A surface magnetic field measurement anomaly data identification and correction method, system, device and medium based on deep learning and adaptive wavelet threshold fusion algorithm

Abstract

This invention discloses a method, system, equipment, and medium for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm. It utilizes a high-sensitivity magnetic field sensor to collect characteristic magnetic field signals generated by buried cables. To address interference from metal pipes, vehicles, industrial electromagnetic noise, and geomagnetic disturbances during the measurement process, a deep learning anomaly detection model is constructed to extract features and identify anomalies in the magnetic field time-series data. An adaptive wavelet threshold denoising algorithm is then used to correct and reconstruct the identified anomaly segments. This method effectively eliminates abnormal observations caused by external interference, retains the true cable signal, and significantly improves the signal-to-noise ratio and reliability of surface magnetic field measurement data, providing high-precision input data for subsequent cable path inversion and 3D positioning. This method overcomes the problems of misjudgment and decreased positioning accuracy in strong interference environments associated with traditional magnetic field measurement methods, thus improving the accuracy of anomaly data identification.

Description

Technical Field

[0001] This invention belongs to the field of geomagnetic measurement and cable monitoring technology, and proposes a method, system, equipment and medium for identifying and correcting abnormal data of surface magnetic field measurement based on deep learning and adaptive wavelet threshold fusion algorithm. Background Technology

[0002] With the rapid development of urban underground cable systems, power transmission and distribution networks are becoming increasingly dense and concealed. Traditional methods relying on manual inspections or cable identification devices are no longer sufficient to meet the precision requirements of modern urban power operation and maintenance. Surface magnetic field measurement technology, as a contact-based, low-interference method for cable condition sensing, reflects the spatial path characteristics of underground cables by analyzing the magnetic field distribution formed on the surface through the injection of DC or low-frequency current. It offers advantages such as ease of implementation, low cost, and wide applicability. Currently, this technology has been applied in scenarios such as cable path location, live-line identification, and partial discharge monitoring. However, due to the weak amplitude of surface magnetic field signals and their significant susceptibility to external factors, the raw measurement data often contains strong noise, environmental drift, and nonlinear disturbances, leading to a significant decrease in the accuracy of subsequent data analysis and spatial inversion. In existing research, commonly used magnetic field signal processing methods include bandpass filtering, Fourier transform, wavelet denoising, and empirical mode decomposition. These methods can effectively suppress noise under static or low-interference conditions. However, in complex environments, the time-frequency characteristics of magnetic field signals change dynamically over time, making it difficult for a single method to simultaneously achieve noise suppression and feature preservation. Fixed-threshold wavelet denoising performs well under low-noise conditions, but it is prone to over-smoothing or under-filtering in scenarios with abrupt changes in interference, resulting in residual anomalous segments or signal distortion. Furthermore, traditional signal analysis methods mostly rely on manually set parameters, making it impossible to automatically identify and model the implicit nonlinear temporal characteristics in magnetic field signals, thus hindering accurate identification and adaptive correction of anomalous signals.

[0003] To address the aforementioned issues, artificial intelligence and adaptive signal analysis techniques have been increasingly introduced into the field of geomagnetic measurement in recent years. By incorporating deep learning models, the time dependence and pattern changes of magnetic field signals can be captured in a multi-dimensional feature space, enabling automatic detection of anomalous fluctuations. Combined with wavelet thresholding algorithms, noise and effective magnetic signals can be separated in a multi-scale space, thus completing signal correction and reconstruction. While existing research has verified the potential of deep learning in anomaly detection, problems remain, including insufficient model generalization, low feature fusion efficiency, and limited reconstruction accuracy. Therefore, there is an urgent need for a comprehensive method that integrates the advantages of deep learning feature extraction and adaptive wavelet thresholding correction to achieve high-precision anomaly identification and adaptive noise correction of surface magnetic field measurement data, providing reliable data support for underground cable path inversion, operational status analysis, and 3D positioning. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe several preferred embodiments. In this section, as well as in the abstract and title of this specification, some content may be simplified or omitted to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be considered as limitations on the scope of protection of this invention.

[0005] In view of the problems existing in the above and / or prior art, the present invention is proposed.

[0006] The technical problem to be solved by this invention is: to address the issues of surface magnetic field measurement signals being easily affected by external interference, significant noise, difficulty in identifying abnormal data, and insufficient correction accuracy, a method for identifying and correcting abnormal data in surface magnetic field measurements that integrates deep learning and adaptive wavelet thresholding is proposed, so as to achieve intelligent detection and high-precision recovery of magnetic field signals.

