A transformer operation state intelligent identification method based on machine learning

Abstract

The application relates to the technical field of machine learning, and discloses a transformer operation state intelligent identification method based on machine learning, which is applied to the edge-cloud collaborative architecture of 110kV-500kV transformers and online monitoring systems thereof. The method extracts multi-modal features based on multi-channel cleaning data, introduces a mode-enhanced TCN model of fusion kernel feature mapping and a Koopman operator, realizes time-frequency coupling modeling and hidden space state representation. The model realizes online recursion and working condition adaptive tracking through a Sherman-Morrison incremental updating mechanism, outputs operation state identification results and risk scores, and realizes millisecond-level alarm linkage through an IEC 61850 GOOSE protocol, thereby providing high-precision and traceable technical support for intelligent monitoring and operation and maintenance of transformers in a transformer substation.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a machine learning-based intelligent identification method for transformer operating status. Background Technology

[0002] With the continuous expansion of the power grid and the increasing automation of substations, transformers are susceptible to the combined effects of various factors such as temperature rise, electrical stress, partial discharge, mechanical vibration, and insulation aging during long-term high-load operation. Their operating status exhibits significant time-varying and multi-physics coupling characteristics. Transformer operating data sources are complex, including multi-source heterogeneous information such as oil temperature, winding hot spot temperature, current, voltage, harmonics, dissolved gas in oil (DGA), vibration, acoustic emission, partial discharge signals, and environmental quantities. The data volume is large, and real-time requirements are high. While existing monitoring systems can achieve parameter acquisition and threshold alarms, they struggle to uncover the dynamic correlations between multimodal features at the time-series level and lack adaptive modeling and risk quantification capabilities for non-stationary processes. Especially in scenarios involving multi-sensor collaborative monitoring, data synchronization errors, noise interference, and signal loss often lead to unstable status identification results, affecting the accuracy and timeliness of operating status judgment. Summary of the Invention

[0003] This invention proposes a machine learning-based intelligent identification method for transformer operating status. Addressing the challenges of multi-physics data coupling, signal heterogeneity, and strong feature dynamics in transformers under complex operating conditions, it constructs an edge-cloud collaborative intelligent identification system. This method achieves nanosecond-level synchronization of multi-source data based on IEEE 1588 PTP high-precision time synchronization. Noise suppression and data completion are achieved through extended Kalman filtering and spline interpolation algorithms, ensuring the integrity and temporal consistency of monitoring data. Based on multi-channel cleaned data, multi-modal features such as thermal, electrical, gaseous, mechanical, discharge, and actuator features are extracted to establish a multi-physics feature set reflecting the full state of the transformer. At the model level, a modal-enhanced temporal convolutional network (TCN) model integrating kernel feature mapping and the Koopman operator is introduced. This transforms the non-stationary operating process into a linearly evolving spectral dynamic feature space, realizing time-frequency coupled modeling and latent space state representation of multi-modal features. The model achieves online recursive updates of the stable inverse metric matrix and the Koopman operator through a Sherman–Morrison incremental update mechanism, thus possessing adaptive tracking capabilities for abrupt and gradual changes in operating conditions. Finally, the system outputs operational status identification results and risk scores through a modal energy-driven feature gating mechanism, achieving unified identification and intelligent alarm of multiple physical fields (thermal, electrical, gas, and mechanical) of the transformer. This invention, while ensuring real-time performance and robustness, realizes full-dimensional dynamic perception and intelligent judgment of the transformer's operational status, providing sustainable and optimized technical support for the intelligent operation and maintenance of high-voltage substation equipment.

[0004] This invention provides a machine learning-based intelligent identification method for transformer operating status, applied to an edge-cloud collaborative architecture of transformers and their online monitoring systems. The online monitoring system includes an edge computing unit, a substation control system, and a control layer. The method includes the following steps:

[0005] Step S1: The edge computing unit continuously collects multi-source sensor data and operating status data from the transformer, and performs nanosecond-level time synchronization and alignment with IEEE1588 PTP to form multi-channel raw data with a unified timestamp.

