Intelligent identification method for magnetic field disturbance mode based on convolutional network

Abstract

The invention discloses a magnetic field disturbance mode intelligent identification method based on a convolutional network, and the method comprises the following steps: collecting magnetic field time sequence data, and carrying out the preprocessing to generate a multi-axis magnetic field time sequence data set; extracting a baseline component, and generating disturbance residual data; extracting a frequency position, a frequency bandwidth and a time stability parameter, and constructing an interference fingerprint; performing interference suppression processing according to the interference fingerprint to generate compensation residual data; constructing three-channel input data based on the baseline component, the disturbance residual data and the compensation residual data; inputting the three-channel input data into the improved SincNet to generate multi-scale disturbance characteristic representation; three-axis cross coupling processing is executed, and disturbance characteristic representation is generated; and generating a magnetic field disturbance mode recognition result based on the disturbance characteristic representation. Through multichannel magnetic field feature construction and an improved SincNet multi-scale coupling identification method, stable identification of a magnetic field disturbance mode is realized, and the method has the advantages of high interference resistance and high cross-axis consistency.

Description

Technical Field

[0001] This invention relates to the field of intelligent recognition of magnetic field perturbation patterns, and in particular to an intelligent recognition method for magnetic field perturbation patterns based on convolutional networks. Background Technology

[0002] Existing magnetic field disturbance identification technologies typically rely on frequency domain analysis, time domain feature extraction, or pattern classification processing based on general convolutional networks to identify magnetic field anomalies or disturbance types from time-series signals acquired by single-axis or multi-axis magnetic field sensors. These methods often employ fixed filters or unified feature extraction structures to suppress or model power frequency harmonics, stable narrowband electromagnetic interference, and environmental background changes in magnetic field signals. They also use single-time-scale or simple multi-scale features to distinguish disturbance patterns and have been applied in some scenarios.

[0003] Existing technologies generally suffer from insufficient ability to distinguish between slowly varying magnetic field backgrounds and weak disturbance signals, and poor adaptability to differences in magnetic field response across different spatial axes. In particular, when there is superposition of multi-source interference and inconsistent sensor mounting orientation, disturbance features are easily mixed and pattern discrimination is reduced. At the same time, existing methods often fail to effectively characterize the interaction between multi-timescale information and spatial axes during feature extraction and discrimination, making it difficult to form consistent disturbance feature representations across scales and axes, thus affecting the accuracy and stability of magnetic field disturbance pattern recognition. Summary of the Invention

[0004] One objective of this invention is to propose an intelligent identification method for magnetic field perturbation patterns based on convolutional networks. This invention achieves stable identification of magnetic field perturbation patterns by constructing and improving the SincNet multi-scale coupling identification method using multi-channel magnetic field features, and has the advantages of strong anti-interference and high cross-axis consistency.

[0005] A method for intelligent recognition of magnetic field perturbation patterns based on convolutional networks according to an embodiment of the present invention includes the following steps: Collect magnetic field time series data of magnetic field sensor in multiple spatial axes, and preprocess to generate multi-axis magnetic field time series dataset; Within a preset time window, the multi-axis magnetic field time series dataset is decomposed, the baseline component is extracted, and the difference between the multi-axis magnetic field time series dataset and the baseline component is calculated to generate perturbation residual data. Based on the perturbation residual data, frequency location parameters, bandwidth parameters, and time stability parameters are extracted to construct the interference fingerprint; Based on the interference fingerprint, a corresponding notch template is selected from the pre-established notch template library. The notch template is then used to suppress the interference in the disturbance residual data and generate compensated residual data. Three-channel input data is constructed based on baseline components, perturbation residual data, and compensation residual data; The three-channel input data is fed into the improved SincNet, and features are extracted from the three-channel input data at different time scales to generate multi-scale perturbation feature representations. A triaxial cross-coupling process is performed on the feature components corresponding to different spatial axes in the multi-scale perturbation feature representation to generate the perturbation feature representation; Based on the perturbation feature representation, magnetic field perturbation pattern recognition results are generated using an improved SincNet.

[0006] Optionally, the preprocessing specifically includes: time synchronization, anomaly removal, amplitude normalization, resampling, and time sequence arrangement.

[0007] Optionally, the generation of the perturbation residual data specifically includes: Within a preset time window, the multi-axis magnetic field time series dataset is extracted in chronological order to generate multi-axis magnetic field time series data within the corresponding time window; Based on the multi-axis magnetic field time series data within the corresponding time window, the magnetic field time series data of each spatial axis are decomposed to generate baseline components corresponding to the time window. Based on the baseline component, time alignment processing is performed on the multi-axis magnetic field time series data within the corresponding time window in each spatial axis to generate aligned data. Based on the aligned data, the multi-axis magnetic field time series dataset and the baseline components are subtracted along each spatial axis to generate perturbation residual data corresponding to each spatial axis. The disturbance residual data corresponding to each spatial axis are aggregated and processed to generate disturbance residual data.

