000 clean signals and 300,000 samples containing strong noise are then superimposed one by one with the sliced ​​low-frequency noise to construct a comprehensive sample containing both low-frequency and strong noise, resulting in a total of 302,000 mixed noise samples.

[0043] Based on the requirements of subsequent tasks, the completed samples will be divided into three datasets:

[0044] (1) Dataset A: 302,000 samples containing both low-frequency noise and strong noise, used to train the first improved U-Net network to achieve low-frequency noise suppression;

[0045] (2) Dataset B: In order to avoid the impact of imbalanced sample distribution on classification performance, 2000 samples of each of the three types of noise samples were uniformly extracted and 2000 clean signal samples were added to train the ROCKET classifier to achieve strong noise type recognition and classification.

[0046] (3) Dataset C: 300,000 samples containing strong noise, used to train the second improved U-Net network to achieve class-based denoising. Among them, Dataset A is used to train the first-stage low-frequency noise suppression model, Dataset B is used to train the strong noise recognition and classification model, and Dataset C is further divided into power frequency interference sample set, triangular wave interference sample set and pulse interference sample set according to noise type, which are used to train the corresponding second-stage directional denoising models. Figure 5 It showcases representative sample types from the sample library.

[0047] This embodiment provides a multi-source electromagnetic noise suppression method based on noise classification and deep learning, such as... Figure 1 As shown, it includes the following steps:

[0048] Step 1: Data Segmentation and Preprocessing

[0049] In aerial electromagnetic observation scenarios, the system typically uses an aerial platform to acquire signals. Environmental disturbances or platform movement can introduce significant low-frequency noise into the geomagnetic field. To obtain stable baseline data, the original electromagnetic observation time series is first acquired. The original series is then segmented into 1-second (2400 points) segments to obtain data fragments, and each data fragment is preprocessed.

[0050] Since the amplitude of electromagnetic signals changes relatively smoothly, to improve the distinguishability between different samples, all sequences are first mean-reduced during sample construction. Subsequently, to eliminate the influence of amplitude scale differences and improve the stability of model training, Max-Abs normalization is applied to the samples. The normalization process is as follows:

[0051] set up For samples to be normalized, Indicates the first in the sample One element, N is the number of samples; the elements of normalization Defined as:

[0052]

[0053] The normalized sample is ,and This normalization method can preserve the original waveform structure and relative amplitude relationship, and is applicable to one-dimensional magnetic field data containing positive and negative values, thereby improving the model's generalization ability and training stability.

[0054] Step 2, First Stage Noise Reduction—Low-Frequency Noise Suppression:

[0055] The preprocessed data segment is input into the first improved U-Net network for low-frequency noise suppression, and the data segment after low-frequency noise suppression is output.

[0056] The traditional U-Net network is a typical encoder-decoder structure that simultaneously captures local details and global contextual information through multi-scale convolutions and skip connections. Based on its advantages in feature representation and reconstruction, this invention first employs a one-dimensional U-Net network as the baseline network to achieve end-to-end initial denoising of electromagnetic observation data. However, the convolutional operations of the U-Net network rely on a fixed receptive field, extracting features only from local neighborhoods and lacking the ability to model long-range dependencies over time. The main noise in electromagnetic observation data exhibits significant temporal structure characteristics. Traditional U-Net, relying solely on local convolutions, struggles to capture multiple temporal patterns simultaneously, potentially leading to over-smoothing of effective information in a highly coherent noise background or noise residue during detail recovery, thus limiting its denoising performance in complex temporal contexts.

[0057] To address the shortcomings of traditional U-Net networks in temporal modeling capabilities, this invention introduces the efficient Mamba temporal modeling module into the U-Net encoding stage to enhance the model's ability to express temporally relevant structures. Mamba is a temporal modeling architecture based on the Selective State Space Model (SSM). Its core lies in a selective state update mechanism, where the model adaptively adjusts the information update intensity based on input features during state evolution, thereby maintaining a high response to key task-related temporal features while suppressing noise or irrelevant inputs. This mechanism, while maintaining linear computational complexity, enables the model to effectively capture dependencies in long-term series, and is particularly suitable for one-dimensional time-series signals containing trend changes, periodic structures, and abrupt changes.