[0006] Step S2: For each channel of the multi-channel raw data, perform outlier detection and extended Kalman filtering denoising to generate multi-channel denoised data; for data with discontinuities or packet loss in the multi-channel denoised data, according to the operating conditions of the transformer equipment, use spline interpolation and Kalman smoothing algorithm to complete the data; at the same time, perform sensor self-test and zero-point drift calibration to ensure the integrity and accuracy of the data and obtain multi-channel cleaning data;

[0007] Step S3: Based on the multi-channel cleaning data, extract multi-source physical quantities and construct multi-modal feature data;

[0008] Step S4: Based on the TCN model, train the TCN model using multimodal feature data to obtain a trained modally enhanced TCN model; output the transformer operating status identification result and risk score using the trained modally enhanced TCN model.

[0009] Step S5: Preset risk threshold. When the risk score exceeds the risk threshold, trigger the alarm linkage mechanism: publish event and alarm information to the substation control system through the IEC61850 GOOSE communication protocol to achieve millisecond-level real-time response; at the same time, write the transformer operation status identification result into the local database and the central time series database in the form of hash signature to form a traceable full life cycle record chain to support subsequent operation and maintenance analysis and responsibility tracing.

[0010] Furthermore, step S3 specifically includes the following steps:

[0011] Step S31: Obtain temperature channel and load channel data from multi-channel cleaning data, calculate top oil temperature rise, winding hot spot temperature estimation, thermal gradient and thermal inertia factor to characterize the thermal characteristic response of the transformer and obtain transformer thermal characteristic data.

[0012] Step S32: Extract electrical quantity channel data from the multi-channel cleaning data, calculate the load rate, unbalance coefficient, total harmonic distortion (THD), and interharmonic energy to characterize the electrical operating state of the transformer and obtain the transformer electrical characteristic data;

[0013] Step S33: Extract dissolved gas (DGA) channel data from the multi-channel cleaning data, calculate Rogers ratio and Doernenburg ratio, generate Duval triangle and pentagonal centroid coordinate codes, and calculate the generation rate of each characteristic gas to characterize the thermal decomposition characteristics and discharge evolution trend of the insulating oil, and obtain transformer gas characteristic data.

[0014] Step S34: Extract vibration channel and acoustic emission channel data from multi-channel cleaning data, and use envelope spectrum analysis to calculate kurtosis, spectral kurtosis and fault bandpass energy index of specific frequency bands to identify abnormal vibration characteristics of transformer mechanical components and obtain transformer mechanical characteristic data.

[0015] Step S35: Extract ultra-high frequency (UHF) and ultrasonic partial discharge channel data from the multi-channel cleaning data, construct a phase distribution discharge (PRPD) statistical map, and extract density moment characteristics, phase concentration characteristics, and discharge energy distribution characteristics to characterize the partial discharge characteristics of the insulation system and obtain transformer discharge characteristic data;

[0016] Step S36: Extract the on-load tap changer (OLTC) operating current channel and stroke displacement channel data from the multi-channel cleaning data, calculate the stroke-current fingerprint similarity and extract the contact bounce feature to reflect the operating status of the tap changer actuator and obtain the transformer actuator feature data;

[0017] Step S37: Combine transformer thermal characteristic data, transformer electrical characteristic data, transformer gas characteristic data, transformer mechanical characteristic data, transformer discharge characteristic data, and transformer actuator characteristic data to obtain multimodal characteristic data.

[0018] Furthermore, step S4 specifically includes the following steps:

[0019] Step S41: Perform time windowing processing on the multimodal feature data, and establish a block-Hankel embedding matrix within the sliding window to maintain the local temporal correlation of the time series; based on the block-Hankel embedding matrix, use the stochastic Fourier feature mapping method with explicit approximation of Gaussian kernel to perform finite-dimensional kernel space mapping to obtain kernel feature snapshots;

