Self-adaptive frequency spectrum monitoring and interference suppression method for railway power transformer

Abstract

The invention discloses a self-adaptive frequency spectrum monitoring and interference suppression method for a railway power transformer. The method comprises 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, and extracting a time-frequency feature tensor; s4, inputting the time-frequency feature tensor into the interference identification network fused with the attention mechanism; s5, dynamically activating an interference suppression module according to an 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 the embedded communication module. According to the invention, multi-dimensional perception and adaptive modeling are fused, and intelligent identification and remote monitoring of railway transformer faults are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent operation and maintenance of power systems and state monitoring of rail transit equipment, and particularly relates to an adaptive spectrum monitoring and interference suppression method for railway power transformers. BACKGROUND

[0002] In the current railway power system, as the core power transmission and distribution equipment, the real-time monitoring and fault early warning of the operation state of the power transformer are of key significance to the safety of train traction power supply. Traditional transformer state monitoring relies on offline detection, single-point sensing and experience-based judgment, which is difficult to adapt to the complex characteristics of high-intensity interference, mixed multi-source signals and rapid fault evolution in the railway operating environment. Especially in high-speed railways, heavy-load railways and urban rail transit, electrical equipment is subjected to the combined effects of frequent start-stop, high-frequency impact, electromagnetic disturbance and structural vibration for a long time, making the early features of typical faults such as partial discharge, insulation deterioration and winding looseness extremely weak and easily covered by background noise, seriously affecting the sensitivity and accuracy of existing state monitoring systems.

[0003] In the prior art, spectrum analysis is often used as a key means for diagnosing faults in electrical signals, but it mostly uses Fourier transform with fixed window length, short-time analysis or static wavelet transform methods, which lack adaptability to dynamic changes in signals. Such techniques have insufficient resolution and weak feature mutation recognition when faced with non-stationary, multi-frequency interference signals in the railway environment, and cannot effectively extract key diagnostic information. In addition, in terms of interference suppression, traditional methods usually use simple filters or set threshold rules to remove interference signals, but have limited processing capacity for multiple types of superimposed interference, spectrum overlap or non-Gaussian background, which easily causes loss of useful signals or misidentification.

[0004] In view of the feature extraction requirements of transformer partial discharge, harmonic anomaly and mechanical fault and other typical faults, existing methods also have insufficient modeling of the coordinated changes between 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-substation scenarios, current methods mostly do not have a unified data structure, intelligent analysis and remote collaborative scheduling mechanism, and cannot 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 is just to solve the problems of poor spectrum sensing ability, weak interference adaptability, low feature modeling capability and lack of remote linkage mechanism in the prior art. By constructing a multi-channel high-precision acquisition mechanism, fusing a multi-scale residual network and an attention-enhanced recognition module, deep deconstruction and interference recognition of the signal are realized; then, by combining variational modal decomposition and a gated recurrent unit network, an adaptive filtering mechanism is constructed to effectively improve the target signal extraction capability; further, by using a graph neural network, fault trend modeling is realized, and finally, a full-link intelligent monitoring closed loop that can be remotely uploaded and cooperated by multiple stations is formed, which significantly improves the state perception and abnormal response capability of railway power transformers in complex operating scenarios.

[0006] Therefore, how to provide an adaptive spectrum monitoring and interference suppression method for railway power transformers is a problem that those skilled in the art need to solve. SUMMARY

[0007] One object of the present application is to provide an adaptive spectrum monitoring and interference suppression method for railway power transformers. The present application fuses multi-channel perception, spectrum intelligent extraction and deep interference suppression technology, realizes state trend modeling by combining a graph neural network, and constructs a closed-loop monitoring system through a remote communication module, thereby solving the problems of difficult fault identification, weak feature extraction and delayed early warning response of railway transformers in complex interference environments, and improving the diagnostic accuracy and system intelligence level.

[0008] The adaptive spectrum monitoring and interference suppression method for railway power transformers according to the embodiment of the present application comprises the following steps: S1, a multi-channel signal acquisition unit is arranged on the primary side and the secondary side of the railway power transformer to simultaneously acquire original multi-source signal data including voltage signals, current signals, partial discharge signals, mechanical vibration signals and electromagnetic radiation signals; S2, the original multi-source signals are preprocessed, and the preprocessing includes removing DC bias by using a high-order statistical filter and normalizing the amplitude interval by Z-score normalization; S3, the preprocessed signals are input into a multi-scale residual fusion time-frequency transform network, which includes parallel multi-scale convolution submodules, residual connection modules and variable window length short-time Fourier transformers, for extracting time-frequency feature tensors of the signals under different time windows and frequency ranges; S4, the time-frequency feature tensors are input into an interference recognition network A-DINet that fuses attention mechanisms, which is constructed based on a multi-head attention mechanism, a multi-channel time series convolution network and a category correlation analyzer, to identify interferences including but not limited to 50 Hz and its integer multiple power frequency harmonics, transient pulse interferences caused by high-voltage switches, non-stationary low-frequency disturbances and wideband non-Gaussian noise; S5, dynamically activating an interference suppression module according to the identification result, the module being composed of a signal decomposer based on variational mode decomposition (VMD) and a gated recurrent unit (GRU) sequence learning network, and realizing by a two-stage strategy: in the first stage, the original signal is decomposed by VMD and the high correlation component is selected based on the shift energy criterion; in the second stage, the GRU model is used to model the frequency domain sequence to realize amplitude-frequency response prediction and target reconstruction, so as to realize target signal enhancement and interference suppression; S6, constructing the signal after interference suppression into a multi-dimensional tensor data structure, extracting pulse signal morphological features based on sparse dictionary learning method, combining spectral entropy, kurtosis factor, instantaneous energy fluctuation rate, frequency drift amplitude, spectral distribution skewness and other statistical and spectral domain features to form a composite feature vector set for diagnostic modeling; S7, inputting the composite feature vector into a health state evaluation module, the module being composed of a graph attention network (GAT) and a multi-channel graph convolutional neural network (GCN), and completing abnormal identification, classification discrimination and operation trend prediction of the transformer state by constructing a dynamic evolution graph structure with nodes being monitoring time points and edges being feature similarity; S8, uploading the diagnostic results including spectral graph, interference category, fault label and state trend score to a remote monitoring platform through an embedded communication module, and supporting linkage with a dispatching system based on MQTT protocol or IEC 61850 standard to complete intelligent early warning, trend visualization and closed-loop operation of multi-substation point collaborative control.