[0007] To address this problem, the first aspect of this invention provides a method for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm. The method includes the following steps: S1. Use a magnetic field sensor to collect the original magnetic field timing signal generated by the surface cable; S2. Construct a deep learning anomaly detection model to extract features and identify anomalies in the collected magnetic field time series data, and mark the abnormal data segments that are disturbed by external factors. S3. An adaptive wavelet threshold denoising algorithm is used to perform multi-scale decomposition, threshold quantization and wavelet reconstruction on the identified abnormal data segment signals to achieve the correction and recovery of abnormal signals. S4. The original signal and the corrected signal are fused together to output the denoised and reconstructed magnetic field data.

[0008] Furthermore, step S1 includes: A high-sensitivity triaxial magnetic field sensor array is deployed in the area where surface cables operate to collect power frequency or carrier magnetic field signals generated by the current in the cable conductors. The sensor array is arranged at predetermined spatial intervals, and the sensor sampling frequency is not less than 1kHz to obtain the time series of the three components of the magnetic field. B x ( t ) 、B y ( t )and B z ( t ); The original signal is preprocessed, the preprocessing including: A unified time series is constructed by removing DC bias, using a bandpass filtering algorithm to retain the main frequency component, and interpolating and aligning multi-point measurement signals.

[0009] Furthermore, step S2 includes: A deep learning model based on a joint structure of convolutional neural network and long short-term memory network is used to perform multidimensional feature mapping and time-series dependency modeling on magnetic field time-series signals. The model predicts the magnetic field strength vector at each time t by inputting a signal sample sequence with a sliding window length of T, thus obtaining a prediction sequence. and the actual measurement sequence Error comparison is performed, and the goodness of fit of the model is determined by the coefficient of determination. Characterization, used to evaluate the model's ability to interpret the variance of the observed signal, is calculated using the following formula:

[0010] in, The actual observed magnetic field value at time t; These are the model's predicted values; is the signal mean; RMSE is the root mean square error, used to characterize the magnitude of the prediction residual; Var is the observed signal variance, used to characterize the overall fluctuation of the magnetic field signal; like The decrease exceeds the threshold Δ within a continuous time window If so, the magnetic field signal during that period is determined to be in an abnormal range.

[0011] Furthermore, in step S3, the abnormal data segment signal... Perform J-level wavelet decomposition to obtain approximate coefficients. and detail coefficient :

[0012] in, For scaling function, This is the form of the mother wavelet function after scaling and translation. For the j-th layer, these are the detail coefficients, reflecting the high-frequency disturbance characteristics of the signal. The number of decomposition layers, This is a time-shift index.

[0013] Furthermore, in step S3, an adaptive thresholding process is applied to the detail coefficients of each layer. Layer threshold Defined as:

[0014] in, Let j be the standard deviation of the noise at layer j. For the number of coefficients in this layer, when Set to zero when the threshold is reached, otherwise shrink according to the soft threshold. The coefficients after thresholding are... The correction signal is obtained by reconstructing the signal using inverse wavelet transform. :

[0015] in, The corrected signal after wavelet reconstruction. represents the approximation coefficients of the J-th layer, corresponding to the low-frequency principal components of the signal.

[0016] Furthermore, in step S4, the corrected signal and the original signal are weighted and fused to obtain high signal-to-noise ratio output magnetic field data. The fusion result... Represented as:

[0017] in, The corrected signal after wavelet reconstruction. This is the original magnetic field signal. The fusion coefficients are adaptively adjusted based on the anomaly confidence level. When the signal anomaly level is high, the weight of the correction signal is increased; when the signal is close to normal, more original components are retained.

[0018] Secondly, the present invention also provides a system for identifying and correcting abnormal data of surface magnetic field measurement based on a deep learning and adaptive wavelet threshold fusion algorithm. The system includes a surface magnetic field data acquisition unit, a deep learning anomaly detection unit, an adaptive wavelet threshold correction unit, and a data fusion output unit. The surface magnetic field data acquisition unit is used to acquire the magnetic field time-series signal of the surface cable area using a magnetic field sensor array, and to filter and synchronize the raw data. The deep learning anomaly detection unit is used to extract temporal features based on convolutional neural network and long short-term memory network models, identify anomalies in magnetic field signals, and generate anomaly segment markers. The adaptive wavelet threshold correction unit is used to perform multi-scale wavelet decomposition and adaptive threshold denoising on abnormal data segments to achieve correction and recovery of magnetic field signals. The data fusion output unit is used to perform weighted fusion of the original signal and the correction signal to output magnetic field data with a high signal-to-noise ratio.