[0008] Optionally, the generation of the interfering fingerprint specifically includes: The disturbance residual data is segmented according to a preset time window to generate disturbance residual segment data corresponding to each time window. The frequency scanning process is performed on the segmented data of the disturbance residual one by one, and the amplitude distribution of each frequency component is calculated within the preset frequency range to generate the frequency amplitude distribution data within the corresponding time window. In the frequency amplitude distribution data, the amplitudes of each frequency component are compared, and the frequency value corresponding to the frequency component with the largest amplitude is selected to generate the frequency location parameter. Centered on the frequency location parameter, search for frequency intervals in the frequency amplitude distribution data with continuously higher amplitude values ​​than the preset amplitude threshold in both the high-frequency and low-frequency directions to generate the bandwidth parameter; The frequency location parameters generated within multiple adjacent time windows are compared, and the changes in the frequency location parameters over time are calculated to generate time stability parameters. Interference fingerprints are generated by combining frequency location parameters, bandwidth parameters, and time stability parameters.

[0009] Optionally, the generation of the compensation residual data specifically includes: Multiple notch templates are read from a pre-established notch template library. Each of the multiple notch templates is pre-configured with a corresponding frequency position range and bandwidth range to generate a notch template set. The frequency position parameters and bandwidth parameters in the interference fingerprint are compared one by one with the frequency position range and bandwidth range of each notch template in the notch template set. The notch templates whose frequency position parameters fall into the corresponding frequency position range and whose bandwidth parameters fall into the corresponding bandwidth range are selected to generate the target notch template. Frequency domain mapping is performed on the perturbation residual data to convert it into a frequency representation. Based on the frequency location range and bandwidth range defined by the target notch template, the frequency components to be suppressed are identified in the frequency representation. The frequency components to be suppressed are subjected to amplitude attenuation processing, while the unidentified frequency components retain their original amplitudes, generating disturbance residual data after interference suppression. The disturbance residual data after interference suppression is processed by inverse frequency domain mapping to generate compensated residual data.

[0010] Optionally, the generation of the three-channel input data specifically includes: Within the same time window, the baseline components, perturbation residual data, and compensation residual data are time-aligned to generate time-aligned baseline components, perturbation residual data, and compensation residual data. Amplitude normalization is performed on the time-aligned baseline components, perturbation residual data, and compensation residual data along each spatial axis to generate normalized baseline components, perturbation residual data, and compensation residual data. According to the preset channel division order, the normalized baseline component is used as the first channel data, the normalized disturbance residual data is used as the second channel data, and the normalized compensation residual data is used as the third channel data. Under each spatial axis and time index, the corresponding first channel data, second channel data, and third channel data are combined to generate three-channel input data.

[0011] Optionally, the generation of the multi-scale perturbation feature representation specifically includes: In the improved SincNet, the three-channel input data are mapped to independent channel processing branches, and independent convolution processing paths are established in each channel processing branch. The improvements to SincNet specifically include: The input layer is improved to receive three-channel input data consisting of baseline components, perturbation residual data, and compensation residual data. In the network structure layer, independent channel processing branches are set up for each of the three channels of input data. Independent convolutional processing paths are established within each channel processing branch. Within each convolutional processing path, a hierarchical convolutional structure is constructed, consisting of parameterized bandpass convolution processing and multi-timescale convolutional processing. In the feature organization layer, the perturbation feature components generated by different channel processing branches at different timescales are concatenated to generate joint perturbation features across channels and timescales. Furthermore, scale uniformity processing is performed on the joint perturbation features to generate multi-scale perturbation feature representations. Parameterized bandpass convolution kernels are configured in the convolution processing path of each channel processing branch. The parameterized bandpass convolution kernels are used to perform bandpass convolution processing on the three-channel input data of the corresponding channel, and frequency response constraints are applied to the input data to generate channel-level initial perturbation features corresponding to each channel. Multiple convolution windows of different lengths are set within the convolution processing path of each channel processing branch. Convolution processing is performed on the initial perturbation features at the channel level at different time scales to generate perturbation feature components corresponding to the channel and time scale. Within the convolutional processing path of each channel processing branch, the perturbation feature components generated at different time scales are organized under the same time index condition to generate multi-scale perturbation feature components in each channel. The multi-scale perturbation feature components of different channel processing branches are spliced ​​in both channel and time scale dimensions to generate joint perturbation features across channels and time scales. The joint perturbation features are scale-consistentized to unify the feature dimensions corresponding to different time scales and generate multi-scale perturbation feature representations.

