Method for evaluating mechanical strength of transformer winding under reclosing working condition

Abstract

The invention relates to the technical field of mechanical performance testing of power transformer winding materials, in particular to a method for evaluating the mechanical strength of a transformer winding under a reclosing working condition, and the method comprises the steps: obtaining and preprocessing historical operation data, finite element simulation data and laboratory simulation test data; deploying a micro foil strain gauge, an MEMS acceleration sensor and a flexible piezoresistive sensor based on the preprocessed data, and synchronously collecting radial strain, axial acceleration and cushion block pressure data; after noise filtering, time alignment and feature extraction, multi-directional deformation cooperative influence is analyzed through methods of Pearson's correlation coefficients, partial least square regression and the like; and in combination with a historical damage sample training evaluation model, outputting a mechanical strength degradation level in real time. According to the method, the problems of synchronous acquisition and correlation analysis of multi-direction deformation data under reclosing impact are solved, and accurate evaluation of overall mechanical strength change caused by accumulated deformation is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical performance testing of power transformer winding materials, and in particular to a method for evaluating the mechanical strength of transformer windings under reclosing conditions. Background Art

[0002] The reclosing condition refers to the automatic reclosing of a circuit breaker in a power system after it trips due to a fault. This process generates transient overvoltages and surge currents. When acting on transformer windings, these can cause imbalances in the electromagnetic force distribution within the windings, electric field distortion in the insulation, and abnormal local temperature rises. By analyzing changes in winding inductance parameters to reflect the degree of interturn displacement, the resonant frequency formed by the winding inductance and interturn capacitance shifts due to winding deformation. High-frequency response analysis is used to collect frequency domain characteristics at different closing times. Combined with a model correlating winding deformation and frequency response characteristics established through finite element simulation, this allows for a quantitative assessment of the impact of reclosing conditions on the mechanical integrity and insulation performance of the windings. Furthermore, the time domain characteristics of the partial discharge signal correspond to the degradation state of the insulation. By capturing parameters such as the amplitude and phase distribution of the discharge pulse using ultra-high frequency sensors, it is possible to indirectly determine the evolution of damage to the winding insulation layer under reclosing shocks.

[0003] In operating scenarios where power transformers are repeatedly subjected to reclosing shock currents, multi-dimensional testing and analysis methods for the degradation of winding mechanical strength present technical pain points: existing technologies primarily test and analyze deformation of the winding in a single radial or axial direction under short-circuit shocks. However, in actual operation, reclosing shocks simultaneously induce multiple mechanical deformations of the winding, such as radial compression and axial vibration. The coupling of these deformations accelerates the degradation of the winding's mechanical properties. However, there is a lack of testing methods that can simultaneously collect and correlate multi-directional deformation data, making it difficult to accurately assess changes in the winding's overall mechanical strength due to accumulated deformation. For example, existing tests only monitor radial deformation through strain gauges and are unable to simultaneously obtain data on the degree of pad loosening caused by axial vibration. This makes it impossible to fully analyze the synergistic effects of radial compression and axial loosening on the winding's mechanical strength, making it difficult to determine whether the winding has lost its ability to operate safely due to the accumulation of multi-dimensional deformations. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method for evaluating the mechanical strength of transformer windings under reclosing conditions, which solves the problem of synchronous collection and correlation analysis of multi-directional deformation data of power transformer windings under reclosing impact, so as to accurately evaluate the overall mechanical strength changes caused by cumulative deformation.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows: The present invention provides a method for evaluating the mechanical strength of transformer windings under reclosing conditions, comprising: Step 1: Obtain basic data including historical operation data, finite element simulation data, and laboratory simulation test data. After deduplication and formatting preprocessing, generate high-stress area data and vulnerable location data for clarifying sensor deployment locations, as well as historical damage samples and simulation and test verification data for model training; Step 2: Based on the preprocessed high-stress area data and vulnerable location data, micro-foil strain gauges, MEMS accelerometers, and flexible piezoresistive sensors are deployed in the winding area. The original time series data of radial strain, axial acceleration, and pad pressure are synchronously collected through an FPGA synchronization controller. Step 3: The collected original time series data is subjected to sliding window mean filtering, 1ms time window alignment based on synchronized timestamps, and feature extraction. Step 4: Based on the extracted multi-dimensional deformation features, such as the maximum radial deformation, the main axial vibration frequency, and the pad pressure fluctuation range, the Pearson correlation coefficient is used to analyze the correlation between the radial deformation and the axial vibration amplitude. The partial least squares regression is used to analyze the correlation between the main axial vibration frequency and the pad pressure fluctuation to identify the parameters that affect the multi-directional deformation synergy. In step 5, the identified multi-directional deformation synergistic influence parameters and the actual damage samples in the historical operation data are input into the training model to train an evaluation model that outputs the mechanical strength degradation level. Based on the deformation characteristics and synergistic influence parameters collected in real time, the evaluation model outputs the real-time winding mechanical strength evaluation results.