[0019] A third aspect of the present invention provides a computer device, the computer device including a memory, a processor, and a transceiver, which are connected to each other via a bus; the memory is used to store a set of computer program instructions and data, and to transmit the stored data to the processor, the processor executing the program instructions stored in the memory to perform the method as described in the first aspect. A fourth aspect of the present invention provides a computer device, the computer device including a memory, a processor, and a transceiver, which are connected to each other via a bus; the memory is used to store a set of computer program instructions and data, and to transmit the stored data to the processor, the processor executing the program instructions stored in the memory to perform the method described in the first aspect. Compared with the prior art, the beneficial effects achieved by the present invention are as follows: This invention, through a fusion algorithm, can effectively eliminate abnormal observations caused by external interference, retain true cable signals, significantly improve the signal-to-noise ratio and reliability of surface magnetic field measurement data, and provide high-precision input data for subsequent cable path inversion and three-dimensional positioning.

[0020] This invention overcomes the problems of misjudgment and decreased positioning accuracy of traditional magnetic field measurement methods under strong interference environments, and improves the accuracy of abnormal data identification. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a method for identifying and correcting abnormal surface magnetic field measurement data based on a deep learning and adaptive wavelet threshold fusion algorithm proposed in this invention.

[0022] Figure 2 This is a flowchart of the deep learning anomaly detection process for a surface magnetic field measurement anomaly data identification and correction method based on deep learning and adaptive wavelet threshold fusion algorithm proposed in this invention.

[0023] Figure 3 This is a flowchart of the adaptive wavelet threshold correction and reconstruction process for an adaptive wavelet threshold correction method for identifying and correcting surface magnetic field measurement anomaly data based on deep learning and adaptive wavelet threshold fusion algorithm proposed in this invention.

[0024] To more clearly illustrate the technical solutions of the present invention, the accompanying drawings used in the present invention will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Detailed Implementation

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0027] Secondly, the present invention will be described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure will be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.

[0028] Furthermore, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0029] This invention provides a method for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm. Compared with traditional algorithms, it has higher anomaly identification accuracy and signal correction effect, and can effectively improve the reliability and positioning accuracy of surface magnetic field data.

[0030] Specifically: The present invention provides a method for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm, comprising the following steps: Earth's surface magnetic field data acquisition and preprocessing.

[0031] A high-sensitivity triaxial magnetic field sensor array is deployed in the area where surface cables operate to collect power frequency or carrier magnetic field signals generated by the current in the cable conductors. Each sensor can simultaneously measure the components of the magnetic field in three orthogonal directions (x, y, z), denoted as [x, y, z]. B x (t), B y (t), B z (t) This allows us to obtain complete vector information about the Earth's surface magnetic field.

[0032] The placement of the sensor array is planned based on the cable routing, burial depth, and environmental noise distribution. The spacing between measuring points is generally controlled within the range of 0.5 to 2 meters to ensure the spatial resolution of the magnetic field. Each measuring point is connected to the data acquisition terminal through a time synchronization module to achieve multi-point synchronous sampling. The sampling frequency is not less than 1 kHz to ensure the continuity of time-series data and time-domain resolution.

[0033] After the signal acquisition is completed, the system first performs preprocessing operations on the raw data, including the following three steps: (1) DC bias removal: eliminate the influence of sensor zero drift and static components of geomagnetic field, so that the signal center is stabilized at zero mean; (2) Bandpass filtering: use a bandpass filter to retain the main frequency component of cable current, such as power frequency 50Hz or carrier 9–12kHz, and filter out low frequency geomagnetic changes and high frequency industrial noise; (3) Time synchronization and interpolation alignment: perform timestamp correction and data interpolation on multi-point sampling signals to ensure that all measurement points are aligned under the same time reference.

[0034] The preprocessed magnetic field signal possesses a high signal-to-noise ratio and temporal continuity, effectively reflecting the dynamic characteristics of the magnetic field formed by the cable conductor current on the ground surface. The system integrates the data into a multi-channel input matrix based on time sequence and measurement point numbering.

[0035] Anomaly Detection Based on Deep Learning After completing the acquisition and preprocessing of the surface magnetic field signal, the multi-channel time-series sample composed of three-axis components (with a sliding window length set to...) is then... The input is fed into a CNN+LSTM spatiotemporal fusion model. The convolutional layers use multiple sets of one-dimensional convolutions and pooling to scan short-term local structures, extracting perturbation patterns and suppressing high-frequency noise. The time-series modeling layer uses a multi-layer LSTM to characterize long-term dependencies and gradual background changes, avoiding gradient decay in conventional RNNs. The feature fusion layer concatenates and weights spatial and temporal features before feeding them into the output layer for sequence prediction, resulting in the predicted sequence. and the measured sequence Comparison is used to measure anomalies.

[0036] Anomaly measurement uses the coefficient of determination Two indicators: RMSE

[0037] in, , when A value close to 1 indicates that the model can explain the observation variance and the signal is in a normal state; if sudden interference or sensor drift occurs, the residuals will increase. Reduce. To enhance sensitivity to gradual anomalies, the decrease in the determination coefficient of adjacent sliding windows is defined.