[0058] In terms of network architecture design, the first improved U-Net network of this invention embeds the Mamba temporal modeling module into each downsampling stage of the U-Net encoder, alternating with convolutional layers. The convolutional layers are primarily responsible for extracting local temporal features, while the Mamba temporal modeling module is used to model long-term dependencies across time scales. This structure retains the advantages of the U-Net encoder-decoder structure and skip connections while enhancing the model's overall ability to suppress complex temporal noise in electromagnetic observation data (such as low-frequency drift caused by platform motion, periodic interference, and transient pulses). The structure of the Mamba temporal modeling module embedded in U-Net is as follows: Figure 3 As shown, the Mamba temporal modeling module is placed after each downsampling stage, mainly based on three considerations. First, downsampling shortens the feature length, which reduces the computational cost of Mamba and improves inference speed. Second, downsampling expands the receptive field, making it more suitable for Mamba to model long temporal dependencies. Finally, the shallow layers of the encoder retain the efficient extraction of local features from convolution, while the deeper layers use Mamba to enhance global modeling, which conforms to the feature extraction rule from fine to coarse.

[0059] Figure 3 The overall structure of the Mamba module is shown, which consists of a state space branch and a convolutional branch: the former is used to capture long-term dependencies and global trends, while the latter is used to extract local short-term patterns. The outputs of the two branches are used to achieve adaptive fusion between convolutional features and Mamba features through a gating mechanism and an element-wise weighted approach. The gating weights are generated by the non-linear activation function SiLU, which enables the model to adaptively adjust feature weights at different time scales, thereby exhibiting stronger robustness in complex temporal noise backgrounds.

[0060] The first improved U-Net network described in this invention is a convolutional neural network specifically designed for one-dimensional sequence data. It uses the U-Net architecture to implement end-to-end regression tasks, and its overall architecture is as follows: Figure 4As shown in the figure, the orange module within the red dashed box represents the Mamba temporal modeling module. The first improved U-Net network structure consists of three parts: an encoder, a bottleneck layer, and a decoder. The encoder extracts features progressively through three levels of downsampling, with each level consisting of two residual blocks, a max-pooling layer, and a Mamba temporal modeling module. The Mamba temporal modeling module, as an efficient long sequence modeling component, enhances the network's ability to capture long-term dependencies in sequences, with the number of channels increasing from 1 to 64, 128, and 256 respectively. The bottleneck layer contains two residual blocks and a Mamba module, compressing the features to 512 channels. The decoder restores spatial resolution through three levels of upsampling, with each level containing a deconvolutional layer and a residual block, and fuses the encoder features from corresponding levels through skip connections, gradually reducing the number of channels to 64. Finally, a 1×1 convolutional layer outputs a single-channel prediction result. This structure enables the network to simultaneously capture multi-scale features and long-term dependencies in sequence data, making it particularly suitable for fields such as temporal signal analysis that require consideration of both local details and global context.

[0061] Step 3, Noise Identification and Classification:

[0062] After removing low-frequency noise, only some data segments still contain strong noise such as power frequency interference, triangular wave interference, and pulse interference. Performing uniform strong noise removal on all data segments would not only be inefficient but might also involve redundant processing of high-quality segments, thus reducing the necessity and specificity of the overall processing. Therefore, this invention introduces the ROCKET classification algorithm to identify whether strong noise is present and further determine its specific type, enabling rapid classification of different types of strong noise signals.

[0063] The data segment after low-frequency noise suppression is input into the trained ROCKET classifier to identify whether the data segment contains strong noise and the specific type of strong noise; the strong noise types include power frequency interference, triangular wave interference and pulse interference.

[0064] The ROCKET classifier performs multi-scale feature mapping on the original sequence using a large number of randomly generated one-dimensional convolutional kernels, and extracts simple statistics (such as maximum value and positive value proportion) from the convolution results. Finally, a linear classifier completes the class discrimination. This method does not require manual feature construction, has low training overhead, and is robust to noise with significant amplitude differences and diverse shapes, making it suitable for the identification and classification of strong noise in electromagnetic observation data. Figure 2 It demonstrates the classification principle of ROCKET.