[0009] Optionally, the S1 specifically comprises: S11, arranging a multi-channel signal acquisition unit on the primary side and the secondary side of the railway power transformer, respectively, the multi-channel signal acquisition unit comprising a current and voltage sensor, a partial discharge signal coupling device, a vibration sensor and an electromagnetic radiation receiving device; S12, high-frequency sampling of the voltage signal and the current signal by a high-speed synchronous sampling module to obtain complete electrical transient information; S13, non-contact high-sensitivity acquisition of the partial discharge signal by a high-frequency capacitive coupler or a special current probe to capture the discharge transient pulse signal; S14, monitoring the micro-vibration behavior of the power transformer body structure and its installation foundation by a piezoelectric or optical fiber vibration sensor to form a mechanical response signal; S15, acquiring the electromagnetic radiation signal around the transformer by a wideband radio frequency receiving antenna or a near-field probe to capture the characteristics of potential radiation interference sources in the system; S16, uniformly clock-synchronized control of the above-mentioned multi-channel signal acquisition unit, and starting synchronous acquisition under specified trigger conditions to form original multi-source signal data covering multiple physical field information.

[0010] Optionally, 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 signal type; S22, performing bias correction on each original signal by using high-order statistical filtering method to remove direct current drift and baseline drift components; S23, performing amplitude normalization processing on the signal after bias correction, and using Z-score standardization method to uniformly convert the amplitude distribution of each type of signal to ensure the data scale consistency between different physical quantities; S24, re-aligning the time axis of each type of signal channel based on the adjacent sample block overlap strategy of time window to ensure strict time synchronization of multi-channel data under the same analysis window; S25, constructing the pre-processed multi-channel signal into an input segment signal sequence with fixed length and overlap rate for subsequent frequency spectrum analysis module.

[0011] Optionally, S3 specifically comprises: S31, grouping each type of signal data after preprocessing according to channel type, and constructing an input sequence tensor by using a fixed length sliding window, 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 in the sliding process to avoid sample loss; S32, sending the input sequence tensor into a multi-scale residual fusion time-frequency transform network, wherein the overall structure of the multi-scale residual fusion time-frequency transform network includes parallel multi-scale convolution path modules, deep residual connection modules, inter-layer channel enhancement modules, and short-time Fourier transform modules with adaptive window length control; S33, in the parallel multi-scale convolution path module, setting three or more independent convolution sub-paths, and respectively using convolution kernels and dilation rates of different sizes to extract local features of signals under micro-scale (such as 0.1 ms level), mesoscale (such as 1 ms level), and macro-scale (such as 10 ms level) time windows to form a cross-scale feature map group; S34, concatenating the cross-scale feature map group in the channel dimension to enter the deep residual connection module, wherein the deep residual connection module is composed of standard residual units, multi-layer jump connection paths, layer-by-layer normalization structure, and residual fusion gate units, and is used to improve the stability of deep feature expression and relieve the gradient vanishing problem; S35, introducing an inter-layer channel enhancement module based on the residual fusion output, wherein the inter-layer channel enhancement module includes a channel attention mapping submodule, a spatial compression mapper, and a scale adaptive activation function, and is used to improve the response value of highly correlated channels and suppress irrelevant channel interference; S36, input the channel enhanced feature tensor 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-period spectrum graph; S37, perform frequency domain interpolation alignment, amplitude frequency normalization and spectrum expansion on all multi-period spectrum graph results to construct a three-dimensional time-frequency feature tensor of time×frequency×channel dimension, and send the three-dimensional time-frequency feature tensor to a dimension standardization module for format unification to meet the input requirements of the downstream model; 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, and is used to support the category discrimination and disturbance modeling process of the subsequent interference identification network.