[0038] And monitor its continuous The cumulative changes within each sliding window.

[0039] When the following conditions are met:

[0040] The corresponding time period will be marked as abnormal; this rule also covers impulse interference (short window error spike) and slow drift ( The system outputs an anomaly marker matrix (start and end times, measurement point number, and confidence level) for cases such as continuous positive values ​​and sudden data loss. The confidence level is then used to summarize the anomaly markers. The rate of decrease and the magnitude of RMSE are correlated, and the matrix dimensions are aligned with the original data for easy tracking and subsequent processing; at the same time, error distribution and confidence curves are generated for visualization.

[0041] Adaptive wavelet threshold correction and reconstruction Anomaly label matrix output by deep learning module In the middle, if at a certain moment A value of 1 indicates that the magnetic field data at that moment is considered abnormal. To avoid being sensitive to single-point noise, adjacent samples with a time interval not exceeding one sampling period are used. The "1"s are merged according to their continuity to obtain a set of abnormal time periods. Record the start and end indices and span of each segment. The minimum duration can be set. This can be used to filter out isolated pulses; morphological closing operations can also be introduced at boundary locations to smooth the label sequence, reducing the impact of sporadic mislabeling on subsequent processing. This applies to any abnormal interval in the set. From the preprocessed triaxial magnetic field signal Extract the components or principal components that need to be repaired to form a one-dimensional signal to be corrected. .in These represent the instantaneous amplitudes of the Earth's surface magnetic field in three orthogonal directions. This is a discrete sampling time index, corresponding to the sampling frequency. The determination of the target channel follows the principle of "prioritizing dominant components and supplementing with robustness": when the anomaly manifests as a single-axis spike, step, or short-term drift, the corresponding axis is directly selected as the target channel. When triaxial coupling is significant or the dominant component is unclear, the energy dominance criterion is adopted.

[0042] This ensures that the selected signal carries more effective information. If there is component aliasing due to phase asynchrony or attitude changes in the field, vector synthesis can be used.

[0043] To improve robustness to phase perturbations and directional noise, and to enhance the stability of subsequent decomposition and reconstruction, context is typically introduced by extending the sampling points by δ (δ∈[16,64] adjustable) at both ends of each anomalous segment; and zero-mean and amplitude normalization is performed on the signal within this extended interval.

[0044] To eliminate the bias in threshold estimation caused by amplitude scale differences, while retaining the original timestamps and measurement point numbers, it facilitates accurate backfilling and global stitching of the reconstruction results. For the standardized... Discrete wavelet decomposition is performed to extract its time-frequency features in multi-scale space. Let the mother wavelet be... The scaling function is Then the scale of the j-th layer and the basis functions after translation are defined as follows:

[0045] Where j represents the decomposition level (frequency scale) and k represents the shift index (time position). The decomposition expression for the signal is written as:

[0046] in For the first Layer approximation coefficient, For the first Layer detail coefficients (local high-frequency, spike, and transient characteristics). The mother wavelet is preferably db8 to balance tight support and orthogonality; symmetrical extension is used for boundary treatment to suppress endpoint ringing; layer decomposition. It can be set based on the empirical relationship between the sampling frequency and the target minimum effective frequency.

[0047] And combined with the validation set Perform a small amount of grid search to strike a balance between anomaly detection capability and reconstruction smoothness. If the anomaly segment is short and contains high-frequency transients, it is recommended to appropriately reduce the grid search speed. To avoid over-smoothing; if a noticeable slow drift is observed, the smoothing level can be moderately increased. To enhance low-frequency characterization capabilities.

[0048] Within the wavelet domain, to handle random noise and transient interference at different scales, the process consists of three steps. First, robust noise estimation is performed: for each layer of detail coefficients, robust statistics such as median absolute deviation are used to measure the noise level of that layer. During estimation, statistics are preferentially calculated in the context intervals at both ends of the outlier segment, and cross-referenced with the results within the segment; when the difference is too large, context estimation takes precedence to avoid being biased by outliers. For high-frequency layers with few sample points, slight smoothing from adjacent time blocks or adjacent scales can be introduced to reduce small sample variance; if narrowband fixed interference exists in the field, narrowband suppression can be performed before entering the wavelet domain to avoid misinterpreting interpretable narrowband components as noise. After obtaining the noise levels of each layer, a general threshold that is proportional to the noise intensity and increases slowly with the scale of the coefficients in that layer is constructed.