[0065] Specifically, to ensure the reproducibility of experimental results, all random operations in this step use a fixed random seed. Dataset B contains 8000 samples, which are randomly selected stratified according to noise type: 90% (7200 samples) are used as the training set, and 10% (800 samples) are used as the independent test set. The training set is used for model training and hyperparameter tuning (5-fold cross-validation), while the test set is only used for final performance evaluation, using independent data other than cross-validation to verify the model's generalization performance. All data segments are 2400 points in length and labeled with their corresponding noise categories.

[0066] During the feature extraction stage, the ROCKET classifier generates K=10000 random one-dimensional convolutional kernels. The kernel length L is uniformly distributed logarithmically from 7 to 150, and the dilation rate d is randomly sampled within the range [1, 32]. For each input data... Each convolution kernel Perform a one-dimensional convolution to obtain the response sequence:

[0067]

[0068] in, Represents the response sequence. R represents the number of convolution kernels; R represents the set of real numbers. This represents a 2400-dimensional real vector space; the entire expression This means that x is a one-dimensional real number sequence of length 2400;

[0069] Here, x is actually a 1-second electromagnetic time series data segment because the sampling rate is 2400.

[0070] Subsequently, two statistical features, MAX and PPV (Proportion of Positive Values), are calculated for each response sequence to characterize the peak structure and sign distribution of the response, thus constructing a feature vector with dimension 2K=20000. Global mean and standard deviation are calculated along the feature dimensions on the training set, and z-score standardization is performed on the training set features. The test set is normalized using the same global mean and standard deviation to avoid information leakage.

[0071] The normalized features are input into a Ridge Regression Classifier with L2 regularization, and the regularization strength is optimized on the training set using grid search combined with 5-fold cross-validation. The optimal regularization strength is selected based on the average accuracy of cross-validation. Subsequently, the ROCKET classifier was retrained on the complete training set, and the final model and random convolution kernel parameters were serialized and saved together to ensure the reproducibility of the experiment. The classification results were used to determine whether the current data segment needed to enter the second stage of denoising processing, and which type of strong noise-corresponding denoising model to use.

[0072] Figure 6 and Figure 7 The performance of the ROCKET classifier on a test set, which includes four types of samples: power frequency interference, triangular wave interference, impulse interference, and clean signals, is demonstrated. The confusion matrix heatmap shows that the diagonal classification accuracy for all four signal types is above 98.6%, while the misclassification rate for all off-diagonal elements is below 0.7%. Specifically, there is approximately 0.5% confusion between impulse signals and clean signals, and approximately 0.6% confusion between triangular wave signals and power frequency interference. These misclassifications mainly stem from the similarity of some signals in local temporal characteristics: for example, triangular waves and power frequency interference overlap in their periodic structures, and impulse signals exhibit short-term stationary behavior similar to clean signals in a few segments, thus leading to confusion in a very small number of samples.

[0073] Figure 7 The model further demonstrates the precision, recall, and F1-score (harmonic mean of precision and recall) for various signal types. For clean signals, all three metrics are close to 99.0%, indicating a high degree of consistency in their characteristic patterns and stable recognition by the model. The three metrics for power frequency interference and pulse signals remain around 98.8%, demonstrating the model's good ability to distinguish between periodic and abrupt noise. In contrast, the precision and recall for triangular wave interference are slightly lower (approximately 98.6%), mainly due to the partial overlap of its frequency structure with that of power frequency signals, introducing some uncertainty near the classification boundary.

[0074] Overall, the F1-scores for all four signal types remained above 98.6%, indicating that the model demonstrated excellent stability in controlling false positives and false negatives. The results show that the ROCKET classifier, relying on multi-scale statistical features constructed using random convolutional kernels, can accurately capture the temporal differences between different types of noise, thus achieving extremely high classification accuracy and excellent generalization performance.