[0012] Optionally, the generation of the perturbation feature representation specifically includes: According to the different spatial axes corresponding to the magnetic field sensor, the multi-scale disturbance feature representation is divided to generate axial feature components corresponding to each spatial axis. Alignment is performed on the axial feature components corresponding to each spatial axis under the same time index condition to generate axial feature components that are consistent in the time dimension. Under each time index, the axial feature components corresponding to different spatial axes are combined to generate a joint axial feature that reflects the state of multiple spatial axes under the same time index. Perform cross-coupling processing on the joint axial features in the spatial axial dimension, calculate the mutual influence relationship between each spatial axial feature component, and generate the cross-coupled axial features. The axial features after cross-coupling are arranged in time index order in the time dimension to generate coupled features in the form of a continuous time series. The coupled features in the form of continuous time series are aggregated to generate perturbation feature representations.

[0013] Optionally, the generation of the magnetic field disturbance pattern recognition result specifically includes: The perturbation feature representation is input into the improved SincNet, and feature recombination processing is performed on the perturbation feature representation along the time dimension in the improved SincNet to generate temporal discriminative features; In the improved SincNet, feature mapping is performed on the temporal discriminative features at different discriminative scales to generate perturbation pattern discriminative features corresponding to different discriminative scales; Scale consistency constraint processing is performed on the perturbation pattern discrimination features under different discrimination scales to eliminate discrimination bias between different discrimination scales and generate scale-consistent perturbation pattern discrimination features; Based on scale-consistent perturbation pattern discrimination features, perturbation pattern determination processing is performed in the improved SincNet to generate magnetic field perturbation pattern recognition results.

[0014] The beneficial effects of this invention are: This invention decomposes multi-axis magnetic field time-series data to explicitly separate the slowly varying background and disturbance components of the magnetic field. Based on this, a three-channel input representation including baseline components, disturbance residual data, and compensation residual data is constructed, which fully purifies and structurally expresses the magnetic field disturbance at the data level. By introducing a collaborative mechanism between interference fingerprint and notch filter template library, this invention can achieve adaptive suppression of stable narrowband interference and power frequency harmonics, avoiding excessive attenuation of effective disturbance signals caused by traditional fixed filtering methods in complex electromagnetic environments, thereby significantly improving the detectability and feature reliability of weak magnetic field disturbances.

[0015] This invention constructs a hierarchical convolutional structure based on an improved SincNet. After completing the frequency response constraint, it introduces multi-timescale feature extraction and scale consistency processing, so that the perturbation features have good frequency selectivity and time-scale robustness. Combined with the triaxial cross-coupling processing for the spatial axis, this invention can characterize the mutual influence relationship between magnetic field perturbations of different spatial axes at the feature level, forming a perturbation feature representation with spatial consistency, effectively reducing the recognition bias caused by differences in sensor installation posture or inconsistent axial response.

[0016] Based on the above technical solution, this invention introduces a discriminant scale consistency constraint in the magnetic field disturbance pattern recognition stage, enabling discriminant features from different time scales to participate in pattern determination within a unified semantic space. This reduces the risk of misjudgment caused by multi-scale discriminant conflicts. Overall, this invention can achieve stable and accurate identification of magnetic field disturbance patterns under strong background interference and complex working conditions, significantly improving the discrimination of disturbance types and the reliability of recognition results. It also has good scene transferability and engineering feasibility, making it suitable for intelligent magnetic field disturbance recognition applications in various complex electromagnetic environments. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart of an intelligent recognition method for magnetic field perturbation patterns based on convolutional networks proposed in this invention. Figure 2 This is a schematic diagram of the improved SincNet structure for an intelligent recognition method of magnetic field perturbation patterns based on convolutional networks proposed in this invention. Figure 3 This is a schematic diagram of the multi-scale perturbation features and triaxial cross-coupling of a magnetic field perturbation pattern intelligent recognition method based on convolutional networks proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A method for intelligent recognition of magnetic field perturbation patterns based on convolutional networks includes the following steps: Collect magnetic field time series data of magnetic field sensor in multiple spatial axes, and preprocess to generate multi-axis magnetic field time series dataset; Within a preset time window, the multi-axis magnetic field time series dataset is decomposed, the baseline component is extracted, and the difference between the multi-axis magnetic field time series dataset and the baseline component is calculated to generate perturbation residual data. Based on the perturbation residual data, frequency location parameters, bandwidth parameters, and time stability parameters are extracted to construct the interference fingerprint; Based on the interference fingerprint, a corresponding notch template is selected from the pre-established notch template library. The notch template is then used to suppress the interference in the disturbance residual data and generate compensated residual data. Three-channel input data is constructed based on baseline components, perturbation residual data, and compensation residual data; The three-channel input data is fed into the improved SincNet, and features are extracted from the three-channel input data at different time scales to generate multi-scale perturbation feature representations. A triaxial cross-coupling process is performed on the feature components corresponding to different spatial axes in the multi-scale perturbation feature representation to generate the perturbation feature representation; Based on the perturbation feature representation, magnetic field perturbation pattern recognition results are generated using an improved SincNet.

[0021] In this embodiment, the preprocessing specifically includes: time synchronization, anomaly removal, amplitude normalization, resampling, and time sequence arrangement.