[0049] At a certain scale, when the overall energy of the coefficients is low, smaller coefficients are more likely to consist mainly of noise, and the threshold will be smaller to avoid over-suppression; when the overall energy is high, the threshold will be raised accordingly, thus preserving more effective details. From the aforementioned robust estimate, This represents the number of coefficients for that layer. Finally, a slight scale adjustment is made to the threshold: the high-frequency layer, biased towards cleaning up glitches and random noise, can be appropriately strengthened; the low-frequency layer, bearing the background and slow trends, should be conservative; the mid-frequency layer represents a compromise between the two. In engineering practice, a scale weight within a narrow range can be used to gently adjust the threshold, making the strategy slightly stronger at high frequencies and slightly weaker at low frequencies, but avoiding excessive fluctuations that could lead to instability across scenarios. The operating strategy is as follows: first, use the above thresholds to quickly provide consistent denoising results; if the variance ratio, residual energy ratio, or joint indicators with upstream anomaly confidence do not meet the standards, then make a one-time fine-tuning of the weights for individual scales. The overall computational load remains linear; combined with block processing and cache statistics, the additional latency of a single segment can be controlled within one or two window steps, meeting the real-time requirements of online repair and subsequent fusion.

[0050] After setting the thresholds for each scale, soft thresholding is applied to the detail coefficients to achieve a balance between noise suppression and waveform continuity. The core idea is to directly reduce the coefficient amplitude to zero when it is below the threshold, and to subtract an amplitude equivalent to the threshold while retaining the original positive or negative sign when the amplitude exceeds the threshold. This cleans up random glitches while avoiding energy abrupt changes and step artifacts. At mid-to-high frequency scales, this significantly reduces background noise without smoothing out effective transients, while at low frequency scales, it maintains background trends and physical consistency as much as possible. The thresholding rule is as follows:

[0051] in This represents the original detail coefficient value at the k-th position of the j-th layer. This indicates the magnitude of the coefficient. This represents the result of the detail coefficients after shrinkage. This indicates that the effective threshold value of the j-th layer has been adjusted by combining the noise intensity and coefficient size of that layer with a gentle scale correction. The larger the value of the scale level in wavelet decomposition, the lower the corresponding frequency. This represents the time position index at a given scale. The approximation coefficients remain unchanged to maintain the overall energy level and the authenticity of slow drift. When there is significant low-frequency drift in the data, only the detail coefficients at the lowest few scales can be gently contracted to avoid weakening the background trend. When the anomalous segment is very short and contains small transients, the kurtosis and local variance of that scale are checked first. If it indicates that the signal component accounts for a high proportion, the contraction intensity of that scale is appropriately relaxed. After completing the layer-by-layer contraction, the data is returned to the time domain according to the predetermined inverse transform. The reconstructed segment and the original segment are connected by symmetrical extension and smooth transition to ensure the continuity of amplitude and first derivative. Finally, the results are seamlessly backfilled into the global sequence based on the timestamp.

[0052] After thresholding at each scale, the corrected detail coefficients are recombinated with the invariant approximation coefficients and an inverse discrete wavelet transform is performed to obtain the corrected time-domain signal within the anomaly segment.

[0053] This is the correction signal at time t; This is the k-th approximation coefficient of the J-th layer; Let be the value of the k-th scaling function in the J-th layer at time t; represents the detail coefficient of the kth layer after soft thresholding; Let be the value of the k-th wavelet function at the j-th level at time t; t is the discrete-time index; J is the top-level scale; j is the scale level number, with larger values ​​corresponding to lower frequencies; and k is the time position index at the given scale. During the reconstruction stage, a symmetrical extension boundary strategy is used to suppress endpoint pseudo-ringing and aliasing, and appropriate context is retained at both ends of the anomalous segments to improve reconstruction stability and the smoothness of subsequent stitching. Adjacent segments use cosine-shaped smoothing weights in the overlapping area to achieve seamless stitching; the weight of the left segment gradually decreases from one to zero, while the weight of the right segment gradually increases from zero to one, thus achieving energy continuity and shape smoothness. Before entering the wavelet domain, the segments have been normalized to zero mean and amplitude. After reconstruction, the original dimensions need to be restored using...

[0054] This is the corrected signal after dimensional recovery at time t; This represents the standard deviation of the amplitude of the segment. This is the mean of the segment. After restoring the dimensions, it is calculated based on the original timestamp. Accurately backfill to the corresponding position in the global sequence; when there are multiple consecutive anomalies, first complete the phase and energy level alignment in the short overlap region, and then complete the cascade splicing in sequence to avoid perceptible step and breathing effects.