[0075] Step 4, Second Stage Denoising—Class-Based Directional Denoising:

[0076] Based on the classification results of the ROCKET classifier, data segments without strong noise are directly retained, while data segments containing specific types of strong noise are denoised by calling the second improved U-Net network corresponding to that noise type to obtain the final denoised data segments. This invention adopts a "class-based denoising" strategy to avoid over-smoothing or loss of detail that may be caused by a uniform model.

[0077] In step 4, the structure of the second improved U-Net network corresponding to the noise type is the same as that of the first improved U-Net network, but each second improved U-Net network is trained separately for a specific set of strong noise types. Since the second-stage denoising model is trained separately for different strong noise types, the classification result will affect the directional denoising branch that the data segment enters; however, since the ROCKET classifier has a high classification accuracy, the impact of misclassification on the overall denoising performance is limited.

[0078] Step 5: Data reconstruction: All data fragments processed in Steps 2 and 4 are spliced ​​together in their original time order to obtain complete, high-quality electromagnetic observation data.

[0079] Performance verification and effect comparison:

[0080] To verify the performance of the adaptive denoising method based on the ROCKET classifier and the improved U-Net network proposed in this invention under complex noise conditions, this embodiment designed three experimental schemes and compared them under completely identical training configurations.

[0081] Solution A: Non-adaptive traditional U-Net network, which directly performs one-time denoising on signals containing both low-frequency noise and strong noise;

[0082] Option B: Non-adaptive improved U-Net network, the same as Option A, but with the introduction of the Mamba module in the encoder to enhance the temporal modeling capability;

[0083] Solution C: The adaptive denoising method based on ROCKET noise classification and improved U-Net network provided by this invention first suppresses low-frequency noise, and secondly performs customized strong noise suppression on the classified signal according to the noise type.

[0084] The main differences among the three schemes lie in their temporal modeling capabilities and whether a two-stage structure is employed. All schemes were trained using the hyperparameters shown in Table 2. During training, all random operations were set with the same random seed to ensure the reproducibility of experimental results and fair comparison between different schemes.

[0085] Table 2. Model Parameters

[0086] hyperparameters Parameter settings Optimizer Adam Initial learning rate 0.001 Learning rate adjustment strategy Cosine annealing loss function Mean Squared Error (MSE) Batch size 16 samples / batch Number of training rounds 100

[0087] To evaluate the generalization ability of this invention on unseen data, this embodiment constructs an independent synthetic test set based on high signal-to-noise ratio measured data from Yangcao Town, Suihua City, Heilongjiang Province. The electromagnetic environment in this area has weak background interference, and the original time series has a high signal-to-noise ratio, making it suitable as an approximately clean reference signal for superimposing different types of noise to generate test samples. Based on this, low-frequency noise, power frequency interference, triangular wave interference, and pulse interference are superimposed, and parameters such as amplitude, frequency, and pulse width are randomized to simulate the multi-type noise environment in real electromagnetic observations.

[0088] Subsequently, schemes A, B, and C were used to denoise the synthetic data that did not overlap with the training set. The performance differences of the three schemes in the new scenario were systematically evaluated from three dimensions: whether the denoised waveform is close to the theoretical true value, whether there is excessive smoothing or waveform distortion, and whether different types of strong noise can be accurately suppressed.

[0089] Figure 8 The time-domain comparison between the synthesized data and the theoretical interference-free signal is shown, in which... Figure 8 (a) shows the global waveform for the first 60 seconds. Figure 8 (b) is a magnified view of the first 4 seconds. The solid blue line represents the noisy signal after multiple noises are superimposed, and the dashed red line represents the original data with a high signal-to-noise ratio, which can be considered as a theoretically interference-free reference signal. From Figure 8 As can be seen, the synthesized data after noise superposition is significantly affected by low-frequency noise in its overall trend. At the same time, the local structure is disrupted by typical interferences such as power frequency, triangular waves, and pulses, resulting in overall signal shift and enhanced local oscillations, making it difficult to use directly for subsequent processing. This figure provides a theoretical truth benchmark for subsequent comparison of denoising effects.