[0006] Furthermore, the method for evaluating the mechanical strength of transformer windings under reclosing conditions of the present invention further includes: The historical operation data is the transformer reclosing event record and corresponding disassembly inspection report for more than 5 years, which is used to provide the reclosing operating conditions and actual damage sample library required for model training; The finite element simulation data simulates the electromagnetic force distribution of the winding under different reclosing conditions through a multi-physics field coupling model, outputting the radial maximum stress position, axial vibration frequency, and pad contact pressure threshold to identify high-stress areas for sensor deployment; The laboratory simulation test data simulates the reclosing condition through a 220kV transformer scale model and a controllable impulse current generator, records the winding deformation data and disassembles the transformer to verify the actual damage, which is used to verify the consistency of the simulation data with the actual operation data.

[0007] Furthermore, in the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention, the sensors include a micro foil strain gauge, a MEMS acceleration sensor, and a flexible piezoresistive sensor; The micro foil strain gauge is deployed in the area with the maximum radial stress determined by the finite element simulation data in step 1, and is used to monitor the radial tensile or compressive strain value in real time; The MEMS acceleration sensor is deployed at the wire-piece pad gap where the axial vibration frequency deviation is significant as recorded in the historical disassembly detection report in step 1, to monitor the acceleration amplitude and frequency of the axial vibration; The flexible piezoresistive sensor is deployed in the area with the lowest pad contact pressure threshold determined by the finite element simulation data in step 1, and indirectly reflects the looseness of the pad through pressure changes; All sensors are synchronized with microsecond timestamps through an FPGA synchronization controller, which allows the acquisition of radial strain, axial acceleration, and pad pressure data on the same time basis.

[0008] Furthermore, the method for evaluating the mechanical strength of transformer windings under reclosing conditions of the present invention includes step 1: Noise filtering uses sliding window mean filtering to filter out high-frequency electromagnetic interference noise and retain low-frequency deformation characteristics related to reclosing impact; Time alignment is based on the synchronization timestamp generated by the FPGA synchronization controller in step 2. The radial strain, axial acceleration, and pad pressure data are divided into time windows of 1ms and matched with the reclosing event timestamps of the historical operation data in step 1 to form a time-aligned multi-dimensional data set. Feature extraction includes: Extract the maximum deformation and deformation change rate within each time window; Perform FFT transformation on the acceleration signal to extract the main vibration frequency and vibration amplitude; Extract the fluctuation range and pressure drop rate of the pressure signal; The extracted multi-dimensional deformation features are used as input parameters for the association analysis in step 4.

[0009] Furthermore, the method for evaluating the mechanical strength of transformer windings under reclosing conditions of the present invention further includes: The Pearson correlation coefficient is used to calculate the correlation between the radial deformation change rate and the axial vibration amplitude, and the radial and axial coupling coefficients are output; Partial least squares regression is used to analyze the correlation between the pad pressure fluctuation range and the axial main vibration frequency, and the axial and pad correlation coefficient is output; By establishing a multivariate covariance matrix, the covariance of radial, axial and pad deformation characteristics is calculated to identify the dominant degradation factors; The coupling coefficient, correlation coefficient and dominant degradation factor output by the correlation analysis are used as input parameters for model training in step 5.

[0010] Furthermore, in the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention, the input parameters of the mechanical strength evaluation model include: the extracted maximum radial deformation, the axial main vibration frequency offset, and the pad pressure drop rate; The multi-directional deformation coupling strength output by the correlation analysis includes the radial and axial coupling coefficients, and the axial and pad correlation coefficients; The output parameter of the mechanical strength assessment model is the mechanical strength degradation level, where: First mechanical strength degradation level: maximum radial deformation ≤ preset threshold, axial main vibration frequency offset ≤ 5%, pad pressure drop rate ≤ 0.5 N / s, and multi-directional deformation coupling strength ≤ 0.3; Second mechanical strength degradation level: Any parameter exceeds the threshold of the first mechanical strength degradation level but does not reach the third mechanical strength degradation level standard; The third mechanical strength degradation level: the maximum radial deformation is greater than 2 times the preset threshold, or the axial main vibration frequency offset is greater than 10%, or the pad pressure drop rate is greater than 1.0N / s; The mechanical strength assessment model adopts random forest algorithm or long short-term memory network training, takes actual damage samples of historical operation data as labels, optimizes tree depth or time step parameters through cross-validation, and captures the nonlinear relationship between multi-dimensional features and degradation level.