[0012] Optionally, the S4 specifically comprises: 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 includes three structural dimensions of time, frequency and channel, and is sent into the interference identification process according to a preset data batch; S42, in the interference identification network, feature extraction is performed through a multi-channel time sequence convolution module, the multi-channel time sequence convolution module is composed of multiple channel independent time sequence convolution units, each time sequence convolution unit 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 send the intermediate fusion tensor into a multi-head attention encoding module, the multi-head attention encoding module includes multiple parallel attention heads, each attention head respectively models the feature correlation in different scales and different frequency bands; S44, calculate the information mutual dependence degree between each frequency band by using an attention distribution matrix, and then dynamically adjust the channel weight coefficient to realize enhancement processing of the spectrum band with disturbance significance while suppressing the response value of the irrelevant frequency domain area; S45, input the enhanced feature tensor into a category correlation analyzer module, the category correlation analyzer module includes a disturbance type discrimination unit, a channel feature fusion layer and a class-specific scorer, which are respectively used for extracting category-level prior features, performing cross-channel aggregation in frequency band dimension, and outputting an activation score vector for a specific interference type; S46, according to the category score result, the input tensor is disturbed and labeled, and the disturbance type includes 50 Hz power fundamental wave interference, 50 Hz 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, output the finally identified interference label together with the frequency domain position index as the basis for dynamic scheduling of the subsequent interference suppression module, and call the abnormal event tracking mechanism.

[0013] Optionally, the S5 specifically includes: S51, according to the interference type label and the 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 matched with the current interference characteristics; S52, input the original signal data 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, each intrinsic modal function component represents a signal component at different frequency scales; S53, calculate the energy offset index, the spectral density distribution characteristics and the similarity score between the target characteristic template for all intrinsic modal function components in turn, and select a plurality of modal components to form a target modal set according to a predetermined correlation criterion; S54, construct the target modal set as a time series input and send it 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 according to the frequency domain prediction result output by the sequence modeling network and the original modal signal, and frequency filtering and reconstruction are performed to realize the suppression of non-target interference components and the enhancement of target signals; S56, perform residual evaluation and dynamic update operation on the enhanced signal, if the difference between the current result and the disturbance template is still greater than the threshold, retrigger the decomposition path fine-tuning process until the interference suppression result is stable; S57, output the signal sequence after interference suppression processing, retain the abnormal characteristics with physical meaning in the target frequency band, and significantly reduce the influence of multi-source interference components on the subsequent analysis model.

[0014] Optionally, the S6 specifically includes: S61, arrange the signal after the interference suppression processing of step S5 according to the channel type and the sampling time window, and construct a standardized multi-dimensional tensor data structure based on the time axis, the frequency axis and the channel axis, to ensure that the structure can be directly analyzed by the subsequent feature extraction model; S62, performing a morphology feature extraction operation based on sparse dictionary learning on the multi-dimensional tensor data, the sparse dictionary being pre-learned from a training sample library and capable of expressing typical waveform patterns of pulse signals such as partial discharge in a limited atomic set; S63, decomposing the input signal into a sparse coefficient matrix and an atomic dictionary combination by using a sparse coding algorithm, and identifying the starting point, duration, waveform symmetry and consistency features of the pulse morphology on multiple channels in the coding process; S64, combining the sparse decomposition results to further extract frequency domain and statistical feature indicators from the signal, the feature indicators including a spectral entropy describing the complexity of frequency component distribution, a kurtosis factor measuring the degree of signal peaks, an instantaneous energy fluctuation rate reflecting the range of short-time energy change, and a frequency drift amplitude depicting the dynamic shift of the main frequency position; and a spectral distribution skew rate judging the skew direction and amplitude of the frequency energy distribution; S65, concatenating all the extracted morphology features, frequency domain indicators and statistics in the feature dimension to form a composite feature vector in a unified format, and generating a corresponding feature expression for each time window; S66, performing dimension standardization, outlier rejection and integrity check on the composite feature vector to ensure that it meets the input requirements and accuracy requirements of the subsequent state evaluation module; S67, outputting the processed composite feature vector set and retaining its corresponding time stamp, channel label and source information as the input data basis for railway power transformer state modeling and abnormality identification.

[0015] Optionally, the S7 specifically includes: S71, arranging the composite feature vector set in chronological order and constructing 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, constructing a dynamic evolution graph structure with the monitoring time points as graph nodes and the Euclidean distance, cosine similarity or spectral distance between feature vectors as graph edges to form a time sequence feature graph based on feature correlation; S73, inputting the dynamic evolution graph structure into a health state evaluation module, the health state evaluation module being composed of a graph attention network and a multi-channel graph convolutional neural network, wherein the graph attention network dynamically weights the edge weights in the graph to improve the expression ability of the key state transition path; S74, introducing a multi-head graph attention mechanism in the graph attention network 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 attention of the graph into a multi-channel graph convolutional neural network, which performs deep convolution operation on node features by using a multi-dimensional channel grouping strategy to extract 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 the continuous multiple time periods, construct the operation state trend curve, and calculate the operation trend slope, fluctuation degree and potential inflection point, which are used to predict the state evolution direction in the future period; S78, output the health state classification label, operation trend score index and state transition path at the current time point, which are used for remote monitoring platform to perform early warning response, operation and maintenance scheduling and equipment maintenance strategy optimization.