[0022] In this embodiment, the generation of the perturbation residual data specifically includes: Within a preset time window, the multi-axis magnetic field time series dataset is extracted in chronological order to generate multi-axis magnetic field time series data within the corresponding time window; Based on the multi-axis magnetic field time series data within the corresponding time window, the magnetic field time series data of each spatial axis are decomposed to generate baseline components corresponding to the time window. Based on the baseline component, time alignment processing is performed on the multi-axis magnetic field time series data within the corresponding time window in each spatial axis to generate aligned data. Based on the aligned data, the multi-axis magnetic field time series dataset and the baseline components are subtracted along each spatial axis to generate perturbation residual data corresponding to each spatial axis. The disturbance residual data corresponding to each spatial axis are aggregated and processed to generate disturbance residual data.

[0023] In this embodiment, the generation of the interference fingerprint specifically includes: The disturbance residual data is segmented according to a preset time window to generate disturbance residual segment data corresponding to each time window. The frequency scanning process is performed on the segmented data of the disturbance residual one by one, and the amplitude distribution of each frequency component is calculated within the preset frequency range to generate the frequency amplitude distribution data within the corresponding time window. In the frequency amplitude distribution data, the amplitudes of each frequency component are compared, and the frequency value corresponding to the frequency component with the largest amplitude is selected to generate the frequency location parameter. Centered on the frequency location parameter, search for frequency intervals in the frequency amplitude distribution data with continuously higher amplitude values ​​than the preset amplitude threshold in both the high-frequency and low-frequency directions to generate the bandwidth parameter; The generation of the bandwidth parameter specifically includes: Using the frequency component corresponding to the frequency location parameter as the center frequency, the corresponding center frequency position is determined in the frequency amplitude distribution data. Starting from the center frequency position, the frequency amplitude distribution data is traversed one frequency component along the high-frequency direction to determine whether the amplitude of each frequency component is continuously higher than the preset amplitude threshold, thus determining the highest frequency boundary that meets the conditions. Starting from the center frequency position, the frequency amplitude distribution data is traversed one frequency component along the low-frequency direction to determine whether the amplitude of each frequency component is continuously higher than the preset amplitude threshold, thus determining the lowest frequency boundary that meets the conditions. Based on the frequency interval between the highest and lowest frequency boundaries, the frequency range is calculated, and the bandwidth parameter is generated. The frequency location parameters generated within multiple adjacent time windows are compared, and the changes in the frequency location parameters over time are calculated to generate time stability parameters. The generation of the time stability parameters specifically includes: Frequency position parameters generated within multiple adjacent time windows are selected in chronological order to construct a frequency position parameter sequence for the corresponding time window sequence. The frequency position parameters corresponding to adjacent time windows in the frequency position parameter sequence are compared one by one to calculate the frequency position change between adjacent time windows. The frequency position changes corresponding to multiple adjacent time windows are aggregated to form a set of frequency position changes arranged in chronological order. Each frequency position change in the set of frequency position changes is compared one by one, and the maximum change in the set of frequency position changes is extracted to generate a time stability parameter characterizing the amplitude of frequency position parameter changes within multiple adjacent time windows. Interference fingerprints are generated by combining frequency location parameters, bandwidth parameters, and time stability parameters.

[0024] In this embodiment, the generation of the compensation residual data specifically includes: Multiple notch templates are read from a pre-established notch template library. Each of the multiple notch templates is pre-configured with a corresponding frequency position range and bandwidth range to generate a notch template set. The frequency position parameters and bandwidth parameters in the interference fingerprint are compared one by one with the frequency position range and bandwidth range of each notch template in the notch template set. The notch templates whose frequency position parameters fall into the corresponding frequency position range and whose bandwidth parameters fall into the corresponding bandwidth range are selected to generate the target notch template. Frequency domain mapping is performed on the perturbation residual data to convert it into a frequency representation. Based on the frequency location range and bandwidth range defined by the target notch template, the frequency components to be suppressed are identified in the frequency representation. The frequency components to be suppressed are subjected to amplitude attenuation processing, while the unidentified frequency components retain their original amplitudes, generating disturbance residual data after interference suppression. The disturbance residual data after interference suppression is processed by inverse frequency domain mapping to generate compensated residual data.

[0025] In this embodiment, the generation of the three-channel input data specifically includes: Within the same time window, the baseline components, perturbation residual data, and compensation residual data are time-aligned to generate time-aligned baseline components, perturbation residual data, and compensation residual data. Amplitude normalization is performed on the time-aligned baseline components, perturbation residual data, and compensation residual data along each spatial axis to generate normalized baseline components, perturbation residual data, and compensation residual data. According to the preset channel division order, the normalized baseline component is used as the first channel data, the normalized disturbance residual data is used as the second channel data, and the normalized compensation residual data is used as the third channel data. Under each spatial axis and time index, the corresponding first channel data, second channel data, and third channel data are combined to generate three-channel input data.