[0011] Furthermore, the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention uses 20% of samples from laboratory simulation test data that were not used in training to verify the mechanical strength evaluation model. The degradation level output by the mechanical strength evaluation model is compared with the actual disassembly damage level. When the matching degree is ≥90%, the matching degree is the proportion of samples whose output grade of the mechanical strength assessment model is consistent with the actual grade, and the model calibration is completed; The calibrated mechanical strength assessment model is used for real-time assessment in step 5.

[0012] Furthermore, the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention includes the following steps in real time: Receive the real-time sensor data collected synchronously by the FPGA synchronization controller in step 2; After the real-time data is subjected to noise filtering, time alignment, and feature extraction in step 3, it is input into the correlation analysis in step 4; Input the real-time characteristic parameters after preprocessing and correlation analysis into the mechanical strength evaluation model trained in step 5, and output the real-time mechanical strength degradation level; Real-time mechanical strength degradation levels include: First mechanical strength degradation level: real-time data and evaluation results are recorded in the database to maintain regular inspection cycles; Second mechanical strength degradation level: generates early warning information to prompt shortening of local monitoring cycle; Third mechanical strength degradation level: Generate maintenance recommendations for push to the operation and maintenance system.

[0013] Furthermore, the method for evaluating the mechanical strength of transformer windings under reclosing conditions of the present invention further includes: preprocessing historical operating data in step 1 to form a library of reclosing conditions and actual damage samples, thereby providing real damage labels for model training in step 5; Finite element simulation data is pre-processed in step 1 to identify high stress areas and theoretical deformation thresholds, providing a theoretical basis for sensor deployment in step 2. The laboratory simulation test data is preprocessed in step 1 to verify the consistency between the simulation data and the actual operation data.

[0014] Furthermore, the method for evaluating the mechanical strength of transformer windings under reclosing conditions of the present invention further includes: The FPGA synchronization controller implements hardware clock synchronization through PTP; Combined with software timestamp correction, it eliminates the time offset caused by hardware characteristics differences between different sensors; Through the coordination of hardware synchronization and software correction, the time alignment used in step 3 accurately matches the timestamps of historical reclosing events to avoid correlation analysis deviations caused by time misalignment.

[0015] Beneficial effects of the present invention: The present invention effectively solves the technical pain points of synchronous collection and correlation analysis of multi-directional deformation data under reclosing impact through multi-source data fusion and multi-dimensional analysis framework: pre-processing and coordination of historical operation data, finite element simulation data and laboratory simulation test data provide multi-dimensional basis for sensor deployment in high-stress areas and vulnerable locations, avoiding the limitations of a single data source; synchronous deployment of micro-foil strain gauges, MEMS acceleration sensors and flexible piezoresistive sensors and alignment of FPGA timestamps realize microsecond-level synchronous collection of radial deformation, axial vibration and pad loosening data, solving the problem of time misalignment of multi-directional deformation data; application of sliding window filtering, time window alignment and feature extraction technology The key features reflecting instantaneous stress intensity, structural stiffness changes and contact stability are extracted, providing quantitative parameters for correlation analysis; the joint analysis of Pearson correlation coefficient, partial least squares regression and multivariate covariance matrix quantifies the synergistic effects of multi-directional deformations such as radial and axial coupling, axial and pad correlation, breaking through the one-sidedness of single-directional deformation analysis; the training and real-time evaluation of random forest or LSTM models, by capturing the nonlinear relationship between multi-dimensional features and degradation levels, realizes the accurate evaluation of the overall mechanical strength changes of the winding caused by cumulative deformation, providing a decision-making basis based on multi-dimensional test data for the operation and maintenance of power transformers, and meeting the technical requirements for multi-dimensional analysis and testing of materials or objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0017] Figure 1 This is a flow chart of a method for evaluating the mechanical strength of transformer windings under reclosing conditions provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0019] See also Figure 1 The present invention provides a method for evaluating the mechanical strength of transformer windings under reclosing conditions, comprising: Step 1: Obtain basic data including historical operation data, finite element simulation data, and laboratory simulation test data. After deduplication and formatting preprocessing, generate high-stress area data and vulnerable location data for clarifying sensor deployment locations, as well as historical damage samples and simulation and test verification data for model training; Step 2: Based on the preprocessed high-stress area data and vulnerable location data, micro-foil strain gauges, MEMS accelerometers, and flexible piezoresistive sensors are deployed in key areas of the winding. The original time series data of radial strain, axial acceleration, and pad pressure are synchronously collected through an FPGA synchronization controller. Step 3: The collected original time series data is subjected to sliding window mean filtering, 1ms time window alignment based on synchronized timestamps, and feature extraction. Step 4: Based on the extracted multi-dimensional deformation features, such as the maximum radial deformation, the main axial vibration frequency, and the pad pressure fluctuation range, the Pearson correlation coefficient is used to analyze the correlation between the radial deformation and the axial vibration amplitude. The partial least squares regression is used to analyze the correlation between the main axial vibration frequency and the pad pressure fluctuation to identify the parameters that affect the multi-directional deformation synergy. In step 5, the identified multi-directional deformation synergistic influence parameters and the actual damage samples in the historical operation data are input into the training model to train an evaluation model that outputs the mechanical strength degradation level. Based on the deformation characteristics and synergistic influence parameters collected in real time, the evaluation model outputs the real-time winding mechanical strength evaluation results.

