Adaptive spectrum monitoring and interference suppression method for railway power transformer

Abstract

The application discloses a self-adaptive spectrum monitoring and interference suppression method for railway power transformers, comprising the following steps: S1, collecting original multi-source signal data; S2, performing high-order filtering and Z-score normalization processing on the original multi-source signal; S3, inputting the original multi-source signal into a multi-scale residual fusion time-frequency transformation network to extract a time-frequency feature tensor; S4, inputting the time-frequency feature tensor into an interference identification network with a fusion attention mechanism; S5, dynamically activating an interference suppression module according to the identification result; S6, constructing a multi-dimensional tensor data structure, extracting sparse dictionary morphological features and spectral domain statistics, and generating a composite feature vector set; S7, inputting the composite feature vector into a health state evaluation module; and S8, uploading the diagnosis result to a remote monitoring platform through an embedded communication module. The application realizes intelligent identification and remote monitoring of railway transformer faults by combining multi-dimensional perception and adaptive modeling.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of power systems and condition monitoring of rail transit equipment, and particularly to an adaptive spectrum monitoring and interference suppression method for railway power transformers. Background Technology

[0002] In the current railway power system, power transformers, as core power transmission and distribution equipment, are crucial for ensuring the safety of train traction power supply through real-time monitoring and fault early warning. Traditional transformer condition monitoring relies heavily on offline detection, single-point sensing, and experience-based judgment, which is ill-suited to the complex characteristics of the railway operating environment, including high-intensity interference, mixed multi-source signals, and rapid fault evolution. Especially in high-speed railways, heavy-haul railways, and urban rail transit, electrical equipment is subjected to the combined effects of frequent start-stop cycles, high-frequency impacts, electromagnetic disturbances, and structural vibrations. This makes the early characteristics of typical faults such as partial discharge, insulation degradation, and winding loosening extremely weak and easily masked by background noise, severely affecting the sensitivity and accuracy of existing condition monitoring systems.

[0003] In existing technologies, spectrum analysis is often used as a key diagnostic tool for fault identification in electrical signals. However, it typically employs fixed-window-length Fourier transform, short-time analysis, or static wavelet transform methods, lacking adaptability to dynamic signal changes. These techniques exhibit insufficient resolution and weak identification of abrupt feature changes when facing non-stationary, multi-frequency interference signals in railway environments, failing to effectively extract crucial diagnostic information. Furthermore, in terms of interference suppression, traditional methods usually employ simple filters or set threshold rules to remove interference signals, but their ability to handle multiple types of superimposed interference, spectral overlap, or non-Gaussian backgrounds is limited, easily leading to loss of useful signals or misidentification.

[0004] To address the need for multi-source signal feature extraction in typical transformer faults such as partial discharge, harmonic anomalies, and mechanical faults, existing methods suffer from insufficient modeling of the coordinated changes among different physical quantities. Most methods lack cross-domain feature fusion and time-series graph modeling capabilities, making it difficult to accurately track fault trends, state evolution, and potential anomalies. Especially in large-scale, multi-transformer scenarios, current methods often lack a unified data structure, intelligent analysis, and remote collaborative scheduling mechanisms, failing to form an integrated system with closed-loop control and early warning response capabilities.

[0005] The adaptive spectrum monitoring and interference suppression method for railway power transformers proposed in this invention aims to address the problems of poor spectrum sensing capability, weak interference adaptability, low feature modeling capability, and lack of remote linkage mechanism in the existing technologies. By constructing a multi-channel high-precision acquisition mechanism and integrating a multi-scale residual network and an attention-enhanced recognition module, deep signal deconstruction and interference identification are achieved. Furthermore, an adaptive filtering mechanism is constructed by combining variational mode decomposition and gated cyclic unit networks, effectively improving the target signal extraction capability. Finally, fault trend modeling is achieved through graph neural networks, ultimately forming a closed-loop intelligent monitoring system capable of remote uploading and multi-station collaboration, significantly improving the state perception and anomaly response capabilities of railway power transformers in complex operating scenarios.

[0006] Therefore, how to provide adaptive spectrum monitoring and interference suppression methods for railway power transformers is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an adaptive spectrum monitoring and interference suppression method for railway power transformers. This invention integrates multi-channel sensing, intelligent spectrum extraction, and deep interference suppression technologies, combines graph neural networks to achieve state trend modeling, and constructs a closed-loop monitoring system through a remote communication module. This solves the problems of difficult fault identification, weak feature extraction, and delayed early warning response of railway transformers in complex interference environments, thereby improving diagnostic accuracy and system intelligence.

[0008] An adaptive spectrum monitoring and interference suppression method for railway power transformers according to an embodiment of the present invention includes the following steps:

[0009] S1. Set up multi-channel signal acquisition units on the primary and secondary sides of the railway power transformer to simultaneously acquire raw multi-source signal data including voltage signals, current signals, partial discharge signals, electromechanical vibration signals and electromagnetic radiation signals.

[0010] S2. Preprocess the original multi-source signal, the preprocessing including: removing DC bias using a high-order statistical filter and normalizing the amplitude range by Z-score normalization.

[0011] S3. Input the preprocessed signal into a multi-scale residual fusion time-frequency transform network. The network includes a parallel multi-scale convolution sub-module, a residual connection module, and a variable window-length short-time Fourier decomposer, which is used to extract the time-frequency feature tensors of the signal under different time windows and frequency ranges.

[0012] S4. Input the time-frequency feature tensor into the interference recognition network A-DINet, which integrates the attention mechanism. This network is built based on the multi-head attention mechanism, multi-channel temporal convolutional network and category correlation analyzer. It identifies interference including but not limited to the following: 50 Hz and its integer multiples of power frequency harmonics, transient pulse interference caused by high voltage switches, non-stationary low-frequency disturbances and broadband non-Gaussian noise.

[0013] S5. Dynamically activate the interference suppression module based on the recognition results. This module consists of a signal decomposer based on variational mode decomposition (VMD) and a gated recurrent unit (GRU) sequence learning network. It adopts a two-stage strategy: the first stage uses VMD to perform mode decomposition on the original signal and selects highly correlated components based on the offset energy criterion; the second stage uses the GRU model to perform amplitude-frequency response prediction and target reconstruction based on frequency domain sequence modeling, thereby achieving target signal enhancement and interference suppression.

[0014] S6. Construct the signal after interference suppression into a multidimensional tensor data structure, extract pulse signal morphological features based on sparse dictionary learning method, and combine statistical and spectral domain features such as spectral entropy, kurtosis factor, instantaneous energy fluctuation rate, frequency drift amplitude, and spectral distribution skewness to form a composite feature vector set for diagnostic modeling.

[0015] S7. Input the composite feature vector into the health status assessment module. The module consists of a graph attention network (GAT) and a multi-channel graph convolutional neural network (GCN). By constructing a dynamic evolution graph structure with nodes representing monitoring time points and edges representing feature similarity, it completes the anomaly identification, classification, and operation trend prediction of the transformer status.

[0016] S8. The diagnostic results, including spectrum, interference category, fault label, and status trend score, are uploaded to the remote monitoring platform through the embedded communication module. It also supports linkage with the dispatch system based on the MQTT protocol or IEC 61850 standard to complete the closed-loop operation of intelligent early warning, trend visualization and multi-transformer point collaborative control.

[0017] Optionally, S1 specifically includes:

[0018] S11. Multi-channel signal acquisition units are respectively installed on the primary and secondary sides of the railway power transformer. The multi-channel signal acquisition unit includes a current and voltage sensor, a partial discharge signal coupling device, a vibration sensor, and an electromagnetic radiation receiving device.

[0019] S12. High-frequency sampling of voltage and current signals is performed by a high-speed synchronous sampling module to obtain complete electrical transient information;

[0020] S13. Use a high-frequency capacitive coupler or a special current probe to perform non-contact, high-sensitivity acquisition of partial discharge signals and capture discharge transient pulse signals.