[0090] Figure 9 The temporal denoising effects of schemes A, B, and C on four typical types of strongly noisy samples were compared. Figure 9 (a)–(d) represent samples with power frequency interference, triangular wave interference, pulse interference, and low-frequency noise only, respectively. The blue, green, and red curves represent the outputs of schemes A, B, and C, respectively, while the black curve represents the clean reference signal. Overall, while schemes A and B mitigate low-frequency drift to some extent, they still generally suffer from two problems: first, excessive waveform smoothing leads to the loss of peak-valley structure; and second, varying degrees of residual noise remain. In contrast, scheme C achieves higher waveform fidelity and more complete noise suppression in all scenarios.

[0091] In the case of low-frequency noise only, schemes A and B exhibit trend shifts and loss of local details after denoising, while scheme C effectively removes low-frequency drift while maintaining key structures such as peaks and valleys, almost perfectly matching the theoretical true value. When the signal is simultaneously subjected to triangular wave interference, schemes A and B still show obvious periodic residues and baseline shifts in their outputs, while scheme C can suppress both types of noise simultaneously, with waveform trends and details consistent with the theoretical true value. In power frequency interference and pulse interference scenarios, schemes A and B respectively exhibit problems with power frequency component residues or blurred abrupt change positions, while scheme C can still accurately recover the signal abrupt change characteristics without excessive smoothing.

[0092] In summary, Scheme C can robustly recover the true form of the signal under various noise conditions, which is significantly better than Schemes A and B, verifying the effectiveness of the proposed method in noise suppression and signal fidelity preservation.

[0093] Table 4. Quantitative statistics of denoising results using different methods

[0094] plan SNR (dB) RE NCC Training time Reasoning time Before processing -18.66 8.5665 0.0460 - - A -1.06 1.1304 0.3815 2731s 0.10s B 2.14 0.7818 0.6433 3364s 0.11s C 7.29 0.4320 0.9146 6228s 0.21s

[0095] Table 4 shows the denoising performance of the three methods under the same data conditions. Scheme C achieves the best performance among the three methods, with an SNR of 7.29 dB, an NCC of 0.9146, and a RE of 0.4320. The Mamba module effectively improves sequence modeling capabilities, while ROCKET classification significantly enhances denoising accuracy in noisy environments. Scheme C, combining both, outperforms Schemes A and B in both noise suppression and signal fidelity, fully validating the effectiveness and generalization ability of the denoising method proposed in this invention.

Claims

1. A method for suppressing multi-source electromagnetic noise based on noise classification and deep learning, characterized in that: Includes the following steps: Step 1, Data Segmentation and Preprocessing: Obtain the original electromagnetic observation time series, segment it according to a fixed time window length to obtain data segments, and preprocess each data segment; Step 2, First stage denoising - low frequency noise suppression: The preprocessed data segment is input into the pre-trained first improved U-Net network to suppress low frequency noise in the data segment; the first improved U-Net network embeds the Mamba time series modeling module into each downsampling stage of the U-Net encoder and works alternately with the convolutional layer to enhance the ability to capture long-term dependencies and output the data segment after low frequency noise suppression; Step 3, Noise Identification and Classification: Input the data segment after low-frequency noise suppression into the pre-trained ROCKET classifier to identify whether the data segment contains strong noise and the specific type of strong noise; Step 4, Second Stage Denoising—Class-Based Directional Denoising: Based on the classification results of the ROCKET classifier, data segments without strong noise are directly retained; for data segments containing specific types of strong noise, they are input into a pre-trained second improved U-Net network corresponding to that type of strong noise for directional denoising processing to obtain the final denoised data segments. Step 5, Data Reconstruction: All data fragments processed in Steps 2 and 4 are sequentially spliced ​​together according to their original time order to restore complete electromagnetic observation data with the same length as the original input time series, thus obtaining high-quality electromagnetic observation results.

2. The multi-source electromagnetic noise suppression method based on noise classification and deep learning according to claim 1, characterized in that: In step 1, each data segment is 1 second long and corresponds to 2400 sampling points. The preprocessing includes mean removal and normalization of the data segments. First, mean removal is performed on each data segment to eliminate DC bias. Then, normalization is performed to reduce the impact of amplitude scale differences between different samples on model training and classification results.