[0020] The data fusion preprocessing in step 1 includes: building a multi-dimensional basic database by integrating three types of heterogeneous data sources: historical transformer operation data, finite element simulation data, and laboratory simulation test data. The historical operation data is extracted from the reclosing event sequence recorded by the power grid dispatching system for more than 5 years and the corresponding transformer disassembly inspection report, including physical damage labels such as the actual winding deformation position and the displacement of the insulating pad; the finite element simulation data calculates the electromagnetic force distribution of the winding under different reclosing phase angles based on the multi-physics field coupling model (electromagnetic, structural, and thermal), and outputs the radial stress cloud map, axial vibration modal frequency, and pad contact pressure distribution threshold; the laboratory simulation test data uses a scaled model of a 220kV transformer, reproduces the reclosing transient process through a controllable impulse current generator, and simultaneously records the winding deformation sensor data and calibrates the actual mechanical damage degree after disassembly. The three types of data are subjected to the ETL process for redundant record deduplication, unified timestamp conversion, and dimension standardization, and finally generate: High stress area coordinate set: It is composed of the radial stress peak coordinates and axial vibration energy concentration area coordinates output by finite element simulation.

[0021] Mapping table of vulnerable locations: established based on the frequently appearing wire cake numbers and pad displacement positions in historical dismantling reports.

[0022] Damage sample feature library: historical damage samples are associated with corresponding operating parameters (inrush current amplitude, reclosing interval duration) and stored.

[0023] This step provides a theoretical basis for optimized sensor deployment through multi-source data spatial registration and feature alignment, and constructs a dataset with physical labels for model training.

[0024] Step 2 of multi-sensor collaborative deployment and synchronous acquisition includes: Based on the high-stress area coordinate set and the damage-prone location mapping table output in step 1, multiple types of sensors are embedded and deployed on the transformer winding entity: Micro foil strain gauges are fixed with epoxy resin to the area with the maximum radial stress determined by finite element simulation (such as the outer surface of the coil in the middle of the low-voltage winding) to monitor the transient response of the radial tensile / compressive strain of the winding.

[0025] MEMS accelerometer: Installed in the gap between the wire cakes where the axial vibration frequency deviation exceeds 5% as recorded in the historical disassembly report, the axial vibration time domain waveform is captured through triaxial acceleration measurement.

[0026] Flexible piezoresistive sensor: Embedded in the area where the pad contact pressure shown by finite element simulation is 20% lower than the design threshold, it measures the dynamic pressure changes of the insulating pad.

[0027] All sensor signals are connected via shielded cables to an FPGA synchronization controller. This controller uses the IEEE 1588 PTP protocol for hardware clock synchronization and writes microsecond-precision timestamps when the data acquisition card is triggered. The FPGA's parallel sampling architecture synchronously captures the original time series of radial strain, axial acceleration, and pad pressure signals, eliminating the data phase difference caused by traditional time-sharing sampling.

[0028] The signal processing and feature engineering in step 3 include: A three-stage processing pipeline is performed on the raw time series acquired by the FPGA: Noise filtering: A sliding mean filter with adjustable window width is used to filter out high-frequency noise (>2kHz) caused by magnetostriction of the transformer core, while retaining the mechanical deformation characteristics in the 0 and 500Hz frequency bands.

[0029] Time alignment: Based on the synchronized timestamps written into the FPGA, the three sensor data are re-sliced ​​into 1ms time windows and matched with the time base of historical reclosing events to form a multi-channel dataset with strict time alignment.

[0030] Feature extraction: Radial strain channel: Calculate the maximum absolute value of deformation in each time window and the deformation rate between adjacent windows.