[0020] Step S42: Define the kernel feature space. Based on the kernel feature snapshot, initialize the stable inverse metric matrix in the kernel feature space. The stable inverse metric matrix is ​​constructed by introducing the autocorrelation matrix of the kernel feature and introducing a stable inverse term during the inversion process of the autocorrelation matrix of the kernel feature to suppress the numerical singularity in the autocorrelation matrix of the kernel feature and improve the numerical stability of matrix inversion, thus obtaining the stable inverse metric matrix. Based on the stable inverse metric matrix, construct the Koopman operator. When the time sliding window moves forward, define the input update matrix and the output update matrix, calculate the local weight matrix, and introduce the Sherman-Morrison formula to perform online recursive updates with constant complexity. Update the stable inverse metric matrix and the Koopman operator in an incremental manner to realize real-time tracking and stable modeling of non-stationary friction dynamics processes, generate the updated Koopman operator, and perform spectral decomposition on the updated Koopman operator in the dimension-reduced POD subspace to obtain the main mode feature pairs, forming time-frequency coupled dynamic mode features.

[0021] Step S43: Inject time-frequency coupled dynamic modal features into the TCN model, and use the modal energy in the time-frequency coupled dynamic modal features to generate feature gating weights to highlight the dominant operating mode and suppress the secondary modes, thereby optimizing the modal attention unit of the TCN model; based on the modal decay rate and frequency distribution information in the time-frequency coupled dynamic modal features, perform adaptive normalization and residual fusion on the output of the convolutional layer of the TCN model to ensure the stability and numerical controllability of the model training process. After multiple rounds of training and optimization, the trained modality-enhanced TCN model is obtained.

[0022] Step S44: By training the post-modal augmentation TCN model, output the transformer operating status identification results and risk score.

[0023] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0024] First, this invention introduces the IEEE 1588 PTP high-precision time synchronization mechanism and extended Kalman filter algorithm into an edge-cloud collaborative architecture, achieving synchronization and dynamic denoising of multi-source sensor data on a nanosecond-level time scale. This solves the feature mismatch problems caused by data time-series drift, noise interference, and signal packet loss in traditional monitoring systems. This mechanism ensures precise alignment of thermal, electrical, gas, and mechanical multi-physical field signals in the time dimension and consistency of correlation in the spatial dimension, achieving high-fidelity fusion of monitoring data and significantly improving the accuracy and stability of transformer operating status feature extraction.

[0025] Secondly, the proposed Temporal Modal Enhanced Convolutional Network (TCN) model, which integrates kernel feature mapping and the Koopman operator, constructs a mapping relationship from non-stationary physical processes to a dynamic spectral feature space, realizing time-frequency coupled modeling and latent space state representation of multimodal features. By introducing the Sherman–Morrison incremental update mechanism, online recursive updates of the Koopman operator and the stabilized inverse metric matrix are achieved, enabling the model to continuously learn new features and adaptively track changes in transformer operating conditions during operation. This innovation significantly improves the model's robustness and real-time response capability under scenarios of sudden faults, gradual aging, and complex load fluctuations.

[0026] Finally, this invention, through a modal energy-driven feature gating mechanism, combines the spectral features of the latent space with the physical state to establish an integrated output model for transformer operating status identification and risk scoring. This model enables accurate identification and graded alarms for conditions such as overload, abnormal cooling, partial discharge, OLTC faults, and insulation aging. This mechanism effectively enhances the intelligence and interpretability of the alarm system, achieving a fully intelligent closed loop from data acquisition and feature modeling to risk assessment. It provides technical support for real-time diagnosis, risk prevention and control, and full lifecycle operation and maintenance of substation transformers, significantly improving the safety and economy of power equipment operation. Attached Figure Description

[0027] Figure 1 A flowchart illustrating a machine learning-based intelligent identification method for transformer operating status provided by the present invention;

[0028] Figure 2 The training convergence curves in Example 3: a comparison of the original TCN and the modally enhanced TCN;

[0029] Figure 3 This is a scatter plot of modal attenuation-frequency (Koopman-POD λ spectrum) from Example 3. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] Example 1, according to Figure 1This invention provides a machine learning-based intelligent identification method for transformer operating status, which is applied to the edge-cloud collaborative architecture of oil-immersed transformers (including OLTC tap changers, bushings, and cooling systems) and their online monitoring systems in 110kV to 500kV substations. The online monitoring system includes a multi-source sensing layer, an edge computing unit, a substation control system, and a control layer.