[0055] To quantify the denoising effect and fidelity, the residual and variance ratio are defined as follows:

[0056] For at any time The residual signal; For at any time The original signal to be corrected; The variance is the ratio of the residual variance to the original signal variance; Var represents the variance calculated on the discrete sequence. Engineering criteria are based on... The threshold is lower than the preset threshold, and the cross-verification of the error energy ratio and the upstream anomaly confidence level serves as the basis for compliance. If the threshold is not met, a slight adjustment is made to the aforementioned threshold strength or splicing window width, and then the structure is reconstructed again. This achieves a better trade-off between noise suppression and detail fidelity. The entire process maintains linear computational complexity. By combining block processing and cache statistics, the additional latency of a single segment can be controlled within two window steps, meeting the real-time requirements of subsequent fusion weighting and full-process closed loop.

[0057] Signal fusion and high signal-to-noise ratio output This step focuses on online status awareness and emergency response for regional power distribution networks. The goal is to integrate the reconstructed sequence obtained in the previous stage with the original field measurements point-by-point under conditions of rapid fluctuations in photovoltaic output and external electromagnetic interference. This results in a physically consistent, boundary-continuous, and traceable high signal-to-noise ratio (SNR) surface magnetic field sequence, providing reproducible evaluation and online thresholds consistent with those of the dispatch center. The core idea is to map the discriminative prior and reconstruction consistency to the same scale using anomaly intensity as the horizontal axis and adaptive weights as the vertical axis. Then, a convex combination is used to achieve a compromise between the two signals, and short-window time-domain processing ensures smooth segment splicing and phase continuity.

[0058] Constructing anomaly confidence sequences at time index t

[0059] Used to measure the intensity of the anomaly at time t. This represents the point-level anomaly probability from the upstream model. This refers to the residual signal obtained by reconstructing the signal from the previous stage as a reference. It is a numerical stability term used to suppress instability caused by minimal residuals. and This is the counterweight coefficient, and the sum of the two is one. During the daytime photovoltaic rise and fall phases, the residuals better reflect structural abrupt changes, therefore, it can be... Slightly higher to increase response to abnormal changes; during nighttime periods when the continuity of power supply to critical users is paramount, prior knowledge is more important, therefore, it can be... Slightly higher to maintain a conservative strategy. To improve the generalization across days and weeks, one can first... Perform probability calibration, and then search on the development set using receiver operating characteristics or precision-recall curves. and A reasonable combination.

[0060] Mapping anomaly confidence scores to fusion weights

[0061] This represents the confidence level of the reconstructed component at time t, with a value ranging from zero to one. Control the speed at which the original component is switched to the reconstructed component. Control the inflection point from normal to abnormal. Based on the power grid's safety preferences, select a larger value for the sensitive periods of key users. Prioritize preserving original observations, and select periods with significant fluctuations in new energy output for smaller periods. To enable faster reliance on reconstructed components; to reduce temporal jitter under strong interference, it is possible to... Perform light smoothing or sizing at two to three sampling points. An exponential average is performed to obtain a more stable response without significantly increasing latency. This mapping is mathematically bounded and monotonic, avoiding energy jumps introduced by abrupt weight changes and mitigating sensitivity to downstream thresholds. After weight calculation, linear fusion of the two routes is executed.

[0062] The fused output at time t The wavelet threshold reconstruction result at time t has been obtained, and the dimensions have been restored according to the segment mean and segment standard deviation. This is the original measurement at time t. This convex combination guarantees... It remains within the energy envelope of both the original and reconstructed signals, neither overly relying on data from either side, and can quickly introduce correction to improve the signal-to-noise ratio when abnormal increases occur. When the sensor approaches saturation, it can... A soft upper limit is applied in conjunction with amplitude clipping to avoid flat-top distortion affecting subsequent interpretations. To comply with causal constraints, all calculations use only current and historical information, allowing the results to be directly fed into the emergency command chain.

[0063] To achieve a perceptible smooth transition at the boundaries of the station area and the junction of segments, A short-window time-domain smoothing is applied. Normalized cosine weights are used.

[0064] in This represents the window weight of the nth sampling point in the overlapping region. This represents the total number of samples in the overlapping region. The window length is proportional to the dominant frequency period, not less than one-quarter of the dominant frequency period to eliminate visible steps, and not greater than one-half of the dominant frequency period to preserve edge details. When adjacent abnormal segments appear consecutively, least squares is first used in the overlapping region to align the energy levels and phases, and then complementary weights are used to connect them, thereby maintaining continuity in terms of amplitude and first derivative. This will significantly reduce the misjudgment of manual inspection and reduce the sensitivity of the automatic threshold to boundary spikes.

[0065] To maintain consistency with the acceptance criteria of the scheduling side, the output quality is measured using variance ratio as the core metric.