[0031] Axial acceleration channel: Extract the 0 and 200 Hz main vibration frequency components and their amplitude envelopes through FFT transformation.

[0032] Pad pressure channel: Statistical pressure fluctuation range (maximum and minimum values) and pressure drop slope of linear regression fitting.

[0033] The multi-dimensional feature vector output by this step serves as the input for subsequent correlation analysis. The feature design focuses on characterizing the instantaneous strength, structural stiffness change rate, and contact stability of the winding mechanical response.

[0034] The multi-dimensional deformation coupling analysis in step 4 includes: Based on the eigenvectors extracted in step 3, a coupling analysis model between deformation directions is established: Radial and axial correlation analysis: The Pearson correlation coefficient is used to calculate the linear correlation between the radial deformation rate and the axial vibration amplitude, and the radial and axial coupling coefficients are output (with a 1 to 1 scale) to quantify the energy transfer effect of the radial compression of the winding on the axial vibration.

[0035] Axial and pad correlation modeling: Partial least squares regression (PLSR) is used to analyze the mapping relationship between the axial main vibration frequency and the pad pressure fluctuation range, and the axial and pad correlation coefficient is output to characterize the degree of pad loosening caused by axial vibration.

[0036] Multivariate covariance analysis: A three-dimensional covariance matrix of radial deformation, axial main frequency offset, and pad pressure drop rate is constructed. The primary factor that dominates mechanical degradation (such as the dimension corresponding to the maximum eigenvalue) is identified through eigenvalue decomposition. The coupling coefficient, correlation coefficient, and dominant factor output by this step together constitute the multi-directional deformation synergistic influence parameter set, revealing the interaction mechanism between different deformation modes.

[0037] The mechanical strength degradation assessment in step 5 includes: Establishing a mechanical strength assessment model based on machine learning: Input layer: Receives the basic features of radial maximum deformation, axial main frequency offset, and pad pressure drop rate extracted in step 3, as well as the coupling coefficient and correlation coefficient collaborative parameters output in step 4.

[0038] Model architecture: Use the random forest algorithm to build a multi-decision tree ensemble model, or apply the long short-term memory network (LSTM) to process time series features; use the actual damage level in the historical damage sample library as the supervision label.

[0039] Training mechanism: Optimize model hyperparameters (such as tree depth of random forest and time step of LSTM) through k-fold cross-validation to learn the nonlinear mapping between multidimensional features and mechanical strength degradation.

[0040] Real-time evaluation: Receive pre-processed sensor data streams online, extract features and perform correlation analysis before inputting them into the trained model to output mechanical strength degradation grades I and III: Level I: All parameters are below safety thresholds and routine monitoring is maintained.

[0041] Level II: A single parameter exceeds the limit but does not reach the critical value, and high-frequency local monitoring is initiated.

[0042] Level III: Key parameters exceed the failure threshold, generating maintenance instructions. This step achieves a quantitative assessment of the winding's cumulative mechanical damage by integrating multi-dimensional deformation characteristics and their synergistic effect parameters.

[0043] Specifically, in the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention, the multi-type sensors include micro foil strain gauges, MEMS acceleration sensors, and flexible piezoresistive sensors; The micro foil strain gauge is deployed in the area with the maximum radial stress determined by the finite element simulation data in step 1, and is used to monitor the radial tensile or compressive strain value in real time; The MEMS acceleration sensor is deployed at the wire-piece pad gap where the axial vibration frequency deviation is significant as recorded in the historical disassembly detection report in step 1, to monitor the acceleration amplitude and frequency of the axial vibration; The flexible piezoresistive sensor is deployed in the area with the lowest pad contact pressure threshold determined by the finite element simulation data in step 1, and indirectly reflects the looseness of the pad through pressure changes; All sensors are synchronized with microsecond timestamps through an FPGA synchronization controller, which allows the acquisition of radial strain, axial acceleration, and pad pressure data on the same time basis.

[0044] Sensor selection and deployment logic includes: The deployment of micro-foil strain gauges is based on the radial stress distribution cloud map generated by the finite element simulation in step 1. Areas of the coil where stress concentration exceeds 80% of the material's yield strength are selected as installation locations. Specifically, the strain gauge base is bonded to the outer surface of the winding conductor using a high-temperature resistant epoxy resin, with the grid wires aligned parallel to the radial stress direction. This captures the microstrain signals of radial elastic deformation of the winding caused by reclosing shock in real time. This deployment method directly correlates with the mechanical weaknesses revealed by the finite element simulation, ensuring that the measured data reflects the actual strain state in high-risk areas predicted by theoretical analysis.