[0026] In this embodiment, the generation of the multi-scale perturbation feature representation specifically includes: In the improved SincNet, the three-channel input data are mapped to independent channel processing branches, and independent convolution processing paths are established in each channel processing branch. The improvements to SincNet specifically include: The input layer is improved to receive three-channel input data consisting of baseline components, perturbation residual data, and compensation residual data. In the network structure layer, independent channel processing branches are set up for each of the three channels of input data. Independent convolutional processing paths are established within each channel processing branch. Within each convolutional processing path, a hierarchical convolutional structure is constructed, consisting of parameterized bandpass convolution processing and multi-timescale convolutional processing. In the feature organization layer, the perturbation feature components generated by different channel processing branches at different timescales are concatenated to generate joint perturbation features across channels and timescales. Furthermore, scale uniformity processing is performed on the joint perturbation features to generate multi-scale perturbation feature representations. Parameterized bandpass convolution kernels are configured in the convolution processing path of each channel processing branch. The parameterized bandpass convolution kernels are used to perform bandpass convolution processing on the three-channel input data of the corresponding channel, and frequency response constraints are applied to the input data to generate channel-level initial perturbation features corresponding to each channel. Multiple convolution windows of different lengths are set within the convolution processing path of each channel processing branch. Convolution processing is performed on the initial perturbation features at the channel level at different time scales to generate perturbation feature components corresponding to the channel and time scale. Within the convolutional processing path of each channel processing branch, the perturbation feature components generated at different time scales are organized under the same time index condition to generate multi-scale perturbation feature components in each channel. The multi-scale perturbation feature components of different channel processing branches are spliced ​​in both channel and time scale dimensions to generate joint perturbation features across channels and time scales. The joint perturbation features are scale-consistentized to unify the feature dimensions corresponding to different time scales and generate multi-scale perturbation feature representations.

[0027] In this embodiment, the generation of the perturbation feature representation specifically includes: According to the different spatial axes corresponding to the magnetic field sensor, the multi-scale disturbance feature representation is divided to generate axial feature components corresponding to each spatial axis. Alignment is performed on the axial feature components corresponding to each spatial axis under the same time index condition to generate axial feature components that are consistent in the time dimension. Under each time index, the axial feature components corresponding to different spatial axes are combined to generate a joint axial feature that reflects the state of multiple spatial axes under the same time index. Perform cross-coupling processing on the joint axial features in the spatial axial dimension, calculate the mutual influence relationship between each spatial axial feature component, and generate the cross-coupled axial features. The axial features after cross-coupling are arranged in time index order in the time dimension to generate coupled features in the form of a continuous time series. The coupled features in the form of continuous time series are aggregated to generate perturbation feature representations.

[0028] In this embodiment, the generation of the magnetic field disturbance pattern recognition result specifically includes: The perturbation feature representation is input into the improved SincNet, and feature recombination processing is performed on the perturbation feature representation along the time dimension in the improved SincNet to generate temporal discriminative features; In the improved SincNet, feature mapping is performed on the temporal discriminative features at different discriminative scales to generate perturbation pattern discriminative features corresponding to different discriminative scales; Scale consistency constraint processing is performed on the perturbation pattern discrimination features under different discrimination scales to eliminate discrimination bias between different discrimination scales and generate scale-consistent perturbation pattern discrimination features; Based on scale-consistent perturbation pattern discrimination features, perturbation pattern determination processing is performed in the improved SincNet to generate magnetic field perturbation pattern recognition results.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a complex electromagnetic environment monitoring scenario that operates for a long time. This scenario is located in an area where various power and communication equipment coexist, surrounded by a continuous power frequency magnetic field background, stable narrowband electromagnetic crosstalk, and irregular anthropogenic magnetic field disturbances. Multi-axis magnetic field sensors are deployed within the monitoring area to continuously collect data on changes in the environmental magnetic field. Due to limited space, the magnetic field sensors inevitably exhibit orientation differences during installation, resulting in significant inconsistencies in the amplitude and phase characteristics of the magnetic field response along different spatial axes. Furthermore, the long-term presence of a slowly varying magnetic field background in the environment, combined with transient disturbance signals, makes it difficult for traditional methods to reliably identify weak magnetic field disturbances in actual operation, leading to frequent misjudgments and missed detections.

[0030] In this application scenario, the magnetic field sensor continuously collects time-series magnetic field data along multiple spatial axes and transmits and stores it according to preset sampling rules. Preprocessing operations such as time synchronization, anomaly removal, amplitude normalization, and resampling are performed on the collected raw magnetic field time-series data to form a structured multi-axis magnetic field time-series dataset. Within a continuous operating time window, the multi-axis magnetic field time-series dataset is decomposed to extract baseline components characterizing the slow-changing trend of the magnetic field. The difference between the raw multi-axis magnetic field time-series data and the corresponding baseline components is calculated to obtain disturbance residual data reflecting transient changes. Through this processing, long-term background magnetic field changes in the environment are effectively separated, providing a cleaner data foundation for subsequent disturbance analysis.