[0021] S14. Use piezoelectric or fiber optic vibration sensors to monitor the micro-vibration behavior of the power transformer body structure and its installation foundation, and generate mechanical response signals.

[0022] S15. Use a wideband radio frequency receiving antenna or near-field probe to collect electromagnetic radiation signals around the transformer and capture the characteristics of potential radiation interference sources in the system.

[0023] S16. Perform unified clock synchronization control on the above-mentioned multi-channel signal acquisition units, and start synchronous acquisition under specified trigger conditions to form original multi-source signal data covering multiple physical field information.

[0024] Optionally, S2 specifically includes:

[0025] S21. Perform channel separation processing on the acquired raw multi-source signal data, and construct independent time-series channel buffers according to the signal type;

[0026] S22. Use a high-order statistical filtering method to perform bias correction on each original signal to remove DC drift and baseline drift components;

[0027] S23. Perform amplitude normalization processing on the bias-corrected signal, and use the Z-score normalization method to uniformly transform the amplitude distribution of various signals to ensure that the data scale is consistent between different physical quantities.

[0028] S24. The adjacent sampling block overlap strategy based on time window realigns the time axis of various signal channels to ensure that the data of multiple channels have strict time synchronization under the same analysis window;

[0029] S25. The preprocessed multi-channel signal is constructed into an input segment signal sequence with a fixed length and overlap rate for use by the subsequent spectrum analysis module.

[0030] Optionally, S3 specifically includes:

[0031] S31. The preprocessed signal data are grouped according to the channel type, and the input sequence tensor is constructed by segmenting the data using a fixed-length sliding window. The length and step size of each window are set according to the target frequency distribution and computational complexity requirements. All windows maintain full coverage of the time axis during the sliding process to avoid sample loss.

[0032] S32. The input sequence tensor is fed into a multi-scale residual fusion time-frequency transform network. The overall structure of the multi-scale residual fusion time-frequency transform network includes a parallel multi-scale convolution path module, a deep residual connection module, an inter-layer channel enhancement module, and an adaptive window length control short-time Fourier transform module.

[0033] S33. In the parallel multi-scale convolutional pathway module, set up three or more independent convolutional sub-paths, and use convolutional kernels of different sizes and dilation rates to extract local features of the signal in the time windows of microscale (e.g., 0.1ms level), mesoscale (e.g., 1ms level), and macroscale (e.g., 10ms level) to form a cross-scale feature map group.

[0034] S34. The cross-scale feature map group is spliced ​​in the channel dimension and entered into the deep residual connection module. The deep residual connection module consists of a standard residual unit, a multi-layer skip connection path, a layer-by-layer normalization structure and a residual fusion gate unit, which is used to improve the stability of deep feature representation and alleviate the gradient vanishing problem.

[0035] S35. Based on the residual fusion output, an inter-layer channel enhancement module is introduced. The inter-layer channel enhancement module includes a channel attention mapping submodule, a spatial compression mapper, and a scale adaptive activation function, which are used to improve the response value of highly correlated channels and suppress interference from irrelevant channels.

[0036] S36. Input the enhanced feature tensor of the channel into the window-length adjustable short-time Fourier transform module. The window-length adjustable short-time Fourier transform module calculates the optimal length and overlap rate of the sliding window based on the frame energy fluctuation rate and the instantaneous entropy function, and performs fast Fourier transform in each window to construct a multi-time period spectrum.

[0037] S37. Perform frequency domain interpolation alignment, amplitude-frequency normalization, and spectrum expansion on all multi-time period spectrum results to construct a three-dimensional time-frequency feature tensor with time × frequency × channel dimensions. Then, send the three-dimensional time-frequency feature tensor into the dimension standardization module to unify the format to meet the input requirements of downstream models.

[0038] S38. Output the standardized three-dimensional time-frequency feature tensor. This three-dimensional time-frequency feature tensor retains the temporal dynamic features, frequency structure features and intra-scale dependencies of the signal, which is used to support the subsequent category discrimination and disturbance modeling process of the interference identification network.

[0039] Optionally, S4 specifically includes:

[0040] S41. Input the three-dimensional time-frequency feature tensor into the interference recognition network A-DINet which integrates the attention mechanism. The three-dimensional time-frequency feature tensor includes three structural dimensions: time dimension, frequency dimension and channel dimension. The data is then sent into the interference recognition process in a preset batch.

[0041] S42. In the interference identification network, feature extraction is performed through a multi-channel temporal convolution module. The multi-channel temporal convolution module consists of multiple independent temporal convolution units. Each temporal convolution unit performs one-dimensional or two-dimensional convolution operations on different frequency bands to capture the short-term dependence features of power frequency disturbances, harmonic interference, shock wave bands and broadband noise.

[0042] S43. The feature maps after convolution of each frequency band are concatenated in the channel dimension to construct an intermediate fusion tensor, which is then fed into the multi-head attention encoding module. The multi-head attention encoding module contains multiple parallel attention heads, and each attention head performs weighted modeling of feature correlations in different scales and different frequency bands.

[0043] S44. Calculate the degree of information interdependence between frequency bands using the attention distribution matrix, and then dynamically adjust the channel weight coefficients to achieve enhancement processing of spectral bands with significant perturbations, while suppressing the response values ​​of irrelevant frequency domain regions.

[0044] S45. Input the enhanced feature tensor into the category correlation analyzer module. The category correlation analyzer module includes a perturbation type discrimination unit, a channel feature fusion layer and a class-specific scorer, which are respectively used to extract category-level prior features, perform cross-channel aggregation of frequency band dimensions and output activation score vectors for specific interference types.

[0045] S46. Based on the category scoring results, perform perturbation discrimination and label assignment on the input tensor. The perturbation types include 50 Hz power frequency fundamental wave interference, 50 Hz integer multiple harmonic interference, transient impact interference caused by high voltage switch operation, non-stationary low frequency perturbation, and broadband non-Gaussian noise interference with statistical adaptive characteristics.

[0046] S47. Output the finally identified interference tags and frequency domain location index together as the basis for dynamic scheduling of subsequent interference suppression modules, and for use by the abnormal event tracking mechanism.

[0047] Optionally, S5 specifically includes:

[0048] S51. Based on the interference type label and its corresponding frequency band index obtained in step S4, dynamically call the parameter configuration scheme of the interference suppression module and activate the signal processing path that matches the current interference characteristics.

[0049] S52. Input the original signal data into the signal decomposer module based on variational mode decomposition. The signal decomposer module performs multi-order mode decomposition on the signal to generate multiple intrinsic mode function components. Each intrinsic mode function component represents the signal composition at different frequency scales.

[0050] S53. Calculate the energy shift index, spectral density distribution characteristics, and similarity score with the target feature template for all intrinsic mode function components in sequence, and select several mode components to form a target mode set according to the preset correlation criterion.

[0051] S54. The target mode set is constructed as a time series input and fed into a sequence modeling network composed of a gate control recurrent unit. The sequence modeling network learns the amplitude variation law and frequency domain transfer path between modes through a recursive structure and outputs the predicted frequency response sequence.

[0052] S55. Based on the frequency domain prediction results output by the sequence modeling network, perform difference inversion with the original modal signal, and perform frequency filtering and reconstruction synthesis operations to suppress non-target interference components and enhance the target signal.

[0053] S56. Perform residual evaluation and dynamic update operations on the enhanced signal. If the difference between the current result and the disturbance template is still greater than the threshold, the decomposition path fine-tuning process is re-triggered until the interference suppression result is stable.

[0054] S57. Output the signal sequence after interference suppression processing, retain the physically meaningful anomalous features within the target frequency band, and significantly reduce the impact of multi-source interference components on the subsequent analysis model.