[0045] The placement of the MEMS accelerometers is based on abnormal axial vibration records in historical disassembly inspection reports. To address inter-coil gasket gaps where frequency deviations exceeded ±5% of the baseline value after multiple reclosing events, triaxial accelerometers were embedded and installed at the corresponding coordinate points of the three-phase windings. The sensors are mechanically coupled to the gasket surface via a rigid bracket, with the measurement direction perpendicular to the gasket plane. They are used to collect the time-domain acceleration waveform and spectral characteristics of the winding's axial vibration. This deployment strategy maps historical damage locations to real-time monitoring points, establishing a direct correlation between axial vibration data and known mechanical failure modes.

[0046] The integration location of the flexible piezoresistive sensor is determined by the critical contact pressure zone of the pad, as determined by finite element simulation. Thin-film piezoresistive sensing units are embedded on the pad surface where simulation results indicate pressure values ​​below 15% of the design compression force threshold. The upper and lower surfaces of the sensor are in direct contact with the insulating pad and the end face of the coil, and the dynamic pressure distribution of the pad is inverted by the rate of change of resistance. When the reclosing shock causes the pad to loosen, the pressure drop is converted into a continuous electrical output signal, indirectly indicating the degree of stability degradation of the insulation support structure.

[0047] The multi-source data synchronous collection mechanism includes: The output signals of the three sensor types are connected to the multi-channel acquisition card of the FPGA synchronization controller via shielded twisted-pair cables. The controller has a built-in IEEE 1588 Precision Time Protocol (PTP) module, which receives the master clock signal via the optical fiber network to achieve hardware clock synchronization for each acquisition channel. Each time a reclosing event is triggered, the controller injects a microsecond-accurate timestamp into all data channels. A parallel sampling architecture is used for synchronous capture: Analog voltage signal of radial strain channel (corresponding to 0 and 5000με range); Three-axis acceleration digital signal (±50g range); Piezoresistive sensor bridge output signal (0 and 10MPa range); Raw data is transferred to the buffer in a fixed frame format, with each frame including the sensor ID, synchronization timestamp, and sample value. This mechanism eliminates the phase errors caused by traditional time-sharing sampling, aligning the transient responses of different physical quantities with the same time base, and establishing a consistent timing foundation for the multidimensional feature correlation analysis in step 3.

[0048] Specifically, the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention uses sliding window mean filtering to filter out high-frequency electromagnetic interference noise and retain low-frequency deformation characteristics related to reclosing impact; The time alignment is based on the synchronization timestamp generated by the FPGA synchronization controller in step 2, dividing the radial strain, axial acceleration and pad pressure data into time windows of 1ms, and matching them with the reclosing event timestamps of the historical operation data in step 1 to form a time-aligned multi-dimensional data set; The feature extraction includes: Extract the maximum deformation and deformation change rate within each time window; Perform FFT transformation on the acceleration signal to extract the main vibration frequency and vibration amplitude; Extract the fluctuation range and pressure drop rate of the pressure signal; The extracted multi-dimensional deformation features are used as input parameters for the association analysis in step 4.

[0049] The noise filtering processing mechanism includes: A sliding window mean filtering algorithm is used to process the raw sensor signals, with the window width set based on the spectral characteristics of the transformer core's magnetostrictive noise. The filtering process is performed separately for the radial strain, axial acceleration, and pad pressure signals: consecutive sampling points are grouped according to a fixed window length, and the arithmetic mean of each group of data points is calculated as the output value. This operation effectively suppresses high-frequency noise components (>2kHz) caused by core vibration and electromagnetic interference, while retaining low-frequency deformation characteristics (0 to 500Hz) associated with the mechanical impact of the reclosing circuit breaker. The filtered signal retains effective mechanical response information, such as radial elastic deformation of the winding, axial structural vibration, and pad contact pressure changes, providing a denoising data foundation for subsequent feature extraction.

[0050] Multi-channel time alignment methods include: Based on the microsecond-level precision timestamp generated by the FPGA synchronization controller in step 2, the time base of the three sensor data streams is unified. The specific implementation includes: The filtered radial strain, axial acceleration and pad pressure data are divided into discrete time windows at 1 ms intervals.

[0051] The data in each window is marked with a starting timestamp and matched with the time coordinates of the reclosing event in the historical operation database in step 1.

[0052] Phase compensation is performed using a linear interpolation algorithm for data points across windows.

[0053] This creates a strictly time-aligned multidimensional dataset, establishing a correspondence between the transient responses of different physical quantities at the same time coordinate. This alignment mechanism resolves data misalignment caused by sensor response delays and ensures temporal consistency for collaborative analysis of multi-directional deformations.