[0016] Optionally, the S8 specifically comprises: S81, integrate the health state classification label, operation trend score index and time series node graph embedding vector output in step S7, and construct a structured diagnostic result data package together with the frequency spectrum graph, interference recognition 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 a plurality of 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 on-site deployment requirements, and perform data packaging and transmission through a lightweight message queue transmission protocol or a power protocol based on IEC 61850 communication model to ensure that the system can be compatible with different plant station dispatching systems; S85, upload the diagnostic data to the remote monitoring platform, which supports structured data analysis, spectrum graph visualization, state trend curve reconstruction and disturbance distribution heat map display functions, forming a comprehensive diagnostic interface for multi-dimensional monitoring objects; S86, set the 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, SMS push, control center pop-up window and cloud platform alarm aggregation distribution; S87, the diagnostic result is pushed to the power dispatching system or the equipment operation and maintenance management system, state refreshing, task scheduling, maintenance work order issuing and substation cooperation strategy switching are supported, and an intelligent operation and maintenance control network supporting multi-site closed-loop response is constructed; S88, all uploaded records are recorded in the local cache and periodically backed up to the historical data management module, for fault reproduction, model training optimization and long-term performance evaluation.

[0017] The beneficial effects of the present application are: The present application realizes a multi-faceted technical breakthrough on the basis of the prior art by constructing an adaptive spectrum monitoring and interference suppression method for railway power transformers, significantly improving the adaptability, accuracy and intelligent level of the transformer state monitoring system in a complex electromagnetic interference environment.

[0018] Firstly, the present application realizes multi-physical quantity and multi-channel synchronous acquisition in the signal acquisition link, covering five typical parameters of voltage, current, partial discharge, vibration and electromagnetic radiation, and effectively improves the timeliness and integrity of the original signal acquisition through high-frequency and high-precision acquisition and unified clock control mechanism, providing a high-quality data basis for subsequent spectrum modeling. Compared with the traditional single electric quantity monitoring method, this method constructs a rich information dimension, so that the weak abnormal characteristics are no longer hidden by background interference.

[0019] Secondly, in the aspect of spectrum feature extraction, a multi-scale residual fusion time-frequency transform network is adopted, which realizes fine description of signals in different frequency and time scales through the synergistic effect of multi-scale convolution, residual connection and adaptive window length short-time Fourier decomposition structures. This structure effectively makes up for the problem of slow response of traditional fixed window analysis to non-stationary signals, significantly improves the recognition ability of fine-grained features such as partial discharge and harmonic disturbance. The interference recognition network combined with the attention mechanism further enhances the intelligent discrimination ability of the system to power frequency harmonics, transient pulses and wideband non-Gaussian noise, and realizes automatic classification of signal categories.

[0020] In addition, in terms of interference suppression, the present invention introduces a two-stage processing strategy that combines variational mode decomposition with a gated recurrent neural network, which effectively completes the decomposition, identification and predictive removal of interference components, ensuring that the target signal maintains its integrity while removing invalid information to the maximum extent possible. By modeling and dynamically reconstructing the amplitude-frequency response, this method takes into account both the suppression effect and signal fidelity, and overcomes the problem of erroneous deletion of characteristic signals by traditional filtering methods. In the feature modeling and state assessment link, the present invention extracts pulse morphology features based on sparse dictionary learning, combines a variety of statistical and spectral parameters to construct a composite feature vector, and realizes dynamic identification and trend prediction of the transformer operating state through the joint modeling of graph attention network and graph convolutional neural network. This modeling method has the ability to express states across time scales, which improves the system's early perception of abnormal evolution.

[0021] Finally, the 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 upload of diagnostic results, push notifications, and intelligent, coordinated control of multiple substations, creating a comprehensive, closed-loop intelligent operation and maintenance system from front-end perception to cloud-based control. Overall, this method possesses significant practicality, engineering value, and promising application prospects in intelligent railway power supply systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0023] Fig. 1 This is a flow chart of the adaptive spectrum monitoring and interference suppression method for railway power transformers proposed by the present invention; Fig. 2 This is a schematic diagram of the network architecture of the multi-scale residual fusion time-frequency transform network proposed in the present invention; Fig. 3 This is the model structure diagram of the interference identification network A-DINet that integrates the attention mechanism proposed in this invention. DETAILED DESCRIPTION