[0031] After obtaining the disturbance residual data, further frequency domain analysis is performed to extract frequency location parameters, bandwidth parameters, and time stability parameters that reflect the interference characteristics. These parameters are then combined to construct an interference fingerprint, which can stably characterize the recurring narrowband interference characteristics in the field environment. Based on the interference fingerprint, a corresponding notch template is automatically matched from a pre-established notch template library to suppress the frequency components in the disturbance residual data that correspond to the interference characteristics, generating compensated residual data. In this way, targeted suppression of stable interference is achieved without destroying effective disturbance information, avoiding excessive attenuation of useful signals by traditional fixed filtering methods.

[0032] In the data organization stage, baseline components, perturbation residual data, and compensation residual data are aligned and normalized under the same time window and spatial axis conditions. Three-channel input data is constructed according to preset rules. This three-channel input data simultaneously retains magnetic field background information, original perturbation information, and perturbation information after interference suppression, enabling the subsequent network to comprehensively analyze magnetic field changes from different perspectives. The three-channel input data is input into the improved SincNet network structure. Inside the network, independent channel processing branches are set for each of the three channel input data. In each channel processing branch, parameterized bandpass convolution processing and multi-timescale convolution processing are performed sequentially. This allows the network to further extract perturbation features at different time scales based on the frequency response constraints, thereby generating a multi-scale perturbation feature representation.

[0033] Based on this, a triaxial cross-coupling process is performed on the feature components corresponding to different spatial axes in the multi-scale perturbation feature representation, so that the perturbation features between different spatial axes can participate in the calculation and form a perturbation feature representation that reflects the mutual influence relationship of spatial axes. Through this process, the problem of inconsistent axial response caused by the difference in sensor installation posture is effectively reduced, and the generated perturbation features maintain good consistency and stability in the spatial dimension. Based on the perturbation feature representation, the perturbation pattern recognition processing is completed in the improved SincNet. Through multi-discriminative scale mapping and scale consistency constraints, the discriminative information at different time scales participates in the recognition decision in a unified semantic space, and finally outputs the magnetic field perturbation pattern recognition result.

[0034] In actual operation, the method of this invention is compared and analyzed with existing magnetic field disturbance identification methods based on single time scale features or simple multi-axis stitching. Through long-term observation of continuous monitoring data, it can be found that after adopting the method of this invention, the magnetic field disturbance pattern recognition results show higher stability and consistency under complex electromagnetic background conditions. For weak disturbance events, this invention can maintain good recognition continuity when background magnetic field fluctuations and stable narrowband interference coexist, effectively reducing misjudgments caused by interference superposition. At the same time, after the sensor attitude changes or the environmental conditions are adjusted, the disturbance features generated by this invention can still maintain good discrimination consistency, demonstrating good scene adaptability.

[0035] In summary, the above embodiments demonstrate that the intelligent identification method for magnetic field disturbance patterns based on convolutional networks proposed in this invention can achieve stable identification of magnetic field disturbances in complex electromagnetic environments. It significantly improves the problem of performance degradation in existing technologies under strong background interference and inconsistent multi-axis responses, and has strong engineering practicality and promotional value.

[0036] Table 1. Comparison of Magnetic Field Disturbance Pattern Recognition Performance under Complex Electromagnetic Environments

[0037] As can be seen from Table 1, traditional time-domain feature methods have limited overall recognition capabilities in complex electromagnetic environments. These methods mainly rely on changes in magnetic field amplitude, statistical features, or fixed thresholds for judgment. In the case of long-term presence of power frequency magnetic field background and slowly changing environmental magnetic field, weak disturbance signals are easily masked by background changes, resulting in a low recognition accuracy. At the same time, the false alarm rate and false negative rate are both high, the ability to detect weak disturbances is insufficient, and the cross-axis consistency performance is poor.

[0038] Traditional frequency domain analysis methods, by introducing spectral features, alleviate the problem of time domain methods being sensitive to stable interference to some extent, thus slightly improving the overall recognition accuracy. However, these methods usually employ fixed frequency bands or regularized analysis approaches, making it difficult to take into account the time-varying characteristics of disturbance signals. Especially when the disturbance frequency drifts over time or multiple interferences are superimposed, the phenomenon of missed detection is still quite obvious, and the improvement in recognition consistency between spatial axes is limited.

[0039] Compared with traditional methods, general convolutional network methods have stronger feature learning capabilities and can adapt to complex magnetic field changes to a certain extent. The overall recognition accuracy and weak perturbation detection rate are improved. However, since they usually directly model the original or simply preprocessed multi-axis magnetic field data in a unified manner, they lack a structured distinction between magnetic field background, perturbation residuals and interference suppression results. At the same time, they do not explicitly introduce a correlation mechanism between multiple time scales and spatial axes. Under conditions of complex interference and inconsistent poses, there are still false alarms and false negatives, and the improvement in cross-axis consistency is limited.