[0054] The multi-dimensional feature extraction process includes: Perform feature quantization within the time-aligned data window: Radial strain channel: The maximum absolute value of the strain within the calculation window is used as the instantaneous deformation strength index. The backward difference method is used to calculate the strain change rate between adjacent sampling points to characterize the dynamic change characteristics of the radial stiffness of the winding.

[0055] Axial acceleration channel: Perform a fast Fourier transform (FFT) on the acceleration time domain signal to identify the main vibration frequency component between 0 and 200 Hz, whose energy accounts for more than 60% of the total energy, and simultaneously extract the peak value of the vibration amplitude envelope corresponding to this frequency.

[0056] Pad pressure channel: Calculate the fluctuation range of the pressure sampling values ​​within the statistical window (the difference between the maximum and minimum values), and use the least squares method to fit the slope of the pressure-time curve as the pressure drop rate.

[0057] The extracted characteristic parameters form a multidimensional vector representing the mechanical response, including radial deformation intensity, axial vibration spectrum characteristics, and pad contact stability indicators. This vector serves as the input for the correlation analysis in step 4, transforming the raw signal into quantitative parameters of mechanical behavior.

[0058] Specifically, the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention uses the radial deformation change rate, axial vibration amplitude, pad pressure fluctuation range, and axial main vibration frequency extracted from the multi-dimensional deformation characteristics as input, and includes the following sub-analyses: The Pearson correlation coefficient is used to calculate the correlation between the radial deformation change rate and the axial vibration amplitude, and the radial and axial coupling coefficients are output; Partial least squares regression is used to analyze the correlation between the pad pressure fluctuation range and the axial main vibration frequency, and the axial and pad correlation coefficient is output; By establishing a multivariate covariance matrix, the covariance of radial, axial and pad deformation characteristics is calculated to identify the dominant degradation factors; The coupling coefficient, correlation coefficient and dominant degradation factor output by the correlation analysis are used as input parameters for model training in step 5.

[0059] Quantification of radial and axial coupling effects includes: Based on the radial deformation rate of change and axial vibration amplitude sequence extracted in step 3, the Pearson correlation coefficient algorithm is used to calculate the linear correlation between the two. The specific implementation process is as follows: within the time window of a single reclosing event, the radial deformation rate of change is used as the independent variable X, and the axial vibration amplitude is used as the dependent variable Y, and the ratio of the covariance to the standard deviation is calculated. The radial-axial coupling coefficient is output (with a value range of 1 to 1). A positive value indicates that radial compression exacerbates axial vibration, while a negative value indicates that radial tension suppresses axial vibration. This coefficient quantifies the modulation effect of radial deformation on the axial dynamic behavior of the winding and reveals the law of mechanical energy transfer in the orthogonal directions.

[0060] The axial and pad association modeling methods include: A multivariate mapping model was established using partial least squares regression (PLSR) for the two characteristic parameters of axial main vibration frequency and pad pressure fluctuation range. The axial main vibration frequency was used as the predictor variable matrix, and the pad pressure fluctuation range as the response variable matrix. The regression equation was established by iteratively extracting latent variables. The axial-pad correlation coefficient (standardized regression coefficient) was output, representing the change in pressure fluctuation amplitude caused by a unit frequency change. This coefficient reflects the efficiency of axial vibration energy transfer to the pad contact interface and is used to assess the stability attenuation trend of the insulation support system.

[0061] Identification of multi-dimensional degradation leading factors includes: A three-dimensional covariance matrix is ​​constructed, including the maximum radial deformation, the axial main frequency offset, and the pad pressure drop rate. The calculation process is as follows: first, the three eigenvectors are normalized by Z and score, and then the covariance values ​​between any two features are calculated to fill the matrix elements. The eigenvalues ​​and eigenvectors of the covariance matrix are solved by the Jacobi iteration method, and the eigenvector component corresponding to the maximum eigenvalue is used as the weight of the dominant degradation factor. For example, if the weight of the radial deformation component exceeds 0.7, it is determined that the radial deformation is the primary cause of mechanical strength degradation. This analysis reveals the dominant failure mode in the synergistic effect of multi-directional deformation.