3. The multi-source electromagnetic noise suppression method based on noise classification and deep learning according to claim 2, characterized in that: In step 1, the normalization process uses the maximum absolute value normalization method, as follows: set up For samples to be normalized, Indicates the first in the sample One element, N is the number of samples; the elements of normalization Defined as: ; The normalized sample is ,and This ensures that the normalized sample values ​​fall within the interval [-1, 1].

4. The multi-source electromagnetic noise suppression method based on noise classification and deep learning according to claim 1, characterized in that: The first improved U-Net network described in step 2 includes an encoder, a bottleneck layer, and a decoder. The encoder extracts features progressively through three levels of downsampling. The encoder input is a single-channel one-dimensional time series, with the number of channels increasing sequentially from 1 to 64, 128, and 256. Each level consists of two residual blocks, a max-pooling layer, and a Mamba temporal modeling module. The Mamba temporal modeling module is placed after each downsampling level of the encoder and works alternately with the convolutional layers. The bottleneck layer contains two residual blocks and a Mamba temporal modeling module. The bottleneck layer maps features to 512 channels for extracting higher-level global temporal features. The decoder restores spatial resolution through three levels of upsampling, each level containing a deconvolution layer and a residual block, and fuses encoder features of the corresponding levels through skip connections.

5. The multi-source electromagnetic noise suppression method based on noise classification and deep learning according to claim 1, characterized in that: The Mamba time series modeling module is based on a selective state space model. Its structure includes a state space branch and a convolutional branch. The state space branch is used to capture long-term dependencies and global trends, while the convolutional branch is used to extract local short-term patterns. The two branches are fused at the output through a gating mechanism, enabling the model to adaptively adjust feature weights at different time scales to capture long-term dependencies and global trends in the sequence.

6. The multi-source electromagnetic noise suppression method based on noise classification and deep learning according to claim 1, characterized in that: In step 2, the first improved U-Net network is obtained through supervised training on a sample set containing low-frequency noise; the output result has the same length as the input data segment and is used to represent the reconstructed time series after low-frequency noise suppression; the first improved U-Net network uses mean squared error as the loss function during training, the optimizer is Adam, the initial learning rate is 0.001, the batch size is 16, and the number of training rounds is 100.

7. The multi-source electromagnetic noise suppression method based on noise classification and deep learning according to claim 1, characterized in that: The types of strong noise mentioned in step 3 include power frequency interference, triangular wave interference, and pulse interference.

8. The multi-source electromagnetic noise suppression method based on noise classification and deep learning according to claim 1, characterized in that: The ROCKET classifier described in step 3 is trained using a labeled sample set, where the sample labels include four categories: "no strong noise", "power frequency interference", "triangular wave interference" and "pulse interference". The training and classification process of the ROCKET classifier includes: generating several random one-dimensional convolution kernels, with kernel lengths uniformly distributed logarithmically and dilation rates randomly sampled to cover different time scales; performing convolution operations on the data segments to be classified with each convolution kernel to perform multi-scale feature mapping on the data segments and obtain response sequences; extracting the maximum value and the proportion of positive values ​​from the response sequences of each convolution kernel as statistical features to construct feature vectors; the feature vectors are used to characterize the multi-scale temporal morphological features of the input data segments and serve as input to the subsequent linear classifier; and inputting the feature vectors into a ridge regression classifier with L2 regularization for class discrimination, outputting class labels: "no strong noise", "power frequency interference", "triangular wave interference" or "pulse interference".

9. The multi-source electromagnetic noise suppression method based on noise classification and deep learning according to claim 1, characterized in that: In step 4, the structure of the second improved U-Net network corresponding to the noise type is the same as that of the first improved U-Net network. The second improved U-Net network corresponding to each noise type is obtained through independent supervised training for power frequency interference, triangular wave interference and pulse interference sample sets, and the parameters of each network are independent of each other and do not share weights. The training parameters of each second improved U-Net network use the mean squared error loss function, and all adopt the Adam optimizer, with an initial learning rate of 0.001, a batch size of 16 and a training epoch of 100.

Images