[0032] Multi-source sensing layer: Multi-source sensors and operating status acquisition devices installed on the transformer body and auxiliary equipment, including oil temperature and winding hot spot temperature sensors, online dissolved gas (DGA) chromatography monitoring device (monitoring H2, CH4, C2H6, C2H4, C2H2, CO, CO2 and moisture content), ultrasonic partial discharge sensor, triaxial vibration and acoustic emission sensor, bushing leakage current and dielectric loss online monitoring device, primary side voltage and current transformer sampling unit, ambient temperature and humidity and wind speed sensor, oil level and oil pressure sensor, and OLTC operating stroke and motor current sampling device, used to acquire thermal, electrical, gas and mechanical multi-physical field data of transformer operation;

[0033] Edge computing unit: An industrial computer (8-core CPU, 16GB memory, with GPU / AI acceleration module) installed in the station control room, equipped with IEEE 1588 PTP, supporting IEC 61850 GOOSE communication protocol, Modbus TCP and RTU protocols, as well as Ethernet and fiber optic communication interfaces; this unit has intrinsically safe isolation and electromagnetic compatibility protection functions, meeting the protection requirements of IEC 61000-6-2 and IEC 61000-6-4, and is used to perform data synchronization, preprocessing, feature extraction, model inference and alarm issuance tasks;

[0034] Substation control system: Deployed at the station control layer, it is used to receive event and alarm information issued by the edge computing unit, complete the visualization of the operation status, operation record and protection action linkage, and realize real-time communication and security joint control between the edge side and the dispatch side;

[0035] Station control layer: Used to perform offline training of models and parameter distribution, and responsible for the unified management of local databases and central time series databases; it is also used to realize model update, historical data archiving and remote operation and maintenance management functions to support the continuous optimization and full life cycle operation and maintenance of the system.

[0036] In this invention, the method includes the following steps:

[0037] Step S1: Data Synchronization: The edge computing unit continuously collects multi-source sensor data and operating status data from the transformer, and performs nanosecond-level time synchronization and alignment via IEEE 1588 PTP to form multi-channel raw data with a unified timestamp.

[0038] Step S2: Preprocessing: For each channel of the multi-channel raw data, outlier detection and extended Kalman filtering are performed to generate multi-channel denoised data; for data with discontinuities or packet loss in the multi-channel denoised data, spline interpolation and Kalman smoothing algorithm are used to complete the data according to the operating conditions of the transformer equipment; at the same time, sensor self-test and zero-point drift calibration are performed to ensure the integrity and accuracy of the data and obtain multi-channel cleaned data;

[0039] Step S3: Feature extraction: Based on multi-channel cleaned data, extract multi-source physical quantities and construct multi-modal feature data;

[0040] Step S4: Model Training and Inference: Based on the TCN model, the TCN model is trained using multimodal feature data to obtain a trained modally enhanced TCN model; the trained modally enhanced TCN model is used to output the transformer operating status identification results and risk scores.

[0041] Step S5: Alarm Release: A preset risk threshold is set. When the risk score exceeds the risk threshold, an alarm linkage mechanism is triggered: Event and alarm information are released to the substation control system through the IEC 61850 GOOSE communication protocol to achieve millisecond-level real-time response; at the same time, the transformer operating status identification result is written into the local database and the central time-series database in the form of a hash signature to form a traceable full life cycle record chain to support subsequent operation and maintenance analysis and responsibility tracing.

[0042] In conventional technical fields, this method includes the following steps:

[0043] Step R1: The edge computing unit continuously collects multi-source sensor data and operating status data from the transformer, and performs nanosecond-level time synchronization and alignment with IEEE1588 PTP to form multi-channel raw data with a unified timestamp.

[0044] Step R2: For each channel of the multi-channel raw data, outlier detection and extended Kalman filtering are performed to generate multi-channel denoised data; for data with discontinuities or packet loss in the multi-channel denoised data, spline interpolation and Kalman smoothing algorithm are used to complete the data according to the operating conditions of the transformer equipment; at the same time, sensor self-test and zero-point drift calibration are performed to ensure the integrity and accuracy of the data and obtain multi-channel cleaned data;

[0045] Step R3: Based on the multi-channel cleaned data, normalize and time-series align different physical quantities to extract statistical features, frequency domain features, and trend features representing the thermal, electrical, gas, and mechanical characteristics of the transformer; construct multimodal feature data through feature selection and dimensionality reduction algorithms to characterize the comprehensive operating status of the transformer.