[0055] Optionally, S6 specifically includes:

[0056] S61. The signal after interference suppression processing in step S5 is organized according to channel type and sampling time window, and a standardized multidimensional tensor data structure is constructed based on the time axis, frequency axis and channel axis to ensure that its structure can be directly parsed by the subsequent feature extraction model.

[0057] S62. Perform morphological feature extraction operation based on sparse dictionary learning on multidimensional tensor data. The sparse dictionary is pre-learned from the training sample library and can express typical waveform patterns of pulse-like signals such as partial discharge in a finite set of atoms.

[0058] S63. The input signal is decomposed into a sparse coefficient matrix and an atomic dictionary using a sparse coding algorithm, and the starting point, duration, waveform symmetry and consistency characteristics of the pulse pattern in the encoding process are identified.

[0059] S64. Based on the sparse decomposition results, further extract frequency domain and statistical feature indicators from the signal. The feature indicators include spectral entropy, which describes the complexity of the frequency component distribution; kurtosis factor, which measures the degree of signal spikes; instantaneous energy fluctuation rate, which reflects the range of short-term energy changes; frequency drift amplitude, which characterizes the dynamic shift of the main frequency position; and spectral distribution skewness, which determines the skew direction and amplitude of frequency energy component deviation.

[0060] S65. All extracted morphological features, frequency domain indices and statistics are concatenated along the feature dimension to form a composite feature vector in a unified format, and a set of corresponding feature expressions is generated for each time window.

[0061] S66. Perform dimensionality standardization, outlier removal, and integrity checks on the composite feature vector to ensure that it meets the input requirements and accuracy requirements of the subsequent state evaluation module.

[0062] S67. Output the processed composite feature vector set, and retain its corresponding timestamp, channel label and source information, as the input data basis for railway power transformer state modeling and anomaly identification.

[0063] Optionally, S7 specifically includes:

[0064] S71. Arrange the composite feature vector set in chronological order and construct a discretized state monitoring sequence according to a fixed time granularity, using the feature vector of each time period as a single node input unit.

[0065] S72. Construct a dynamic evolution graph structure using monitoring time points as graph nodes and Euclidean distance, cosine similarity, or spectral distance between feature vectors as graph edges to form a time-series feature graph based on feature correlation.

[0066] S73. Input the dynamic evolution graph structure into the health status assessment module. The health status assessment module is composed of a graph attention network and a multi-channel graph convolutional neural network. The graph attention network dynamically weights the edge weights in the graph to improve the expressive ability of key state transition paths.

[0067] S74. In graph attention networks, a multi-head graph attention mechanism is introduced to differentiate the correlation between different time points, and the node representation is updated through a weighted adjacency matrix.

[0068] S75. The dynamic evolution graph structure updated by graph attention is input into a multi-channel graph convolutional neural network. The multi-channel graph convolutional neural network uses a multi-dimensional channel grouping strategy to perform deep convolution operations on node features and extracts the state change trend and feature evolution pattern across time periods.

[0069] S76. Input the global graph embedding vector output by graph convolution into the fully connected classification layer to perform health level classification and abnormal state identification of the current state of the transformer. The identified state labels include normal, mild abnormality, moderate abnormality and severe fault.

[0070] S77. Based on the node classification results of multiple consecutive time periods, construct the running state trend curve and calculate the running trend slope, volatility and potential inflection point to predict the state evolution direction in the next few cycles.

[0071] S78. Output the health status classification label, operation trend score index and status transition path at the current time point, so that the remote monitoring platform can perform early warning response, operation and maintenance scheduling and equipment maintenance strategy optimization.

[0072] Optionally, S8 specifically includes:

[0073] S81. Integrate the health status classification label, running trend scoring index and time series node graph embedding vector output in step S7, and construct a structured diagnostic result data package together with the spectrum, interference identification label and interference suppression residual value obtained in steps S3 to S6.

[0074] S82. Perform format standardization conversion on the structured diagnostic result data packet to generate a unified remote reporting format containing data timestamp, transformer number, channel source, anomaly type, feature summary, graph compression summary and status trend encoding fields;

[0075] S83. The formatted diagnostic result data is encapsulated through an embedded communication module, which includes multiple data transmission paths such as a cellular communication module, an Ethernet interface, and an industrial fiber optic transmission module.

[0076] S84. Select the communication protocol type according to the on-site deployment requirements, and use the lightweight message queue transmission protocol or the power protocol based on the IEC 61850 communication model for data packaging and transmission to ensure that the system can be compatible with different plant dispatching systems.

[0077] S85. Upload the diagnostic data to the remote monitoring platform, which supports structured data parsing, spectrum visualization, state trend curve reconstruction and interference distribution heat map display functions, forming a comprehensive diagnostic interface for multi-dimensional monitoring objects.

[0078] S86. Set alarm policies in the remote monitoring platform and automatically trigger multi-level early warning mechanisms based on the severity of status labels, trend score slope and disturbance level, including in-station audible and visual alarms, SMS push, control center pop-up windows and cloud platform alarm aggregation and distribution.

[0079] S87. Synchronously push the diagnostic results to the power dispatching system or equipment operation and maintenance management system, support the linkage execution of status refresh, task scheduling, maintenance work order issuance and inter-substation collaborative strategy switching, and build an intelligent operation and maintenance control network that supports closed-loop response at multiple sites.

[0080] S88. All uploaded records are cached locally and periodically backed up to the historical data management module for fault reproduction, model training optimization, and long-term performance evaluation.

[0081] The beneficial effects of this invention are:

[0082] This invention achieves several technological breakthroughs based on existing technologies by constructing an adaptive spectrum monitoring and interference suppression method for railway power transformers, significantly improving the adaptability, accuracy, and intelligence of transformer condition monitoring systems in complex electromagnetic interference environments.

[0083] First, this invention achieves simultaneous acquisition of multiple physical quantities and multiple channels in the signal acquisition stage, covering five typical parameters: voltage, current, partial discharge, vibration, and electromagnetic radiation. Through high-frequency, high-precision acquisition and a unified clock control mechanism, it effectively improves the timeliness and completeness of the original signal acquisition, providing a high-quality data foundation for subsequent spectrum modeling. Compared with traditional single-electrical-quantity monitoring methods, this method constructs rich information dimensions, ensuring that weak anomalies are no longer masked by background interference.

[0084] Secondly, in terms of spectral feature extraction, a multi-scale residual fusion time-frequency transform network was adopted. Through the synergistic effect of multi-scale convolution, residual connections, and adaptive window-length short-time Fourier decomposition, a fine characterization of the signal at different frequencies and time scales was achieved. This structure effectively compensates for the sluggish response of traditional fixed-window analysis to non-stationary signals and significantly improves the ability to identify fine-grained features such as partial discharge and harmonic disturbances. The interference recognition network combined with an attention mechanism further enhances the system's intelligent discrimination capability against interference such as power frequency harmonics, transient pulses, and broadband non-Gaussian noise, achieving automatic signal classification.

[0085] Furthermore, in terms of interference suppression, this invention introduces a two-stage processing strategy combining variational mode decomposition and gated recurrent neural networks, effectively completing the decomposition, identification, and predictive removal of interference components, ensuring that the target signal retains its integrity while removing invalid information to the maximum extent. Through amplitude-frequency response modeling and dynamic reconstruction, this method balances suppression effectiveness and signal fidelity, overcoming the problem of erroneous deletion of feature signals by traditional filtering methods. In the feature modeling and state assessment stages, this invention extracts pulse morphology features based on sparse dictionary learning, constructs composite feature vectors by combining various statistical and spectral parameters, and achieves dynamic identification and trend prediction of transformer operating states through a joint modeling approach using graph attention networks and graph convolutional neural networks. This modeling method possesses state representation capabilities across time scales, enhancing the system's early perception capability of abnormal evolution.