[0040] In contrast, the method of this invention exhibits more stable and balanced performance advantages without pursuing extreme recognition accuracy. By decomposing the multi-axis magnetic field time-series data and explicitly separating the baseline component and perturbation residual, the weak perturbation signal is enhanced at the data level, which is the main reason for the significant improvement in the weak perturbation detection rate. At the same time, based on the interference fingerprint and notch template library interference suppression mechanism, stable narrowband interference is specifically weakened, effectively reducing the false alarm rate. Furthermore, the improved SincNet, through the hierarchical structure of frequency response constraints and multi-time-scale convolution, enables perturbation features to be effectively characterized at different time scales, while the three-axis cross-coupling processing enhances the consistency between features of different spatial axes. Thus, without significantly increasing the system complexity, the overall recognition performance and stability are simultaneously improved.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent recognition of magnetic field perturbation patterns based on convolutional networks, characterized in that, Includes the following steps: Collect magnetic field time series data of magnetic field sensor in multiple spatial axes, and preprocess to generate multi-axis magnetic field time series dataset; Within a preset time window, the multi-axis magnetic field time series dataset is decomposed, the baseline component is extracted, and the difference between the multi-axis magnetic field time series dataset and the baseline component is calculated to generate perturbation residual data. Based on the perturbation residual data, frequency location parameters, bandwidth parameters, and time stability parameters are extracted to construct the interference fingerprint; Based on the interference fingerprint, a corresponding notch template is selected from the pre-established notch template library. The notch template is then used to suppress the interference in the disturbance residual data and generate compensated residual data. Three-channel input data is constructed based on baseline components, perturbation residual data, and compensation residual data; The three-channel input data is fed into the improved SincNet, and features are extracted from the three-channel input data at different time scales to generate multi-scale perturbation feature representations. A triaxial cross-coupling process is performed on the feature components corresponding to different spatial axes in the multi-scale perturbation feature representation to generate the perturbation feature representation; Based on the perturbation feature representation, magnetic field perturbation pattern recognition results are generated using an improved SincNet.

2. The intelligent recognition method for magnetic field perturbation patterns based on convolutional networks according to claim 1, characterized in that, The preprocessing specifically includes: time synchronization, anomaly removal, amplitude normalization, resampling, and time sequence arrangement.

3. The intelligent recognition method for magnetic field perturbation patterns based on convolutional networks according to claim 1, characterized in that, The generation of the perturbation residual data specifically includes: Within a preset time window, the multi-axis magnetic field time series dataset is extracted in chronological order to generate multi-axis magnetic field time series data within the corresponding time window; Based on the multi-axis magnetic field time series data within the corresponding time window, the magnetic field time series data of each spatial axis are decomposed to generate baseline components corresponding to the time window. Based on the baseline component, time alignment processing is performed on the multi-axis magnetic field time series data within the corresponding time window in each spatial axis to generate aligned data. Based on the aligned data, the multi-axis magnetic field time series dataset and the baseline components are subtracted along each spatial axis to generate perturbation residual data corresponding to each spatial axis. The disturbance residual data corresponding to each spatial axis are aggregated and processed to generate disturbance residual data.

4. The intelligent recognition method for magnetic field perturbation patterns based on convolutional networks according to claim 1, characterized in that, The generation of the interference fingerprint specifically includes: The disturbance residual data is segmented according to a preset time window to generate disturbance residual segment data corresponding to each time window. The frequency scanning process is performed on the segmented data of the disturbance residual one by one, and the amplitude distribution of each frequency component is calculated within the preset frequency range to generate the frequency amplitude distribution data within the corresponding time window. In the frequency amplitude distribution data, the amplitudes of each frequency component are compared, and the frequency value corresponding to the frequency component with the largest amplitude is selected to generate the frequency location parameter. Centered on the frequency location parameter, search for frequency intervals in the frequency amplitude distribution data with continuously higher amplitude values ​​than the preset amplitude threshold in both the high-frequency and low-frequency directions to generate the bandwidth parameter; The frequency location parameters generated within multiple adjacent time windows are compared, and the changes in the frequency location parameters over time are calculated to generate time stability parameters. Interference fingerprints are generated by combining frequency location parameters, bandwidth parameters, and time stability parameters.

5. The intelligent recognition method for magnetic field perturbation patterns based on convolutional networks according to claim 1, characterized in that, The generation of the compensation residual data specifically includes: Multiple notch templates are read from a pre-established notch template library. Each of the multiple notch templates is pre-configured with a corresponding frequency position range and bandwidth range to generate a notch template set. The frequency position parameters and bandwidth parameters in the interference fingerprint are compared one by one with the frequency position range and bandwidth range of each notch template in the notch template set. The notch templates whose frequency position parameters fall into the corresponding frequency position range and whose bandwidth parameters fall into the corresponding bandwidth range are selected to generate the target notch template. Frequency domain mapping is performed on the perturbation residual data to convert it into a frequency representation. Based on the frequency location range and bandwidth range defined by the target notch template, the frequency components to be suppressed are identified in the frequency representation. The frequency components to be suppressed are subjected to amplitude attenuation processing, while the unidentified frequency components retain their original amplitudes, generating disturbance residual data after interference suppression. The disturbance residual data after interference suppression is processed by inverse frequency domain mapping to generate compensated residual data.