[0046] Step R4: Construct a state recognition model based on the TCN model, use multimodal feature data as input, and perform supervised training on transformer operation samples; after the model converges, perform inference on the real-time input data and output the transformer operation state recognition result and risk score.

[0047] Step R5: Preset risk threshold. When the risk score exceeds the risk threshold, trigger the alarm linkage mechanism: publish event and alarm information to the substation control system through the IEC61850 GOOSE communication protocol to achieve millisecond-level real-time response; at the same time, write the transformer operation status identification result into the local database and the central time series database in the form of hash signature to form a traceable full life cycle record chain to support subsequent operation and maintenance analysis and responsibility tracing.

[0048] Example 2, this example is based on Example 1. In this example, step S3 specifically includes the following steps:

[0049] Step S31: Obtain temperature channel and load channel data from multi-channel cleaning data, calculate top oil temperature rise, winding hot spot temperature estimation, thermal gradient and thermal inertia factor to characterize the thermal characteristic response of the transformer and obtain transformer thermal characteristic data.

[0050] Step S32: Extract electrical quantity channel data from the multi-channel cleaning data, calculate the load rate, unbalance coefficient, total harmonic distortion (THD), and interharmonic energy to characterize the electrical operating state of the transformer and obtain the transformer electrical characteristic data;

[0051] Step S33: Extract dissolved gas (DGA) channel data from the multi-channel cleaning data, calculate Rogers ratio and Doernenburg ratio, generate Duval triangle and pentagonal centroid coordinate codes, and calculate the generation rate of each characteristic gas to characterize the thermal decomposition characteristics and discharge evolution trend of the insulating oil, and obtain transformer gas characteristic data.

[0052] Step S34: Extract vibration channel and acoustic emission channel data from multi-channel cleaning data, and use envelope spectrum analysis to calculate kurtosis, spectral kurtosis and fault bandpass energy index of specific frequency bands to identify abnormal vibration characteristics of transformer mechanical components and obtain transformer mechanical characteristic data.

[0053] Step S35: Extract ultra-high frequency (UHF) and ultrasonic partial discharge channel data from the multi-channel cleaning data, construct a phase distribution discharge (PRPD) statistical map, and extract density moment characteristics, phase concentration characteristics, and discharge energy distribution characteristics to characterize the partial discharge characteristics of the insulation system and obtain transformer discharge characteristic data;

[0054] Step S36: Extract the on-load tap changer (OLTC) operating current channel and stroke displacement channel data from the multi-channel cleaning data, calculate the stroke-current fingerprint similarity and extract the contact bounce feature to reflect the operating status of the tap changer actuator and obtain the transformer actuator feature data;

[0055] Step S37: Combine transformer thermal characteristic data, transformer electrical characteristic data, transformer gas characteristic data, transformer mechanical characteristic data, transformer discharge characteristic data, and transformer actuator characteristic data to obtain multimodal characteristic data.

[0056] Example 3, according to Figure 2 , Figure 3 This embodiment is based on Embodiment 2. In this embodiment, step S4 specifically includes the following steps:

[0057] Step S41: Perform time windowing processing on the multimodal feature data, and establish a block-Hankel embedding matrix within the sliding window to maintain the local temporal correlation of the time series; based on the block-Hankel embedding matrix, use the stochastic Fourier feature mapping method with explicit approximation of Gaussian kernel to perform finite-dimensional kernel space mapping to obtain kernel feature snapshots;