[0086] Finally, this invention also establishes a data upload mechanism that interfaces with a remote monitoring platform, supporting data transmission based on the MQTT protocol and the IEC 61850 standard. This enables real-time uploading of diagnostic results, alarm push notifications, and intelligent linkage control of multiple substations, comprehensively constructing a closed-loop intelligent operation and maintenance system from front-end perception to cloud-based control. In summary, this method possesses significant practicality, engineering promotion value, and application prospects in intelligent railway power supply systems. Attached Figure Description

[0087] 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:

[0088] Figure 1 The flowchart shows the adaptive spectrum monitoring and interference suppression method for railway power transformers proposed in this invention.

[0089] Figure 2 This is a schematic diagram of the network architecture of the multi-scale residual fusion time-frequency transform network proposed in this invention;

[0090] Figure 3 This is a model structure diagram of A-DINet, the interference recognition network that incorporates the attention mechanism proposed in this invention. Detailed Implementation

[0091] 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.

[0092] refer to Figure 1-3 An adaptive spectrum monitoring and interference suppression method for railway power transformers includes the following steps:

[0093] S1. Set up multi-channel signal acquisition units on the primary and secondary sides of the railway power transformer to simultaneously acquire raw multi-source signal data including voltage signals, current signals, partial discharge signals, electromechanical vibration signals and electromagnetic radiation signals.

[0094] S2. Preprocess the original multi-source signal, the preprocessing including: removing DC bias using a high-order statistical filter and normalizing the amplitude range by Z-score normalization.

[0095] S3. Input the preprocessed signal into a multi-scale residual fusion time-frequency transform network. The network includes a parallel multi-scale convolution sub-module, a residual connection module, and a variable window-length short-time Fourier decomposer, which is used to extract the time-frequency feature tensors of the signal under different time windows and frequency ranges.

[0096] S4. Input the time-frequency feature tensor into the interference recognition network A-DINet, which integrates the attention mechanism. This network is built based on the multi-head attention mechanism, multi-channel temporal convolutional network and category correlation analyzer. It identifies interference including but not limited to the following: 50 Hz and its integer multiples of power frequency harmonics, transient pulse interference caused by high voltage switches, non-stationary low-frequency disturbances and broadband non-Gaussian noise.

[0097] S5. Dynamically activate the interference suppression module based on the recognition results. This module consists of a signal decomposer based on variational mode decomposition (VMD) and a gated recurrent unit (GRU) sequence learning network. It adopts a two-stage strategy: the first stage uses VMD to perform mode decomposition on the original signal and selects highly correlated components based on the offset energy criterion; the second stage uses the GRU model to perform amplitude-frequency response prediction and target reconstruction based on frequency domain sequence modeling, thereby achieving target signal enhancement and interference suppression.

[0098] S6. Construct the signal after interference suppression into a multidimensional tensor data structure, extract pulse signal morphological features based on sparse dictionary learning method, and combine statistical and spectral domain features such as spectral entropy, kurtosis factor, instantaneous energy fluctuation rate, frequency drift amplitude, and spectral distribution skewness to form a composite feature vector set for diagnostic modeling.

[0099] S7. Input the composite feature vector into the health status assessment module. The module consists of a graph attention network (GAT) and a multi-channel graph convolutional neural network (GCN). By constructing a dynamic evolution graph structure with nodes representing monitoring time points and edges representing feature similarity, it completes the anomaly identification, classification, and operation trend prediction of the transformer status.

[0100] S8. The diagnostic results, including spectrum, interference category, fault label, and status trend score, are uploaded to the remote monitoring platform through the embedded communication module. It also supports linkage with the dispatch system based on the MQTT protocol or IEC 61850 standard to complete the closed-loop operation of intelligent early warning, trend visualization and multi-transformer point collaborative control.

[0101] This invention constructs a full-process intelligent monitoring framework, encompassing signal acquisition, preprocessing, time-frequency feature extraction, interference identification, interference suppression, feature modeling, and remote linkage. In the operating environment of railway transformers, it achieves joint processing of multi-source heterogeneous signals and deep learning-driven dynamic spectrum modeling. This effectively overcomes the problems of weak spectrum perception, high model rigidity, and broken response chains in traditional monitoring systems under non-stationary interference scenarios, and enables accurate extraction and closed-loop diagnosis of weak faults under complex operating conditions.

[0102] In this embodiment, S1 specifically includes:

[0103] S11. Multi-channel signal acquisition units are respectively installed on the primary and secondary sides of the railway power transformer. The multi-channel signal acquisition unit includes a current and voltage sensor, a partial discharge signal coupling device, a vibration sensor, and an electromagnetic radiation receiving device.

[0104] S12. High-frequency sampling of voltage and current signals is performed by a high-speed synchronous sampling module to obtain complete electrical transient information;

[0105] S13. Use a high-frequency capacitive coupler or a special current probe to perform non-contact, high-sensitivity acquisition of partial discharge signals and capture discharge transient pulse signals.

[0106] S14. Use piezoelectric or fiber optic vibration sensors to monitor the micro-vibration behavior of the power transformer body structure and its installation foundation, and generate mechanical response signals.

[0107] S15. Use a wideband radio frequency receiving antenna or near-field probe to collect electromagnetic radiation signals around the transformer and capture the characteristics of potential radiation interference sources in the system.

[0108] S16. Perform unified clock synchronization control on the above-mentioned multi-channel signal acquisition units, and start synchronous acquisition under specified trigger conditions to form original multi-source signal data covering multiple physical field information.

[0109] By deploying multi-channel acquisition systems on the primary and secondary sides of railway power transformers and integrating the sensing dimensions of electrical, mechanical, and electromagnetic physical fields, a five-dimensional signal acquisition network covering voltage, current, partial discharge, structural vibration, and radiated interference was formed. Combined with a unified clock synchronization mechanism, not only was millisecond-level signal alignment achieved, but the integrity of the raw data in terms of temporal continuity, spatial coverage, and fault scenario characterization was also greatly improved, providing a reliable engineering foundation for subsequent time-frequency analysis.

[0110] In this embodiment, S2 specifically includes:

[0111] S21. Perform channel separation processing on the acquired raw multi-source signal data, and construct independent time-series channel buffers according to the signal type;

[0112] S22. Use a high-order statistical filtering method to perform bias correction on each original signal to remove DC drift and baseline drift components;

[0113] S23. Perform amplitude normalization processing on the bias-corrected signal, and use the Z-score normalization method to uniformly transform the amplitude distribution of various signals to ensure that the data scale is consistent between different physical quantities.

[0114] S24. The adjacent sampling block overlap strategy based on time window realigns the time axis of various signal channels to ensure that the data of multiple channels have strict time synchronization under the same analysis window;

[0115] S25. The preprocessed multi-channel signal is constructed into an input segment signal sequence with a fixed length and overlap rate for use by the subsequent spectrum analysis module.

[0116] By introducing high-order statistical filtering and normalization strategies, this method can eliminate baseline offset, amplitude inconsistency and time misalignment caused by field equipment aging, inductive coupling differences and sampling drift. It ensures that multi-channel and multi-type signals have high consistency in frequency, amplitude and timing, effectively improves the adaptability and stability of input data to neural network models, and significantly enhances the robustness and feature fidelity of the spectrum extraction process.

[0117] In this embodiment, S3 specifically includes:

[0118] S31. The preprocessed signal data are grouped according to the channel type, and the input sequence tensor is constructed by segmenting the data using a fixed-length sliding window. The length and step size of each window are set according to the target frequency distribution and computational complexity requirements. All windows maintain full coverage of the time axis during the sliding process to avoid sample loss.

[0119] S32. The input sequence tensor is fed into a multi-scale residual fusion time-frequency transform network. The overall structure of the multi-scale residual fusion time-frequency transform network includes a parallel multi-scale convolution path module, a deep residual connection module, an inter-layer channel enhancement module, and an adaptive window length control short-time Fourier transform module.