[0066] here This represents the ratio of residual energy to original energy; a smaller value indicates more effective noise suppression without destroying the principal components. Represents the original sequence. This represents the fused sequence. To explicitly diagnose frequency domain bias, it reports the bandwidth energy percentage to identify over-smoothing and insufficient denoising, and reports the correlation improvement with the reference sequence to measure the retention of structural information. The reference sequence can be a calibrated physical model output or redundant measurement points. A minimum threshold of [specified value] is recommended. The threshold must be no higher than 0.2 and the reference relevance must be improved by no less than 0.1 compared to before fusion; if the threshold is not met, a minor adjustment will be made to meet the real-time constraints. Small adjustments are made to either lean towards the original or towards reconstruction. Make small adjustments to change the switching speed, while increasing or decreasing the smoothing window length by one or two sampling points, and then recalculate. and The re-evaluation was completed. All parameter changes were recorded with timestamps and measurement point numbers to ensure traceability and compliance. This process binds the upstream decision confidence level and the reconstructed consistency metric to the same anomaly intensity scale. Through adaptive weights The convex combination achieves a real-time trade-off between wavelet reconstruction and original observations at a single-point granularity, ensuring that the output maintains the continuity of the geomagnetic dominant frequency and phase while significantly improving the signal-to-noise ratio. Short-window smoothing and overlapping stitching ensure boundary continuity and reduce the abruptness between the visual and algorithmic layers. Quality evaluation and threshold closure enable the results to be readily usable for 3D scene reconstruction, abnormal path inversion, and high-priority alarms. The computational complexity of the entire link increases linearly with the time length. Combined with block segmentation and cache statistics, the end-to-end additional latency can be controlled within two window steps, meeting the real-time requirements of online analysis and emergency command.

[0067] Experimental results demonstrate that the fused output signal exhibits higher robustness and interpretability compared to the results obtained using deep learning detection or wavelet denoising alone. Statistical results show that the average signal-to-noise ratio (SNR) of the fused signal is improved by 8–10 dB, and the Pearson correlation coefficient with the ideal magnetic field reference signal is improved by approximately 0.1–0.15, reflecting a dual optimization effect at both the energy and structural levels. Simultaneously, the fusion algorithm significantly suppresses high-frequency noise components and random interference while maintaining the amplitude deviation of the dominant frequency component within 2%, achieving high-fidelity signal reconstruction. The system completes a closed-loop optimization process from anomaly identification and signal correction to adaptive fusion. Output signal It combines high signal-to-noise ratio, smooth continuity and physical consistency, and can be directly used in scenarios such as three-dimensional imaging of the Earth's surface magnetic field, cable path inversion, anomaly source location and geological structure analysis.

[0068] This method not only improves data quality and algorithm robustness, but also has engineering advantages such as adjustable algorithm parameters, low computational complexity, and suitability for real-time deployment, providing a reliable technical approach for high-precision acquisition of geomagnetic data in complex electromagnetic environments.

[0069] A system for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm is provided. The system includes a surface magnetic field data acquisition unit, a deep learning anomaly detection unit, an adaptive wavelet threshold correction unit, and a data fusion output unit. The surface magnetic field data acquisition unit is used to acquire the magnetic field time-series signal of the surface cable area using a magnetic field sensor array, and to filter and synchronize the raw data. The deep learning anomaly detection unit is used to extract temporal features based on convolutional neural network and long short-term memory network models, identify anomalies in magnetic field signals, and generate anomaly segment markers. The adaptive wavelet threshold correction unit is used to perform multi-scale wavelet decomposition and adaptive threshold denoising on abnormal data segments to achieve correction and recovery of magnetic field signals. The data fusion output unit is used to perform weighted fusion of the original signal and the correction signal to output magnetic field data with a high signal-to-noise ratio.

[0070] A computer device, characterized in that the computer device includes a memory, a processor, and a transceiver, which are connected to each other via a bus; the memory is used to store a set of computer program instructions and data, and to transmit the stored data to the processor; the processor executes the program instructions stored in the memory to perform the aforementioned method for identifying and correcting abnormal data of surface magnetic field measurement based on deep learning and adaptive wavelet threshold fusion algorithm.

[0071] A computer device includes a memory, a processor, and a transceiver connected via a bus. The memory stores a set of computer program instructions and data, and transmits the stored data to the processor. The processor executes the program instructions stored in the memory to perform the aforementioned method for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying and correcting abnormal data of ground magnetic field measurement based on deep learning and adaptive wavelet threshold fusion algorithm, characterized in that, The method includes the following steps: S1. Use a magnetic field sensor to collect the original magnetic field timing signal generated by the surface cable; S2. Construct a deep learning anomaly detection model to extract features and identify anomalies in the collected magnetic field time series data, and mark the abnormal data segments that are disturbed by external factors. S3. An adaptive wavelet threshold denoising algorithm is used to perform multi-scale decomposition, threshold quantization and wavelet reconstruction on the identified abnormal data segment signals to achieve the correction and recovery of abnormal signals. S4. The original signal and the corrected signal are fused together to output the denoised and reconstructed magnetic field data.