[0024] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0025] refer to Figs. 1-3 , an adaptive spectrum monitoring and interference suppression method for railway power transformers, comprising the following steps: S1, a multi-channel signal acquisition unit is arranged on the primary side and the secondary side of the railway power transformer, and original multi-source signal data including voltage signals, current signals, partial discharge signals, electromechanical vibration signals and electromagnetic radiation signals are collected simultaneously; S2, the original multi-source signals are preprocessed, and the preprocessing includes: removing DC bias by using a high-order statistical filter, and normalizing the amplitude interval by Z-score normalization; S3, the preprocessed signals are input into a multi-scale residual fusion time-frequency transform network, the network includes a parallel multi-scale convolution submodule, a residual connection module and a variable window short-time Fourier transformer, and is used for extracting time-frequency feature tensors of the signals under different time windows and frequency ranges; S4, the time-frequency feature tensors are input into an interference identification network A-DINet with a fusion attention mechanism, the network is constructed based on a multi-head attention mechanism, a multi-channel time sequence convolution network and a category correlation analyzer, and identifies interferences including but not limited to the following: 50 Hz and its integer multiple power frequency harmonics, transient pulse interferences caused by high-voltage switches, non-stationary low-frequency disturbances and wideband non-Gaussian noises; S5, according to the identification result, a disturbance suppression module is dynamically activated, the module is composed of a signal decomposer based on a variational mode decomposition (VMD) and a gated recurrent unit (GRU) sequence learning network, and a two-stage strategy is adopted: in the first stage, the original signal is decomposed into modes by using VMD, and a high correlation component is selected based on a shift energy criterion; in the second stage, a GRU model is used to model the frequency domain sequence, to realize amplitude-frequency response prediction and target reconstruction, so as to realize target signal enhancement and interference suppression; S6, the signal after the interference is suppressed is constructed into a multi-dimensional tensor data structure, pulse signal pattern features are extracted based on a sparse dictionary learning method, and statistical and spectral domain features such as spectral entropy, kurtosis factor, instantaneous energy fluctuation rate, frequency drift amplitude and spectral distribution skewness are combined to form a composite feature vector set for diagnosis modeling; S7, the composite feature vector is input into a health state evaluation module, the module is composed of a graph attention network (GAT) and a multi-channel graph convolutional neural network (GCN), a dynamic evolution graph structure with monitoring time points as nodes and feature similarity as edges is constructed, and abnormal identification, classification discrimination and operation trend prediction of the transformer state are completed; S8, the diagnosis results including a spectrum diagram, an interference category, a fault label and a state trend score are uploaded to a remote monitoring platform through an embedded communication module, and linkage with a dispatching system based on an MQTT protocol or an IEC 61850 standard is supported, to complete intelligent early warning, trend visualization and closed-loop operation of multi-substation point collaborative control.

[0026] The application realizes joint processing of multi-source heterogeneous signals and dynamic spectrum modeling driven by deep learning in the running environment of railway transformers through constructing an intelligent monitoring framework covering the whole process from signal acquisition, preprocessing, time-frequency feature extraction, interference identification, interference suppression, feature modeling to remote linkage, effectively overcomes the problems of weak spectrum sensing ability, large model rigidity and broken response chain of traditional monitoring systems in non-stationary interference scenarios, and realizes accurate extraction and closed-loop diagnosis of weak faults under complex operating conditions.

[0027] In the embodiment, the S1 specifically includes: S11, a multi-channel signal acquisition unit is arranged on the primary side and the secondary side of the railway power transformer respectively, 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; S12, a high-frequency sampling module is used to sample the voltage signal and the current signal at high frequency to obtain complete electrical transient information; S13, a high-frequency capacitor coupler or a special current probe is used to non-contact high-sensitivity acquisition of the partial discharge signal to capture the discharge transient pulse signal; S14, a piezoelectric or optical fiber vibration sensor is used to monitor the micro-vibration behavior of the power transformer body structure and the installation foundation thereof to form a mechanical response signal; S15, a wideband radio frequency receiving antenna or a near-field probe is used to acquire the electromagnetic radiation signal around the transformer to capture the characteristics of potential radiation interference sources in the system; S16, the above multi-channel signal acquisition unit is controlled by a unified clock synchronization, and synchronous acquisition is started under a specified trigger condition to form original multi-source signal data covering multiple physical field information.

[0028] By deploying a multi-channel acquisition system on the primary side and the secondary side of the railway power transformer and integrating electrical, mechanical and electromagnetic three types of physical field sensing dimensions, a five-dimensional signal acquisition network covering voltage, current, partial discharge, structure vibration and radiation interference is formed, combined with a unified clock synchronization mechanism, not only the signal alignment with millisecond level precision is realized, but also the integrity of the original data in time sequence continuity, spatial coverage and fault scene representation is greatly improved, providing a reliable basis for subsequent time-frequency analysis.

[0029] In the embodiment, the S2 specifically includes: S21, the original multi-source signal data acquired is subjected to channel separation processing, and independent time sequence channel buffer areas are respectively constructed according to the signal types; S22, a high-order statistical filtering method is used to correct the bias of each original signal to remove the direct current drift and baseline drift components; S23, perform amplitude normalization processing on the bias-corrected signal, and use Z-score standardization method to uniformly convert the amplitude distribution of each type of signal, to ensure that the data scales of different physical quantities are consistent; S24, re-align the time axis of each type of signal channel based on the adjacent sampling block overlap strategy of the time window, to ensure that the multi-channel data has strict time synchronization under the same analysis window; S25, construct the pre-processed multi-channel signal into an input segment signal sequence with a fixed length and overlap rate for subsequent spectral analysis module.

[0030] By introducing high-order statistical filtering and normalization strategy, the method can eliminate the baseline offset, amplitude inconsistency and time misalignment problems caused by aging of field devices, difference in inductive coupling and sampling drift, ensure the high consistency of multi-channel, multi-type signals in frequency, amplitude and time sequence, effectively improve the adaptability and stability of input data to neural network model, and significantly enhance the robustness and feature fidelity of the spectral extraction link.