[0120] S33. In the parallel multi-scale convolutional pathway module, set up three or more independent convolutional sub-paths, and use convolutional kernels of different sizes and dilation rates to extract local features of the signal in the time windows of microscale (e.g., 0.1ms level), mesoscale (e.g., 1ms level), and macroscale (e.g., 10ms level) to form a cross-scale feature map group.

[0121] S34. The cross-scale feature map group is spliced ​​in the channel dimension and entered into the deep residual connection module. The deep residual connection module consists of a standard residual unit, a multi-layer skip connection path, a layer-by-layer normalization structure and a residual fusion gate unit, which is used to improve the stability of deep feature representation and alleviate the gradient vanishing problem.

[0122] S35. Based on the residual fusion output, an inter-layer channel enhancement module is introduced. The inter-layer channel enhancement module includes a channel attention mapping submodule, a spatial compression mapper, and a scale adaptive activation function, which are used to improve the response value of highly correlated channels and suppress interference from irrelevant channels.

[0123] S36. Input the enhanced feature tensor of the channel into the window-length adjustable short-time Fourier transform module. The window-length adjustable short-time Fourier transform module calculates the optimal length and overlap rate of the sliding window based on the frame energy fluctuation rate and the instantaneous entropy function, and performs fast Fourier transform in each window to construct a multi-time period spectrum.

[0124] S37. Perform frequency domain interpolation alignment, amplitude-frequency normalization, and spectrum expansion on all multi-time period spectrum results to construct a three-dimensional time-frequency feature tensor with time × frequency × channel dimensions. Then, send the three-dimensional time-frequency feature tensor into the dimension standardization module to unify the format to meet the input requirements of downstream models.

[0125] S38. Output the standardized three-dimensional time-frequency feature tensor. This three-dimensional time-frequency feature tensor retains the temporal dynamic features, frequency structure features and intra-scale dependencies of the signal, which is used to support the subsequent category discrimination and disturbance modeling process of the interference identification network.

[0126] A multi-scale residual fusion time-frequency transform network is constructed. Local and global features are extracted through parallel convolution at small, medium, and large scales. Residual connections are combined to avoid gradient vanishing and feature destruction. Furthermore, a channel attention mechanism is used to finely control the information flow path. The frequency analysis granularity is dynamically adjusted through an adaptive window length STFT. This enables accurate representation of signals at multiple time scales and frequency levels, effectively solving the technical challenge of feature extraction imbalance in scenarios where partial discharge, high-frequency impact, and slow-varying interference coexist.

[0127] In this embodiment, S4 specifically includes:

[0128] S41. Input the three-dimensional time-frequency feature tensor into the interference recognition network A-DINet which integrates the attention mechanism. The three-dimensional time-frequency feature tensor includes three structural dimensions: time dimension, frequency dimension and channel dimension. The data is then sent into the interference recognition process in a preset batch.

[0129] S42. In the interference identification network, feature extraction is performed through a multi-channel temporal convolution module. The multi-channel temporal convolution module consists of multiple independent temporal convolution units. Each temporal convolution unit performs one-dimensional or two-dimensional convolution operations on different frequency bands to capture the short-term dependence features of power frequency disturbances, harmonic interference, shock wave bands and broadband noise.

[0130] S43. The feature maps after convolution of each frequency band are concatenated in the channel dimension to construct an intermediate fusion tensor, which is then fed into the multi-head attention encoding module. The multi-head attention encoding module contains multiple parallel attention heads, and each attention head performs weighted modeling of feature correlations in different scales and different frequency bands.

[0131] S44. Calculate the degree of information interdependence between frequency bands using the attention distribution matrix, and then dynamically adjust the channel weight coefficients to achieve enhancement processing of spectral bands with significant perturbations, while suppressing the response values ​​of irrelevant frequency domain regions.

[0132] S45. Input the enhanced feature tensor into the category correlation analyzer module. The category correlation analyzer module includes a perturbation type discrimination unit, a channel feature fusion layer and a class-specific scorer, which are respectively used to extract category-level prior features, perform cross-channel aggregation of frequency band dimensions and output activation score vectors for specific interference types.

[0133] S46. Based on the category scoring results, perform perturbation discrimination and label assignment on the input tensor. The perturbation types include 50 Hz power frequency fundamental wave interference, 50 Hz integer multiple harmonic interference, transient impact interference caused by high voltage switch operation, non-stationary low frequency perturbation, and broadband non-Gaussian noise interference with statistical adaptive characteristics.

[0134] S47. Output the finally identified interference tags and frequency domain location index together as the basis for dynamic scheduling of subsequent interference suppression modules, and for use by the abnormal event tracking mechanism.

[0135] By constructing the A-DINet interference identification network, which integrates multi-channel convolution, multi-head attention mechanism and category correlation modeling, automatic classification and spectral domain localization of typical disturbance types in railway transformer environment are realized. It can quickly separate and label power frequency harmonics, transient pulses, low frequency disturbances and broadband non-Gaussian interference under strong background noise conditions, significantly improving the automation level and spectral domain matching accuracy of interference identification, and optimizing the adaptation performance and resource allocation efficiency of subsequent interference suppression links.

[0136] In this embodiment, S5 specifically includes:

[0137] S51. Based on the interference type label and its corresponding frequency band index obtained in step S4, dynamically call the parameter configuration scheme of the interference suppression module and activate the signal processing path that matches the current interference characteristics.

[0138] S52. Input the original signal data into the signal decomposer module based on variational mode decomposition. The signal decomposer module performs multi-order mode decomposition on the signal to generate multiple intrinsic mode function components. Each intrinsic mode function component represents the signal composition at different frequency scales.

[0139] S53. Calculate the energy shift index, spectral density distribution characteristics, and similarity score with the target feature template for all intrinsic mode function components in sequence, and select several mode components to form a target mode set according to the preset correlation criterion.

[0140] S54. The target mode set is constructed as a time series input and fed into a sequence modeling network composed of a gate control recurrent unit. The sequence modeling network learns the amplitude variation law and frequency domain transfer path between modes through a recursive structure and outputs the predicted frequency response sequence.

[0141] S55. Based on the frequency domain prediction results output by the sequence modeling network, perform difference inversion with the original modal signal, and perform frequency filtering and reconstruction synthesis operations to suppress non-target interference components and enhance the target signal.

[0142] S56. Perform residual evaluation and dynamic update operations on the enhanced signal. If the difference between the current result and the disturbance template is still greater than the threshold, the decomposition path fine-tuning process is re-triggered until the interference suppression result is stable.

[0143] S57. Output the signal sequence after interference suppression processing, retain the physically meaningful anomalous features within the target frequency band, and significantly reduce the impact of multi-source interference components on the subsequent analysis model.

[0144] By constructing a two-stage interference suppression module combining variational mode decomposition and gated recurrent unit (GRU) networks, this approach utilizes VMD to perform frequency domain decomposition of signal modes, accurately selecting highly correlated modes to form a target subset. Then, a GRU is used to model and predict the frequency domain evolution trend, achieving the preservation of key features and the filtering of invalid disturbances under complex interference environments. Compared to traditional filters, this scheme possesses adaptive, learnable, and adjustable characteristics, improving the system's robustness and signal fidelity in dynamic interference scenarios.

[0145] In this embodiment, S6 specifically includes:

[0146] S61. The signal after interference suppression processing in step S5 is organized according to channel type and sampling time window, and a standardized multidimensional tensor data structure is constructed based on the time axis, frequency axis and channel axis to ensure that its structure can be directly parsed by the subsequent feature extraction model.

[0147] S62. Perform morphological feature extraction operation based on sparse dictionary learning on multidimensional tensor data. The sparse dictionary is pre-learned from the training sample library and can express typical waveform patterns of pulse-like signals such as partial discharge in a finite set of atoms.