[0058] Step S42: Define the kernel feature space. Based on the kernel feature snapshot, initialize the stable inverse metric matrix within the kernel feature space. The stable inverse metric matrix is ​​constructed by introducing the autocorrelation matrix of the kernel feature and introducing a stable inverse term during the inversion process of the kernel feature autocorrelation matrix to suppress numerical singularity in the kernel feature autocorrelation matrix and improve the numerical stability of matrix inversion, thus obtaining the stable inverse metric matrix. Based on the stable inverse metric matrix, construct the Koopman operator. When the time sliding window moves forward (discarding the earliest sample and introducing the latest sample), define the input update matrix and output update matrix, calculate the local weight matrix, and introduce the Sherman–Morrison formula to perform online recursive updates with constant-time complexity. Update the stable inverse metric matrix and Koopman operator incrementally to achieve real-time tracking and stable modeling of non-stationary friction dynamics processes, generating the updated Koopman operator. Perform spectral decomposition on the updated Koopman operator in the reduced-dimensional POD subspace to obtain the main mode feature pairs. ,in Characterizing the mode's decay / oscillation properties, The modal coefficient vector of the current state forms the time-frequency coupled dynamic modal features. ; For the updated Koopman operator;

[0059] Initialize the stable inverse metric matrix and the Koopman approximation operator in the kernel feature space:

[0060] , ;

[0061] in, Indicates the index of the current sliding window. , Represents block-Hankel snapshot matrix pairs. and Represents the kernel feature snapshot matrix. Represents the autocorrelation matrix in the kernel feature space; Represents the stabilization parameter. Represents the identity matrix. Represents the stabilized inverse metric matrix. This represents the Koopman approximation operator;

[0062] When the sliding window moves forward, the following definition applies:

[0063] ;

[0064] in, This represents the input update matrix. Indicates time index The corresponding sliding window contains the first (oldest) block-Hankel snapshot vector; This represents the newly introduced block-Hankel snapshot vector (physical space) during the same window forward movement, with the column index denoted as... Indicates "the next column outside the window"; Represents a random Fourier feature (RFF) mapping;

[0065] ;

[0066] in, This indicates the output update matrix. This represents the snapshot in the second column of the current sliding window (i.e., the moment after the oldest snapshot), which is related to... Form a pair of time-adjacent samples; This represents a new snapshot outside the current sliding window, corresponding to "the next moment after the newly introduced sample";

[0067] , ;

[0068] in, Represents a symbol matrix, express The inverse of the matrix, express transpose, Represents the local weight matrix;

[0069] Perform online recursion with constant time complexity based on the Sherman–Morrison formula:

[0070] , ;

[0071] in, This represents the updated stabilized inverse metric matrix. This represents the updated Koopman approximation operator;

[0072] Step S43: Inject time-frequency coupled dynamic modal features into the TCN model, and use the modal energy in the time-frequency coupled dynamic modal features to generate feature gating weights to highlight the dominant operating mode and suppress the secondary modes, thereby optimizing the modal attention unit of the TCN model; based on the modal decay rate and frequency distribution information in the time-frequency coupled dynamic modal features, perform adaptive normalization and residual fusion on the output of the convolutional layer of the TCN model to ensure the stability and numerical controllability of the model training process. After multiple rounds of training and optimization, the trained modality-enhanced TCN model is obtained.

[0073] During the training process, Figure 2 Training convergence curves: Comparison of original TCN vs. modality-enhanced TCN; Figure 2 In the diagram, the horizontal axis (Epoch) represents the number of iterations in model training; the vertical axis (loss value) reflects the magnitude of the model's error on the training and validation sets; the blue and orange curves (original TCN) show the changes in loss on the training and validation sets, exhibiting slow convergence and large fluctuations on the validation set; the green and red curves (modal augmentation TCN) show that after introducing dynamic modal features and gating mechanisms, the model reaches lower loss in earlier epochs, and the validation set curve is smoother, indicating faster convergence, better generalization performance, and a more stable model training process.

[0074] produce Figure 3 Modal attenuation-frequency scatter plot (Koopman-POD λ spectrum); Figure 3 In the diagram, the horizontal axis (frequency) represents the characteristic frequencies corresponding to different modes; the vertical axis (attenuation rate) represents the degree of energy attenuation of a mode, and the closer the value is to 1, the more stable the mode is; the dot (early training stage): the mode distribution is relatively dispersed and the attenuation rate is low, indicating that the model still responds to unstable modes (such as noise and short-term disturbances); the cross point (post-convergence stage): the modes are concentrated in the low attenuation, high energy region, indicating that the model has focused on the dominant mode and the response to high-frequency noise modes has weakened.