[0031] In the embodiment, the S3 specifically includes: S31, group the pre-processed signal data of each type according to the channel type, and use a fixed length sliding window to segment and construct an input sequence tensor, 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 to avoid sample loss; S32, input the input sequence tensor into a multi-scale residual fusion time-frequency transform network, and 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 a short-time Fourier transform module with adaptive window length control; S33, in the parallel multi-scale convolution path module, three or more independent convolution sub-paths are set, different sizes of convolution kernels and dilation rates are used to extract local features of signals in micro-scale (such as 0.1ms level), mesoscale (such as 1ms level) and macro-scale (such as 10ms level) time windows, and a cross-scale feature map group is formed; S34, splice the cross-scale feature map group in the channel dimension and input it into the deep residual connection module, and the deep residual connection module 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, which is used to improve the stability of deep feature expression and relieve the gradient vanishing problem; S35, introduce an inter-layer channel enhancement module based on the residual fusion output, and the inter-layer channel enhancement module includes a channel attention mapping submodule, a spatial compression mapper and a scale adaptive activation function, which is used to improve the response value of highly correlated channels and suppress irrelevant channel interference; S36, input the channel enhanced feature tensor 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, performs fast Fourier transform in each window to construct a multi-time period spectrum graph; S37, perform frequency domain interpolation alignment, amplitude frequency normalization and spectrum graph expansion on all multi-time period spectrum graph results to construct a three-dimensional time-frequency feature tensor of time x frequency x channel dimensions, and send the three-dimensional time-frequency feature tensor to a dimension standardization module for format unification to meet the input requirements of the downstream model; 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, and is used to support the category discrimination and disturbance modeling process of the subsequent interference identification network.

[0032] A multi-scale residual fusion time-frequency transform network is constructed, local and global features are extracted through small-scale, medium-scale and large-scale parallel convolution, residual connection is combined to avoid gradient disappearance and feature damage, channel attention mechanism is used to fine control the information flow path, and the frequency analysis granularity is dynamically adjusted through adaptive window length STFT, which realizes accurate representation of the signal in multiple time scales and multiple frequency levels, and effectively solves the technical problem of unbalanced feature extraction in the coexistence scene of partial discharge, high-frequency impulse and slow-changing interference.

[0033] In the embodiment, the S4 specifically comprises: 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 includes three structural dimensions of time dimension, frequency dimension and channel dimension, and is sent into the interference identification process according to a preset data batch; S42, in the interference identification network, feature extraction is performed through a multi-channel time sequence convolution module, the multi-channel time sequence convolution module is composed of multiple channel independent time sequence convolution units, each time sequence convolution unit performs one-dimensional or two-dimensional convolution operation on different frequency bands to capture short-term dependence features of power frequency disturbance, harmonic interference, impulse 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 send the intermediate fusion tensor into a multi-head attention encoding module, the multi-head attention encoding module includes multiple parallel attention heads, each attention head respectively models the feature correlation in different scales and different frequency bands; S44, calculate the information mutual dependence degree between each frequency band by using an attention distribution matrix, and then dynamically adjust the channel weight coefficient to realize enhancement processing of the spectrum band with disturbance significance while suppressing the response value of the irrelevant frequency domain area; 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, respectively used for extracting a class-level prior feature, performing cross-channel aggregation in a frequency band dimension, and outputting an activation score vector for a specific disturbance type; S46, according to the category score result, perform disturbance discrimination and label assignment on the input tensor, and the disturbance types include 50 Hz fundamental wave disturbance, 50 Hz harmonic disturbance, transient impulse disturbance caused by high-voltage switch operation, non-stationary low-frequency disturbance, and wideband non-Gaussian noise disturbance with statistical adaptive characteristics; S47, output the finally recognized disturbance label together with a frequency domain position index as a basis for dynamic scheduling of a subsequent disturbance suppression module and for calling an abnormal event tracking mechanism.

[0034] By constructing the disturbance recognition network A-DINet which integrates multi-channel convolution, multi-head attention mechanism, and category correlation modeling, automatic classification and spectral domain positioning of typical disturbance types in the environment of a railway transformer are realized, the 50 Hz harmonic, transient pulse, low-frequency disturbance, and wideband non-Gaussian noise can be quickly separated and labeled under strong background noise conditions, the automation degree and spectral domain matching accuracy of disturbance recognition are significantly improved, and the adaptive performance and resource calling efficiency of the subsequent disturbance suppression link are optimized.

[0035] In the embodiment, the S5 specifically includes: S51, according to the disturbance type label and the corresponding frequency band index obtained in step S4, dynamically call the parameter configuration scheme of the disturbance suppression module, and activate the signal processing path matched with the current disturbance feature; S52, input the original signal data into a 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, each intrinsic modal function component representing a signal component at different frequency scales; S53, calculate the energy offset index, the spectral density distribution feature, and the similarity score between each intrinsic modal function component and the target feature template in sequence, and select a plurality of modal components to form a target modal set according to a preset correlation criterion; S54, construct the target modal set as a time series input and send it 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 modal components through a recursive structure, and outputs a predicted frequency response sequence; S55, according to the frequency domain prediction result output by the sequence modeling network and the original modal signal, perform difference inversion, frequency filtering, and reconstruction to suppress non-target disturbance components and enhance target signals. S56, perform residual evaluation and dynamic update operation on the enhanced signal, if the difference between the current result and the perturbation template is still greater than the threshold, re-trigger the decomposition path fine-tuning process until the interference suppression result is stable; S57, output the signal sequence after interference suppression processing, retain the abnormal characteristics with physical meaning in the target frequency band, and significantly reduce the influence of multi-source interference components on the subsequent analysis model.