[0148] S63. The input signal is decomposed into a sparse coefficient matrix and an atomic dictionary using a sparse coding algorithm, and the starting point, duration, waveform symmetry and consistency characteristics of the pulse pattern in the encoding process are identified.

[0149] S64. Based on the sparse decomposition results, further extract frequency domain and statistical feature indicators from the signal. The feature indicators include spectral entropy, which describes the complexity of the frequency component distribution; kurtosis factor, which measures the degree of signal spikes; instantaneous energy fluctuation rate, which reflects the range of short-term energy changes; frequency drift amplitude, which characterizes the dynamic shift of the main frequency position; and spectral distribution skewness, which determines the skew direction and amplitude of frequency energy component deviation.

[0150] S65. All extracted morphological features, frequency domain indices and statistics are concatenated along the feature dimension to form a composite feature vector in a unified format, and a set of corresponding feature expressions is generated for each time window.

[0151] S66. Perform dimensionality standardization, outlier removal, and integrity checks on the composite feature vector to ensure that it meets the input requirements and accuracy requirements of the subsequent state evaluation module.

[0152] S67. Output the processed composite feature vector set, and retain its corresponding timestamp, channel label and source information, as the input data basis for railway power transformer state modeling and anomaly identification.

[0153] A tensor-level feature extraction model is constructed using a sparse dictionary learning method, which enables high-dimensional modeling of pulse morphology, spectral shape, and statistical changes in signals, demonstrating significant advantages, especially in the detection of transient signals such as partial discharge. By combining parameters such as spectral entropy, kurtosis, and frequency drift to construct a composite feature vector, the model effectively enhances the discriminative power and expressive ability of abnormal features, providing key input features with structural, time-varying, and engineering interpretability for health status assessment.

[0154] In this embodiment, S7 specifically includes:

[0155] S71. Arrange the composite feature vector set in chronological order and construct a discretized state monitoring sequence according to a fixed time granularity, using the feature vector of each time period as a single node input unit.

[0156] S72. Construct a dynamic evolution graph structure using monitoring time points as graph nodes and Euclidean distance, cosine similarity, or spectral distance between feature vectors as graph edges to form a time-series feature graph based on feature correlation.

[0157] S73. Input the dynamic evolution graph structure into the health status assessment module. The health status assessment module is composed of a graph attention network and a multi-channel graph convolutional neural network. The graph attention network dynamically weights the edge weights in the graph to improve the expressive ability of key state transition paths.

[0158] S74. In graph attention networks, a multi-head graph attention mechanism is introduced to differentiate the correlation between different time points, and the node representation is updated through a weighted adjacency matrix.

[0159] S75. The dynamic evolution graph structure updated by graph attention is input into a multi-channel graph convolutional neural network. The multi-channel graph convolutional neural network uses a multi-dimensional channel grouping strategy to perform deep convolution operations on node features and extracts the state change trend and feature evolution pattern across time periods.

[0160] S76. Input the global graph embedding vector output by graph convolution into the fully connected classification layer to perform health level classification and abnormal state identification of the current state of the transformer. The identified state labels include normal, mild abnormality, moderate abnormality and severe fault.

[0161] S77. Based on the node classification results of multiple consecutive time periods, construct the running state trend curve and calculate the running trend slope, volatility and potential inflection point to predict the state evolution direction in the next few cycles.

[0162] S78. Output the health status classification label, operation trend score index and status transition path at the current time point, so that the remote monitoring platform can perform early warning response, operation and maintenance scheduling and equipment maintenance strategy optimization.

[0163] By constructing composite feature vectors into a time-series graph structure and employing graph attention networks and multi-channel graph convolutional neural networks for modeling, the evolution trend of transformer operating status across time periods was modeled, classified, and predicted. This structure possesses the ability to efficiently capture the spatiotemporal correlation between features and automatically identify potential fault nodes and their propagation paths. It significantly improves the system's early perception level of trend abrupt changes, degradation evolution, and abnormal clustering behavior, thereby enhancing prediction accuracy and interpretability.

[0164] In this embodiment, S8 specifically includes:

[0165] S81. Integrate the health status classification label, running trend scoring index and time series node graph embedding vector output in step S7, and construct a structured diagnostic result data package together with the spectrum, interference identification label and interference suppression residual value obtained in steps S3 to S6.

[0166] S82. Perform format standardization conversion on the structured diagnostic result data packet to generate a unified remote reporting format containing data timestamp, transformer number, channel source, anomaly type, feature summary, graph compression summary and status trend encoding fields;

[0167] S83. The formatted diagnostic result data is encapsulated through an embedded communication module, which includes multiple data transmission paths such as a cellular communication module, an Ethernet interface, and an industrial fiber optic transmission module.

[0168] S84. Select the communication protocol type according to the on-site deployment requirements, and use the lightweight message queue transmission protocol or the power protocol based on the IEC 61850 communication model for data packaging and transmission to ensure that the system can be compatible with different plant dispatching systems.

[0169] S85. Upload the diagnostic data to the remote monitoring platform, which supports structured data parsing, spectrum visualization, state trend curve reconstruction and interference distribution heat map display functions, forming a comprehensive diagnostic interface for multi-dimensional monitoring objects.

[0170] S86. Set alarm policies in the remote monitoring platform and automatically trigger multi-level early warning mechanisms based on the severity of status labels, trend score slope and disturbance level, including in-station audible and visual alarms, SMS push, control center pop-up windows and cloud platform alarm aggregation and distribution.

[0171] S87. Synchronously push the diagnostic results to the power dispatching system or equipment operation and maintenance management system, support the linkage execution of status refresh, task scheduling, maintenance work order issuance and inter-substation collaborative strategy switching, and build an intelligent operation and maintenance control network that supports closed-loop response at multiple sites.

[0172] S88. All uploaded records are cached locally and periodically backed up to the historical data management module for fault reproduction, model training optimization, and long-term performance evaluation.

[0173] A remote upload mechanism is constructed, encompassing standardized data formats, compatible communication protocols, and cloud-based linkage. Combining MQTT and IEC 61850 protocols, it enables efficient interaction of diagnostic results between edge devices and the cloud platform. The system supports spectrum visualization, status scoring display, and early warning control linkage, effectively establishing a data loop between on-site detection, intelligent diagnosis, and dispatch execution. This enhances the intelligent joint control capabilities and dispatch coordination response efficiency among multiple substations, meeting the high reliability and real-time intelligent operation and maintenance requirements of railway scenarios.

[0174] Example 1:

[0175] To verify the feasibility of this invention in practice, it was applied to a traction substation on a main railway line. The substation was equipped with an intelligent monitoring system for the operation status of power transformers. As a typical high-load, frequently starting and stopping traction power supply node, the substation operates in an environment with complex background noise such as strong power frequency disturbances, high-frequency pulse interference, and electromagnetic coupling interference. Traditional condition monitoring equipment is difficult to effectively identify early abnormal discharge signals, impulsive instability signals, and multiple superimposed interference characteristics. In addition, it suffers from data upload delays, low remote response efficiency, high equipment misjudgment rate, delayed maintenance, and unclear safety boundaries.

[0176] In this embodiment of the invention, the monitoring system deploys five signal acquisition channels—voltage, current, partial discharge, vibration, and electromagnetic radiation—on the primary side, secondary side, and key parts of the target transformer's casing. A multi-channel synchronous sampling mechanism is employed, with a sampling frequency reaching 10 MHz. Data from each channel is spliced ​​and buffered in real-time using a sliding window. Through unified clock synchronization control and anti-aliasing sampling strategies, the system can achieve high-precision time alignment of multi-source signals without relying on a third-party synchronization module, establishing a unified input foundation for subsequent analysis.