[0075] Step S44: After training the modal augmentation TCN model, perform temporal modeling and feature alignment to output the latent space state representation. The latent space state representation reflects the coupled dynamic characteristics of the multi-physics fields of transformer thermal-electrical-gas-mechanical-discharge, which can be used for subsequent operation status identification and risk assessment. Input the latent space state representation into the classification-regression joint head of the modal augmentation TCN model to output the transformer operation status identification results and risk score. The transformer operation status identification results include: normal, overload, abnormal cooling, thermal fault T1 / T2 / T3, electrical fault (ionization discharge / arc), partial discharge, OLTC fault, bushing insulation abnormality, and early insulation paper aging. The risk score ranges from 0 to 100 points, and the handling level and suggested strategy are given.

[0076] Example 4, this example is based on Example 3, in this example,

[0077] Step S4: Model Training and Inference: Based on the TCN model, the TCN model is trained using multimodal feature data to obtain a trained modally enhanced TCN model; the trained modally enhanced TCN model is used to output the transformer operating status identification results and risk scores.

[0078] In this embodiment:

[0079] Equipment: 220 / 110 kV, 180 MVA main transformer (oil-immersed, ONAF), on-load tap changer (OLTC).

[0080] Transformer operating status identification results and risk score:

[0081] Identification results: Cooling anomaly (fan A failure / performance degradation) accompanied by "increased risk of overheating T2" (no obvious arc discharge, partial discharge at low to medium level);

[0082] Risk score: 83 / 100;

[0083] Contribution breakdown (readable): Hot aisles +38, Cooling links +28, DGA (T2 / T3 edge) +12, Partial discharge +5, others 0;

[0084] Alarm Level: Level II (Requires immediate repair; recommended to be addressed within 2 hours).

[0085] Recommended strategy:

[0086] Immediately force the standby fan circuit to be switched on, and check the current of the fan A contactor / motor and the bearing noise;

[0087] Maintain the load at ≤0.85 pu until maintenance is complete;

[0088] Retest DGA within 48 hours and monitor the growth rates of ethylene and acetylene;

[0089] Increase the frequency of oil temperature inspections (every 15 minutes). If the hot spot is ≥95 ℃ or the score is ≥90, implement Level I upgrade and load transfer.

[0090] Step S5: Alarm Release: A preset risk threshold is set. When the risk score exceeds the risk threshold, an alarm linkage mechanism is triggered: Event and alarm information are released to the substation control system through the IEC 61850 GOOSE communication protocol to achieve millisecond-level real-time response; at the same time, the transformer operating status identification result is written into the local database and the central time-series database in the form of a hash signature to form a traceable full life cycle record chain to support subsequent operation and maintenance analysis and responsibility tracing.

[0091] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A machine learning-based intelligent identification method for transformer operating status, applied in an edge-cloud collaborative architecture of transformers and their online monitoring systems, wherein the online monitoring system includes: Edge computing unit, substation control system, and control layer; characterized in that the method includes the following steps: Step S1: The edge computing unit continuously collects multi-source sensor data and operating status data of the transformer, and performs nanosecond-level time synchronization and alignment through IEEE1588 PTP to form multi-channel raw data; Step S2: For each channel of the multi-channel raw data, perform outlier detection, extended Kalman filtering for noise reduction, and data completion to obtain multi-channel cleaned data; Step S3: Based on the multi-channel cleaning data, extract multi-source physical quantities and construct multi-modal feature data; Step S4: Based on the TCN model, train the TCN model using multimodal feature data to obtain a trained modally enhanced TCN model; output the transformer operating status identification result and risk score using the trained modally enhanced TCN model. Step S5: Preset a risk threshold. When the risk score exceeds the risk threshold, trigger an alarm linkage mechanism. Step S4 specifically includes the following steps: Step S41: Perform time windowing processing on the multimodal feature data, establish a block-Hankel embedding matrix within the sliding window, and perform finite-dimensional kernel space mapping based on the block-Hankel embedding matrix using the stochastic Fourier feature mapping method with explicit approximation of Gaussian kernel to obtain kernel feature snapshots. Step S42: Define the kernel feature space. Based on the kernel feature snapshot, initialize the stable inverse metric matrix in the kernel feature space. Construct the Koopman operator based on the stable inverse metric matrix. When the time sliding window moves forward, define the input update matrix and the output update matrix, calculate the local weight matrix, and introduce the Sherman-Morrison formula to perform online recursive updates with constant complexity. Update the stable inverse metric matrix and the Koopman operator incrementally to generate the updated Koopman operator. Perform spectral decomposition on the updated Koopman operator to obtain the main modal feature pairs, forming time-frequency coupled dynamic modal features. Step S43: Inject the time-frequency coupled dynamic modal features into the TCN model. After multiple rounds of training and optimization, the trained modality-enhanced TCN model is obtained. Step S44: By training the post-modal augmentation TCN model, output the transformer operating status identification results and risk score.