[0036] By constructing a two-stage interference suppression module composed of a variational modal decomposition and a gated recurrent unit network, the signal modal is decomposed in the frequency domain by VMD, and the high correlation modal is accurately selected to form a target subset, and then the frequency domain evolution trend is modeled and reconstructed by GRU, which realizes the retention of key features and the filtering of invalid disturbances in complex interference environment. Compared with the traditional filter, this scheme has the characteristics of self-adaptation, learning and adjustment, which improves the robustness and signal fidelity of the system to dynamic interference scenarios.

[0037] In this embodiment, the S6 specifically includes: S61, arrange the signal after interference suppression processing in step S5 according to the channel type and the sampling time window, and construct a standardized multi-dimensional tensor data structure based on the time axis, the frequency axis and the channel axis, to ensure that the structure can be directly analyzed by the subsequent feature extraction model; S62, perform morphological feature extraction operation on the multi-dimensional tensor data based on sparse dictionary learning, the sparse dictionary is learned from the training sample library in advance, and can express the typical waveform mode of pulse signals such as partial discharge in a limited atomic set; S63, decompose the input signal into a sparse coefficient matrix and an atomic dictionary combination by using sparse coding algorithm, and identify the starting point, duration, waveform symmetry and consistency features of pulse morphology in multiple channels during the coding process; S64, further extract frequency domain and statistical feature indexes of the signal in combination with the sparse decomposition result, the feature 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, and frequency drift amplitude depicting the dynamic shift of main frequency position; the spectral distribution skew rate of frequency energy component skew direction and amplitude; S65, splice all the extracted morphological features, frequency domain indexes and statistics in the feature dimension to form a composite feature vector in a unified format, and generate a corresponding feature table for each time window; S66, perform dimension standardization, outlier rejection and integrity check on the composite feature vector to ensure that it meets the input requirements and accuracy requirements of the subsequent state evaluation module; S67, output the processed composite feature vector set, and retain the corresponding time stamp, channel label and source information as the input data basis for railway power transformer state modeling and abnormality identification.

[0038] The sparse dictionary learning method is adopted to construct a tensor-level feature extraction model, which can model the pulse shape, spectrum shape and statistical change in the signal in high dimension, and has a significant advantage in partial discharge transient signal detection. The composite feature vector composed of spectrum entropy, kurtosis, frequency drift and other parameters effectively enhances the distinguishability and expression ability of abnormal features, and provides key input features with structural, time-varying and engineering interpretation for health state evaluation.

[0039] In the embodiment, the S7 specifically comprises: S71, arrange the composite feature vector set in time sequence, 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 taking the monitoring time point as a graph node, and the Euclidean distance, cosine similarity or spectral distance between feature vectors as a 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, wherein the graph attention network dynamically weights the edge weight in the graph to improve the expression ability of the key state transition path; S74, in the graph attention network, a multi-head graph attention mechanism is introduced to model the correlation between different time points differently, and the node representation is updated 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 the node features, to extract the 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 identification of the current state of the transformer, and the state label includes normal, mild abnormality, moderate abnormality and severe failure; S77, based on the node classification results of a plurality of consecutive time periods, construct a running state trend curve, and calculate the running trend slope, fluctuation degree and potential inflection point, to predict the state evolution direction in the future period; S78, output the health state classification label, running trend score index and state transition path of the current time point, for the remote monitoring platform to perform early warning response, operation and maintenance scheduling and equipment maintenance strategy optimization.

[0040] By constructing the composite feature vector into a time series graph structure and using a graph attention network and a multi-channel graph convolutional neural network for modeling, the evolution trend of the transformer operating state across time periods is modeled, classified and predicted. The structure has the ability to efficiently capture the spatio-temporal correlation between features, automatically identify potential fault nodes and their propagation paths, significantly improve the early perception level of the system for trend mutations, degradation evolution and abnormal aggregation behavior, and enhance the prediction accuracy and interpretability.

[0041] In the embodiment, the S8 specifically includes: S81, the health status classification label, the operation trend score index and the time series node graph embedding vector output in step S7 are integrated, and a structured diagnostic result data package is constructed together with the frequency spectrum graph, the interference identification label and the interference suppression residual value obtained in steps S3 to S6; S82, format standardization conversion is performed 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, the formatted diagnostic result data is packaged through an embedded communication module, which includes a cellular communication module, an Ethernet interface, an industrial optical fiber transmission module and various data transmission paths; S84, according to the field deployment requirements, select the communication protocol type, through the lightweight message queue transmission protocol or the power protocol based on IEC 61850 communication model for data packaging and transmission, to ensure that the system can be compatible with different substation dispatching systems; S85, upload the diagnostic data to the remote monitoring platform, which supports structured data analysis, spectrum graph visualization, state trend curve reconstruction and disturbance distribution heat map display functions, forming a comprehensive diagnostic interface for multi-dimensional monitoring objects; S86, set the alarm strategy in the remote monitoring platform, automatically trigger the multi-level early warning mechanism according to the state label severity, trend score slope and disturbance level, including in-station sound and light alarm, SMS push, control center pop-up window and cloud platform alarm aggregation distribution; S87, push the diagnostic results to the power dispatching system or the equipment operation and maintenance management system, support linkage execution state refresh, task scheduling, maintenance work order issuance and substation interworking strategy switching, build an intelligent operation and maintenance control network supporting multi-site closed-loop response; S88, all uploaded records are cached locally and backed up to the historical data management module periodically, for fault reproduction, model training optimization and long-term performance evaluation.