2. The method according to claim 1, wherein, Step S1 includes: In the surface cable operation area, a high sensitivity triaxial magnetic field sensor array is arranged to collect the power frequency or carrier magnetic field signals generated by the cable conductor current, the sensor array is arranged according to a predetermined spatial interval, and the sensor sampling frequency is not less than 1 kHz to obtain the time series of three components of the magnetic field B x ( t ) 、B y ( t ) and B z ( t ) The original signal is preprocessed, and the preprocessing includes: A unified time series is constructed by removing DC bias, using a bandpass filtering algorithm to retain the main frequency component, and interpolating and aligning multi-point measurement signals.

3. The method for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm as described in claim 1, characterized in that, Step S2 includes: A deep learning model based on a joint structure of convolutional neural network and long short-term memory network is used to perform multidimensional feature mapping and time-series dependency modeling on magnetic field time-series signals. The model predicts the magnetic field strength vector at each time t by inputting a signal sample sequence with a sliding window length of T, thus obtaining a prediction sequence. and the actual measurement sequence Error comparison is performed, and the goodness of fit of the model is determined by the coefficient of determination. Characterization, used to evaluate the model's ability to interpret the variance of the observed signal, is calculated using the following formula: in, The actual observed magnetic field value at time t; These are the model's predicted values; is the signal mean; RMSE is the root mean square error, used to characterize the magnitude of the prediction residual; Var is the observed signal variance, used to characterize the overall fluctuation of the magnetic field signal; like The decrease exceeds the threshold Δ within a continuous time window If so, the magnetic field signal during that period is determined to be in an abnormal range.

4. The method for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm as described in claim 1, characterized in that, In step S3, the abnormal data segment signal is processed. Perform J-level wavelet decomposition to obtain approximate coefficients. and detail coefficient : in, For scaling function, This is the form of the mother wavelet function after scaling and translation. For the j-th layer, these are the detail coefficients, reflecting the high-frequency disturbance characteristics of the signal. The number of decomposition layers, This is a time-shift index.

5. The method for identifying and correcting anomaly data in surface magnetic field measurement based on a deep learning and adaptive wavelet threshold fusion algorithm according to claim 4, characterized in that, In step S3, adaptive thresholding is applied to the detail coefficients of each layer. Layer threshold Defined as: in, Let j be the standard deviation of the noise at layer j. For the number of coefficients in this layer, when Set to zero when the threshold is reached, otherwise shrink according to the soft threshold. The coefficients after thresholding are... The correction signal is obtained by reconstructing the signal using inverse wavelet transform. : in, This is the corrected signal after wavelet reconstruction. represents the approximation coefficients of the J-th layer, corresponding to the low-frequency principal components of the signal.

6. The method for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm as described in claim 1, characterized in that, In step S4, the correction signal and the original signal are weighted and fused to obtain high signal-to-noise ratio output magnetic field data. The fusion result is... Represented as: in, This is the corrected signal after wavelet reconstruction. This is the original magnetic field signal. The fusion coefficients are adaptively adjusted based on the anomaly confidence level. When the signal anomaly level is high, the weight of the correction signal is increased; when the signal is close to normal, more original components are retained.

7. A system for identifying and correcting anomaly data in surface magnetic field measurements based on a deep learning and adaptive wavelet threshold fusion algorithm, characterized in that, The system includes a surface magnetic field data acquisition unit, a deep learning anomaly detection unit, an adaptive wavelet threshold correction unit, and a data fusion output unit. The surface magnetic field data acquisition unit is used to acquire the magnetic field time-series signal of the surface cable area using a magnetic field sensor array, and to filter and synchronize the raw data. The deep learning anomaly detection unit is used to extract temporal features based on convolutional neural network and long short-term memory network models, identify anomalies in magnetic field signals, and generate anomaly segment markers. The adaptive wavelet threshold correction unit is used to perform multi-scale wavelet decomposition and adaptive threshold denoising on abnormal data segments to achieve correction and recovery of magnetic field signals. The data fusion output unit is used to perform weighted fusion of the original signal and the correction signal to output magnetic field data with a high signal-to-noise ratio.

8. A computer device, characterized in that, The computer device includes a memory, a processor, and a transceiver connected to each other via a bus; the memory stores a set of computer program instructions and data, and transmits the stored data to the processor, which executes the program instructions stored in the memory to perform the method as described in any one of claims 1 to 6.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a transceiver connected to each other via a bus; the memory stores a set of computer program instructions and data, and transmits the stored data to the processor, which executes the program instructions stored in the memory to perform the method as described in any one of claims 1 to 6.

Images