2. The intelligent identification method for transformer operating status based on machine learning according to claim 1, characterized in that: Step S3 specifically includes the following steps: Step S31: Obtain temperature channel and load channel data from the multi-channel cleaning data, calculate the top oil temperature rise, winding hot spot temperature estimation, thermal gradient and thermal inertia factor, and obtain transformer thermal characteristic data; Step S32: Extract electrical quantity channel data from the multi-channel cleaning data, calculate the load rate, unbalance coefficient, total harmonic distortion rate and interharmonic energy to obtain transformer electrical characteristic data; Step S33: Extract dissolved gas channel data from the multi-channel cleaning data, calculate Rogers ratio and Doernenburg ratio, generate Duval triangle and pentagon centroid coordinate codes, and calculate the generation rate of each characteristic gas to obtain transformer gas characteristic data. Step S34: Extract vibration channel and acoustic emission channel data from multi-channel cleaning data, and use envelope spectrum analysis to calculate kurtosis, spectral kurtosis and fault bandpass energy index to obtain transformer mechanical characteristic data; Step S35: Extract the discharge channel data of ultra-high frequency and ultrasonic parts from the multi-channel cleaning data, construct the phase distribution discharge statistics map, extract the density moment characteristics, phase concentration characteristics and discharge energy distribution characteristics, and obtain the transformer discharge characteristic data; Step S36: Extract the on-load tap changer operation current channel and travel displacement channel data from the multi-channel cleaning data, calculate the travel-current fingerprint similarity and extract the contact bounce feature to obtain the transformer execution unit feature data; Step S37: Combine transformer thermal characteristic data, transformer electrical characteristic data, transformer gas characteristic data, transformer mechanical characteristic data, transformer discharge characteristic data, and transformer actuator characteristic data to obtain multimodal characteristic data.

3. The intelligent identification method for transformer operating status based on machine learning according to claim 1, characterized in that: Step S5 specifically involves: setting a risk threshold; when the risk score exceeds the risk threshold, triggering an alarm linkage mechanism: using the IEC 61850 GOOSE communication protocol to send event and alarm information to the substation control system. The transformer operating status identification results are written into the station control layer database in the form of hash signatures to form a traceable full life cycle record chain.

4. The intelligent identification method for transformer operating status based on machine learning according to claim 1, characterized in that: The stable inverse metric matrix is ​​constructed by introducing the autocorrelation matrix of the kernel feature and introducing a stable inverse term during the inversion process of the autocorrelation matrix of the kernel feature to suppress the numerical singularity in the autocorrelation matrix of the kernel feature and improve the numerical stability of matrix inversion, thus obtaining the stable inverse metric matrix.

5. The intelligent identification method for transformer operating status based on machine learning according to claim 1, characterized in that: Step S43 specifically includes: injecting time-frequency coupled dynamic modal features into the TCN model, using modal energy in the time-frequency coupled dynamic modal features to generate feature gating weights, and optimizing the modal attention unit of the TCN model; based on the modal decay rate and frequency distribution information in the time-frequency coupled dynamic modal features, performing adaptive normalization and residual fusion on the convolutional layer output of the TCN model to ensure the stability and numerical controllability of the model training process, and after multiple rounds of training and optimization, obtaining the trained modally enhanced TCN model.

Images