[0042] A remote uploading mechanism including data format standardization, communication protocol adaptation and cloud linkage is constructed, combining MQTT and IEC 61850 protocol to realize efficient interaction of diagnosis results between edge devices and cloud platforms. The system supports spectrum graph visualization, state score display and early warning control linkage, effectively breaking the data loop between on-site detection, intelligent diagnosis and dispatch execution, improving the intelligent joint control capability and dispatch collaborative response efficiency among multiple substations, and meeting the intelligent operation and maintenance requirements of high reliability and high real-time in the railway scene.

[0043] Embodiment 1 In order to verify the feasibility of the application in implementation, the application is applied to a traction substation of a certain railway main line, which deploys a set of intelligent monitoring system for the running state of power transformers. As a typical high-load and frequently started and stopped traction power supply node, there are strong power frequency disturbance, high frequency pulse interference and electromagnetic coupling interference and other complex background noise in the running environment. Traditional state monitoring equipment cannot effectively identify early abnormal discharge signals, impact instability signals and multiple superimposed interference characteristics, and there are problems of high equipment misjudgment rate, maintenance lag and unclear safety boundary.

[0044] In the embodiment of the application, the monitoring system is arranged with five types of signal acquisition channels of voltage, current, partial discharge, vibration and electromagnetic radiation at key positions of the primary side, the secondary side and the shell of the target transformer. A multi-channel synchronous sampling mechanism is adopted, and the sampling frequency can reach 10 MHz. The data of each channel is real-time spliced and cached according to the sliding window. Through unified clock synchronization control and anti-aliasing sampling strategy, the system can realize high-precision time alignment of multi-source signals without relying on third-party synchronization modules, thereby establishing a unified input basis for subsequent analysis.

[0045] The collected signals are first subjected to high-order statistical filter for offset and trend removal, and then subjected to amplitude standardization through Z-score normalization, and then input into the multi-scale residual fusion time-frequency transformation network constructed by the application. Three groups of convolution paths with different receptive fields are used inside the network to extract micro-disturbance, periodic fluctuation and trend evolution of the signals, and residual connection and channel attention mechanism are used to effectively prevent feature degradation, and the short-time Fourier window length is dynamically adjusted based on frame energy and entropy function, and finally a three-dimensional time-frequency feature tensor is generated.

[0046] The interference recognition part discriminates the interference types through the recognition network constructed by multi-channel time sequence convolution and multi-head attention mechanism. The training set contains multiple interference category samples, including power frequency harmonic, impact disturbance, low frequency background interference and non-Gaussian wideband interference. The test results show that the interference classification accuracy is stable at more than 96%, the average interference recognition delay is not more than 1 second, and it shows strong adaptability and stability under different equipment load and weather conditions.

[0047] Subsequently, the system automatically calls the interference suppression module composed of variational mode decomposition and gated recurrent unit, and performs two-stage processing on the identified interference section. The VMD part decomposes the signal into several modal components and selects high correlation components according to the energy shift criterion, and the GRU model models and predicts the reconstruction of the frequency domain evolution trend. The average signal-to-noise ratio of the signal after interference suppression is improved from 8.7 dB to 16.4 dB, effectively enhancing the identifiability of low-amplitude weak pulses in the original signal.

[0048] The system further adopts a sparse dictionary learning method to extract structural features of the signal after interference suppression, combines spectral entropy, kurtosis factor, energy fluctuation rate and other spectral and statistical features to construct a composite feature vector, and classifies and predicts the device operating state through a graph attention network and a graph convolution model. In the field evaluation, the model can realize stable division of four states of "normal", "mild abnormality", "moderate abnormality" and "severe abnormality", and the average classification accuracy reaches 94.2%, and the trend prediction has an early warning rate of more than 85% within 3 cycles.

[0049] All diagnostic results are finally encapsulated and uploaded to the remote monitoring platform through the edge computing terminal, supporting the linkage of the dispatching system based on the MQTT protocol and the IEC 61850 standard. The platform can automatically generate alarms, trend maps and recommended maintenance work orders, and realize cooperative strategy switching with adjacent sites. The operation data show that after the system is deployed, the average fault recognition time delay is reduced by about 72%, and the false alarm rate is reduced by more than 60%, which provides strong support for the continuous visualization management of operation and maintenance efficiency and transformer health status.

[0050] This embodiment verifies that the present application has full-link, high-precision and intelligent fault perception capability in a strong interference railway operating environment, which can significantly improve the practicality, precision and automation level of the transformer monitoring system, and has good engineering adaptability and promotion prospect.

[0051] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

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 pulse signals such as partial discharge in a limited atomic set; S63, the input signal is decomposed into a sparse coefficient matrix and an atomic dictionary combination by using 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 a cellular communication module, an Ethernet interface, an industrial optical fiber transmission module and various data transmission paths; 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