[0177] The acquired signal is first processed by a high-order statistical filter for offset and detrending removal, and then normalized by Z-score normalization. It is then input into the multi-scale residual fusion time-frequency transform network constructed in this invention. The network employs three sets of convolutional paths with different receptive fields to extract the signal's micro-perturbations, periodic fluctuations, and trend evolution. Combined with residual connections and channel attention mechanisms, it effectively prevents feature degradation. Furthermore, it dynamically adjusts the short-time Fourier window length based on frame energy and entropy functions, ultimately generating a three-dimensional time-frequency feature tensor.

[0178] The interference identification component uses a recognition network constructed with multi-channel temporal convolution and multi-head attention mechanisms to distinguish interference types. The training set contains samples of multiple interference categories, including power frequency harmonics, impulsive disturbances, low-frequency background interference, and non-Gaussian broadband interference. Test results show that the interference classification accuracy is consistently above 96%, the average interference identification delay is less than 1 second, and it exhibits strong adaptability and stability under different equipment loads and weather conditions.

[0179] The system then automatically invokes an interference suppression module based on variational mode decomposition and gated cyclic units to perform two-stage processing on the identified interference segments. The VMD part decomposes the signal into several modal components and filters highly correlated components according to the energy shift criterion, while the GRU model models and predicts the frequency domain evolution trend for reconstruction. After interference suppression, the average signal-to-noise ratio of the signal is improved from the original 8.7 dB to 16.4 dB, effectively enhancing the identifiability of low-amplitude weak pulses in the original signal.

[0180] The system further employs a sparse dictionary learning method to extract structural features from the interference-suppressed signal. It constructs a composite feature vector by combining multiple spectral domains and statistical features such as spectral entropy, kurtosis factor, and energy fluctuation rate. Furthermore, it classifies and predicts the operating status of equipment using a graph attention network and a graph convolution model. In on-site evaluation, the model can stably classify four states: "normal," "mildly abnormal," "moderately abnormal," and "severely abnormal," with an average classification accuracy of 94.2%. The trend prediction early warning rate exceeds 85% within three periods.

[0181] All diagnostic results are finally encapsulated by the edge computing terminal and uploaded to the remote monitoring platform, supporting the linkage of the dispatching system based on the MQTT protocol and IEC 61850 standard. The platform can automatically generate alarms, trend graphs and recommended maintenance work orders, and implement the collaborative strategy switching with adjacent sites. The operation data shows that after the system is deployed, the average fault identification delay is reduced by about 72%, and the false alarm rate is reduced by more than 60%, providing strong support for the continuous visual management of the operation and maintenance efficiency and the health status of the transformer.

[0182] This embodiment verifies that the present invention has the fault perception ability of full-link, high-precision and intelligent in the strong interference railway operation environment, which can significantly improve the practicability, precision and automation level of the transformer monitoring system, and has good engineering adaptability and popularization prospect.

[0183] The above is only the preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent replacement or change, and should be covered within the protection scope of the present invention.

Claims

1. A method of adaptive spectrum monitoring and interference mitigation for railway power transformers, characterized in that, The method comprises the following steps: S1, setting a multi-channel signal acquisition unit on the primary side and the secondary side of a railway power transformer to acquire original multi-source signal data; S2, preprocessing the original multi-source signal; S3, inputting the preprocessed original multi-source signal into a multi-scale residual fusion time-frequency transformation network to extract a time-frequency feature tensor of the original multi-source signal under different time windows and frequency ranges; S4, inputting the time-frequency feature tensor into an interference identification network with a fusion attention mechanism; S5, dynamically activating an interference suppression module according to the identification result, wherein the interference suppression module is composed of a signal decomposer based on variational modal decomposition and a gated recurrent unit sequence modeling network; S6, constructing the signal after interference suppression into a multi-dimensional tensor data structure, extracting pulse signal morphological features based on a sparse dictionary learning method to form a composite feature vector set; S7, inputting the composite feature vector into a health state evaluation module; S8, uploading the diagnosis result to a remote monitoring platform through an embedded communication module.

2. The adaptive spectrum monitoring and interference mitigation method for railway-oriented power transformers according to claim 1, characterized in that, The S1 specifically comprises: S11, setting a multi-channel signal acquisition unit on the primary side and the secondary side of a railway power transformer, wherein the multi-channel signal acquisition unit comprises a current and voltage sensor, a partial discharge signal coupling device, a vibration sensor and an electromagnetic radiation receiving device; S12, sampling the voltage signal and the current signal through a synchronous sampling module to obtain complete electrical transient information; S13, using a capacitor coupler or a special current probe to non-contact collect the partial discharge signal to capture the discharge transient pulse signal; S14, using a piezoelectric or optical fiber vibration sensor to monitor the micro-vibration behavior of the power transformer body structure and its installation foundation to form a mechanical response signal; S15, using a radio frequency receiving antenna or a near-field probe to collect the electromagnetic radiation signal around the transformer to capture the characteristics of potential radiation interference sources in the system; S16, uniformly clock synchronizing the above multi-channel signal acquisition unit, and starting synchronous collection under specified trigger conditions to form original multi-source signal data covering multiple physical field information.

3. The adaptive spectrum monitoring and interference mitigation method for railway-oriented power transformers according to claim 1, characterized in that, The S2 specifically comprises: S21, performing channel separation processing on the collected original multi-source signal data, and constructing independent time sequence channel buffer according to the signal type; S22, using a high-order statistical filtering method to correct the bias of each original signal to remove the direct current drift and baseline drift component; S23, performing amplitude normalization processing on the signal after bias correction, and using the Z-score standardization method to uniformly convert the amplitude distribution of various signals; S24, re-aligning the time axis of various signal channels based on the adjacent sampling block overlap strategy of the time window; S25, constructing the preprocessed multi-channel signal into an input segment signal sequence with a fixed length and an overlap rate.

4. The adaptive spectrum monitoring and interference mitigation method for railway-oriented power transformers according to claim 1, characterized in that, The S3 specifically comprises: S31, grouping the preprocessed signal data of various types according to the channel type, and constructing an input sequence tensor using a fixed length sliding window segmentation, wherein the length and step of each window are set according to the target frequency distribution and the calculation complexity requirement, and all windows maintain full coverage of the time axis during sliding. S32, input the input sequence tensor into a multi-scale residual fusion time-frequency transform network, the overall structure of the multi-scale residual fusion time-frequency transform network including a parallel multi-scale convolution path module, a deep residual connection module, an inter-layer channel enhancement module and a short-time Fourier transform module with adaptive window length control; S33, in the parallel multi-scale convolution path module, three independent convolution sub-paths are arranged, and different sizes of convolution kernels and expansion rates are used to extract local features of the signal under the time window, forming a cross-scale feature map group; S34, the cross-scale feature map group is spliced in the channel dimension and input into the deep residual connection module, which is composed of a standard residual unit, a multi-layer jump connection path, a layer-by-layer normalization structure and a residual fusion gate unit; S35, on the basis of the residual fusion output, an inter-layer channel enhancement module is introduced, which includes a channel attention mapping submodule, a spatial compression mapper and a scale adaptive activation function; S36, input the channel enhanced feature tensor into the window length adjustable short-time Fourier transform module, which calculates the optimal length and overlap rate of the sliding window based on the frame energy fluctuation rate and the instantaneous entropy function, and performs fast Fourier transform in each window to construct a multi-time period spectrum; S37, perform frequency domain interpolation alignment, amplitude frequency normalization and spectrum expansion on all multi-time period spectrum results to construct a three-dimensional time-frequency feature tensor, and input the three-dimensional time-frequency feature tensor into a dimension standardization module for format unification; S38, output the three-dimensional time-frequency feature tensor after standardization processing, which retains the time dynamic characteristics, frequency structure characteristics and scale internal dependence of the signal.