6. The intelligent recognition method for magnetic field perturbation patterns based on convolutional networks according to claim 1, characterized in that, The generation of the three-channel input data specifically includes: Within the same time window, the baseline components, perturbation residual data, and compensation residual data are time-aligned to generate time-aligned baseline components, perturbation residual data, and compensation residual data. Amplitude normalization is performed on the time-aligned baseline components, perturbation residual data, and compensation residual data along each spatial axis to generate normalized baseline components, perturbation residual data, and compensation residual data. According to the preset channel division order, the normalized baseline component is used as the first channel data, the normalized disturbance residual data is used as the second channel data, and the normalized compensation residual data is used as the third channel data. Under each spatial axis and time index, the corresponding first channel data, second channel data, and third channel data are combined to generate three-channel input data.

7. The intelligent recognition method for magnetic field perturbation patterns based on convolutional networks according to claim 1, characterized in that, The generation of the multi-scale perturbation feature representation specifically includes: In the improved SincNet, the three-channel input data are mapped to independent channel processing branches, and independent convolution processing paths are established in each channel processing branch. The improvements to SincNet specifically include: The input layer is improved to receive three-channel input data consisting of baseline components, perturbation residual data, and compensation residual data. In the network structure layer, independent channel processing branches are set up for each of the three channels of input data. Independent convolutional processing paths are established within each channel processing branch. Within each convolutional processing path, a hierarchical convolutional structure is constructed, consisting of parameterized bandpass convolution processing and multi-timescale convolutional processing. In the feature organization layer, the perturbation feature components generated by different channel processing branches at different timescales are concatenated to generate joint perturbation features across channels and timescales. Furthermore, scale uniformity processing is performed on the joint perturbation features to generate multi-scale perturbation feature representations. Parameterized bandpass convolution kernels are configured in the convolution processing path of each channel processing branch. The parameterized bandpass convolution kernels are used to perform bandpass convolution processing on the three-channel input data of the corresponding channel, and frequency response constraints are applied to the input data to generate channel-level initial perturbation features corresponding to each channel. Multiple sets of convolution windows of different lengths are set in the convolution processing path of each channel processing branch. Convolution processing is performed on the initial perturbation features at the channel level at different time scales to generate perturbation feature components corresponding to the channel and time scale. Within the convolutional processing path of each channel processing branch, the perturbation feature components generated at different time scales are organized under the same time index condition to generate multi-scale perturbation feature components in each channel. The multi-scale perturbation feature components of different channel processing branches are spliced ​​in both channel and time scale dimensions to generate joint perturbation features across channels and time scales. The joint perturbation features are scale-consistentized to unify the feature dimensions corresponding to different time scales and generate multi-scale perturbation feature representations.

8. The intelligent recognition method for magnetic field perturbation patterns based on convolutional networks according to claim 1, characterized in that, The generation of the perturbation feature representation specifically includes: According to the different spatial axes corresponding to the magnetic field sensor, the multi-scale disturbance feature representation is divided to generate axial feature components corresponding to each spatial axis. Alignment is performed on the axial feature components corresponding to each spatial axis under the same time index condition to generate axial feature components that are consistent in the time dimension. Under each time index, the axial feature components corresponding to different spatial axes are combined to generate a joint axial feature that reflects the state of multiple spatial axes under the same time index. Perform cross-coupling processing on the joint axial features in the spatial axial dimension, calculate the mutual influence relationship between each spatial axial feature component, and generate the cross-coupled axial features. The axial features after cross-coupling are arranged in time index order in the time dimension to generate coupled features in the form of a continuous time series. The coupled features in the form of continuous time series are aggregated to generate perturbation feature representations.

9. The intelligent recognition method for magnetic field perturbation patterns based on convolutional networks according to claim 1, characterized in that, The generation of the magnetic field disturbance pattern recognition result specifically includes: The perturbation feature representation is input into the improved SincNet, and feature recombination processing is performed on the perturbation feature representation along the time dimension in the improved SincNet to generate temporal discriminative features; In the improved SincNet, feature mapping is performed on the temporal discriminative features at different discriminative scales to generate perturbation pattern discriminative features corresponding to different discriminative scales; Scale consistency constraint processing is performed on the perturbation pattern discrimination features under different discrimination scales to eliminate discrimination bias between different discrimination scales and generate scale-consistent perturbation pattern discrimination features; Based on scale-consistent perturbation pattern discrimination features, perturbation pattern determination processing is performed in the improved SincNet to generate magnetic field perturbation pattern recognition results.

Images