5. The adaptive spectrum monitoring and interference mitigation method for railway-oriented power transformers according to claim 1, characterized in that, The S4 specifically includes: S41, input the three-dimensional time-frequency feature tensor into the interference identification network A-DINet with fusion attention mechanism, the three-dimensional time-frequency feature tensor including three structural dimensions of time, frequency and channel, and input into the interference identification process according to the preset data batch; S42, in the interference identification network, feature extraction is performed through a multi-channel time series convolution module, which is composed of multiple channel-independent time series convolution units, each of which performs one-dimensional or two-dimensional convolution operation on different frequency bands to capture short-term dependence features of power frequency disturbance, harmonic interference, shock wave band and broadband noise; S43, map the features after convolution of each frequency band to the channel dimension for splicing to construct an intermediate fusion tensor, and input into a multi-head attention encoding module, which includes multiple parallel attention heads, each of which models the feature correlation in different scales and different frequency bands; S44, calculate the information mutual dependence degree between each frequency band using the attention distribution matrix, and dynamically adjust the channel weight coefficient while suppressing the response value of irrelevant frequency domain regions; S45, input the enhanced feature tensor into a category correlation analyzer module, which includes a disturbance type discrimination unit, a channel feature fusion layer and a class-specific scorer; S46, according to the category score result, the input tensor is disturbed and the label is assigned, and the disturbance type includes 50Hz power fundamental wave interference, 50Hz integer harmonic interference, transient impact interference caused by high-voltage switch operation, non-stationary low-frequency disturbance, and wideband non-Gaussian noise interference with statistical adaptive characteristics; S47, the final identified interference label is output together with the frequency domain position index.

6. The adaptive spectrum monitoring and interference mitigation method for railway-oriented power transformers according to claim 1, characterized in that, The S5 specifically comprises: S51, according to the interference type label and the corresponding frequency band index obtained in step S4, the parameter configuration scheme of the interference suppression module is dynamically called, and the signal processing path matched with the current interference characteristics is activated; S52, the original signal data is input into the signal decomposer module based on variational modal decomposition, the signal decomposer module performs multi-order modal decomposition on the signal to generate a plurality of intrinsic modal function components, and each intrinsic modal function component represents a signal component at different frequency scales; S53, the energy offset index, the spectral density distribution characteristics and the similarity score between the target characteristic template are calculated for all intrinsic modal function components in turn, and a plurality of modal components are selected to form a target modal set according to a preset correlation criterion; S54, the target modal set is constructed as a time series input and sent to a sequence modeling network composed of a gated recurrent unit, the sequence modeling network learns the amplitude variation law and frequency domain transfer path between the modes through a recursive structure, and outputs a predicted frequency response sequence; S55, difference inversion is performed on the original modal signal according to the frequency domain prediction result output by the sequence modeling network, and frequency filtering and reconstruction operations are performed; S56, residual evaluation and dynamic update operation is performed on the enhanced signal, if the difference between the current result and the disturbance template is still greater than the threshold, the decomposition path fine tuning process is triggered again until the interference suppression result is stable; S57, the signal sequence after interference suppression processing is output, and the abnormal characteristics with physical meaning in the target frequency band are retained.

7. The adaptive spectrum monitoring and interference mitigation method for railway-oriented power transformers according to claim 1, characterized in that, The S6 specifically comprises: S61, the signal after interference suppression processing in step S5 is sorted according to the channel type and the sampling time window, and is constructed into a standardized multi-dimensional tensor data structure based on the time axis, the frequency axis and the channel axis; S62, morphological feature extraction operation is performed on the multi-dimensional tensor data based on sparse dictionary learning, the sparse dictionary is obtained by pre-learning in the training sample library, and can express the typical waveform mode of partial discharge pulse class signal in a limited atomic set; S63, the input signal is decomposed into a sparse coefficient matrix and an atomic dictionary combination by using a sparse coding algorithm, and the starting point, duration, waveform symmetry and consistency characteristics of the pulse form in multiple channels are identified in the coding process; S64, combined with the sparse decomposition result, frequency domain and statistical characteristic indexes of the signal are further extracted, the characteristic indexes include spectral entropy describing the complexity of frequency component distribution, kurtosis factor measuring the degree of signal peak, instantaneous energy fluctuation rate reflecting the change range of short-time energy, frequency drift amplitude describing the dynamic offset of main frequency position, and spectral distribution skewness judging the skewness direction and amplitude of frequency energy distribution; S65, splice all extracted morphological features, frequency domain indicators and statistics in the feature dimension to form a composite feature vector in a unified format, and generate a corresponding feature expression set for each time window; S66, perform dimension standardization, outlier rejection and integrity check on the composite feature vector; S67, output the processed composite feature vector set, and retain its corresponding time stamp, channel label and source information.

8. The adaptive spectrum monitoring and interference mitigation method for railway-oriented power transformers according to claim 1, characterized in that, The S7 specifically comprises: S71, arrange the composite feature vector set in chronological order, and construct a discretized state monitoring sequence according to a fixed time granularity, taking the feature vector of each time period as a single node input unit; S72, construct a dynamic evolution graph structure with the monitoring time point as the graph node, and the Euclidean distance, cosine similarity or spectral distance between feature vectors as the graph edge, to form a time sequence feature graph based on feature correlation; S73, input the dynamic evolution graph structure into a health state evaluation module, which is composed of a graph attention network and a multi-channel graph convolutional neural network; S74, in the graph attention network, introduce a multi-head graph attention mechanism to model the correlation between different time points differently, and update the node representation through a weighted adjacency matrix; S75, input the dynamic evolution graph structure updated by the graph attention into the multi-channel graph convolutional neural network, which uses a multi-dimensional channel grouping strategy to perform deep convolution operation on node features, extracts state change trend and feature evolution mode across time periods; S76, input the global graph embedding vector output by the graph convolution into a fully connected classification layer to perform health level classification and abnormal state recognition of the transformer current state, and the state label includes normal, mild abnormal, moderate abnormal and severe failure; S77, based on the node classification results of continuous multiple time periods, construct a running state trend curve, and calculate the running trend slope, fluctuation degree and potential inflection point; S78, output the health state classification label, running trend score index and state transition path of the current time point.

9. The adaptive spectrum monitoring and interference mitigation method for railway-oriented power transformers according to claim 1, characterized in that, The S8 specifically comprises: S81, integrate the health state classification label, running trend score index and time sequence node graph embedding vector output in step S7, and construct a structured diagnostic result data package together with the frequency spectrum, interference identification label and interference suppression residual value obtained in steps S3 to S6; S82, perform format standardization conversion on the structured diagnostic result data package to generate a unified remote reporting format containing data timestamp, transformer number, channel source, abnormal type, feature abstract, graph compression abstract and state trend coding field; S83, package the formatted diagnostic result data through an embedded communication module, which includes multiple data transmission paths such as a cellular communication module, an Ethernet interface and an industrial optical fiber transmission module; S84, select a communication protocol type according to the field deployment requirements, and perform data packaging and transmission through a lightweight message queue transmission protocol or a power protocol based on the IEC61850 communication model; S85, upload the diagnostic data to a remote monitoring platform, which supports functions of structured data analysis, spectrum diagram visualization, state trend curve reconstruction, and disturbance distribution heat map display, to form a comprehensive diagnostic interface for multi-dimensional monitoring objects; S86, set an alarm strategy in the remote monitoring platform, and automatically trigger a multi-level early warning mechanism according to the state label severity, trend score slope, and disturbance level, including in-station sound and light alarm, short message push, control center pop-up window, and cloud platform alarm aggregation and distribution; S87, push the diagnostic results to a power dispatching system or a device operation and maintenance management system, support linkage execution state refresh, task scheduling, maintenance work order issuance, and substation interworking strategy switching, and build an intelligent operation and maintenance control network supporting multi-site closed-loop response; S88, record all uploaded data in a local cache and periodically backup to a historical data management module.

Images