[0062] Specifically, the method for evaluating the mechanical strength of transformer windings under reclosing conditions of the present invention comprises the following input parameters of the mechanical strength evaluation model: the extracted maximum radial deformation, the axial main vibration frequency offset, and the pad pressure drop rate; The multi-directional deformation coupling strength output by the correlation analysis includes the radial and axial coupling coefficients, and the axial and pad correlation coefficients; The output parameter of the mechanical strength assessment model is the mechanical strength degradation level, where: First mechanical strength degradation level: maximum radial deformation ≤ preset threshold, axial main vibration frequency offset ≤ 5%, pad pressure drop rate ≤ 0.5 N / s, and multi-directional deformation coupling strength ≤ 0.3; Second mechanical strength degradation level: Any parameter exceeds the threshold of the first mechanical strength degradation level but does not reach the third mechanical strength degradation level standard; The third mechanical strength degradation level: the maximum radial deformation is greater than 2 times the preset threshold, or the axial main vibration frequency offset is greater than 10%, or the pad pressure drop rate is greater than 1.0N / s; The mechanical strength assessment model adopts random forest algorithm or long short-term memory network training, takes actual damage samples of historical operation data as labels, optimizes tree depth or time step parameters through cross-validation, and captures the nonlinear relationship between multi-dimensional features and degradation level.

[0063] Model input parameter construction includes: The input layer of the mechanical strength assessment model receives two sets of parameters: the first is the basic deformation features output by the feature extraction module in step 3, including the absolute value of the maximum deformation in the radial strain channel, the percentage offset of the main vibration frequency in the axial acceleration channel relative to the baseline value, and the rate of pressure drop per unit time in the pad pressure channel. The second is the synergistic influence parameters generated by the correlation analysis in step 4, including the Pearson coupling coefficient between the rate of change of the radial deformation and the axial vibration amplitude, and the partial least squares correlation coefficient between the axial main vibration frequency and the pad pressure fluctuation range. After normalization, the input parameters form a five-dimensional feature vector, which comprehensively characterizes the transient mechanical response of the winding under reclosing shock and the strength of multi-directional deformation interaction.

[0064] The degradation levels are based on: The model output is divided into discrete mechanical strength degradation levels: The first level corresponds to the state of maintaining the integrity of the winding structure, requiring that the radial deformation does not exceed the elastic limit threshold of the material, the axial vibration frequency offset is within the design tolerance band (≤5%), the pad pressure attenuation rate is maintained in the linear viscoelastic deformation stage (≤0.5N / s), and the multi-directional coupling effect does not cause abnormal energy accumulation (coupling strength ≤0.3).

[0065] The second level indicates local mechanical property degradation, which is triggered when any basic parameter exceeds the first level threshold but does not reach the critical failure value, for example, the radial deformation enters the plastic deformation range but does not exceed twice the elastic limit.

[0066] The third level is determined as a structural failure risk state and is activated when the radial deformation exceeds a threshold of twice the material yield strength, or the axial vibration frequency shift causes mechanical resonance (>10%), or the pad pressure decay rate reflects the instability of the supporting structure (>1.0N / s).

[0067] The grading logic is established based on the threshold of mechanical properties of materials and the theory of dynamic stability of mechanical systems to achieve quantitative grading of damage degree.

[0068] The machine learning model training mechanism includes: Build evaluation models using Random Forest or Long Short-Term Memory network architectures: Random Forest Implementation: Using actual damage samples from the historical operation database as supervisory labels (Levels I and III), multiple decision trees are constructed for ensemble learning. Grid search and k-fold cross-validation are used to optimize hyperparameters such as tree depth and the minimum number of samples per node for splitting, enabling the model to learn the nonlinear decision boundary between multi-dimensional features and degradation levels.

[0069] LSTM implementation: Time window sequence features are fed into a recurrent neural network, using hidden state memory units to capture the temporal dependencies of deformation parameters. Time step adjustment and gradient clipping are used to optimize the training process and extract the dynamic evolution of mechanical performance degradation.

[0070] The trained model can output the corresponding mechanical strength degradation level probability distribution based on the real-time input five-dimensional feature vector, and take the maximum probability level as the evaluation result.

[0071] Specifically, the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention uses 20% of samples from laboratory simulation test data that were not used in training to verify the mechanical strength evaluation model. The degradation level output by the mechanical strength evaluation model is compared with the actual disassembly damage level. When the matching degree is ≥90%, the matching degree is the proportion of samples whose output grade of the mechanical strength assessment model is consistent with the actual grade, and the model calibration is completed; The calibrated mechanical strength assessment model is used for real-time assessment in step 5.

[0072] The verification sample construction mechanism includes: The model validation phase uses an independent data set generated by laboratory simulation tests, which is separated from the original test data in the preprocessing step 1. The specific selection principle is as follows: in the reclosing simulation test conducted on a scaled model of a 220kV transformer, 20% of the test samples are randomly selected as the validation set. The deformation characteristic parameters and synergistic influence parameters corresponding to these samples are not involved in the model training process. Each validation sample is associated with the actual mechanical damage level recorded in the winding disassembly inspection report after the test, including the I and III level damage labels converted from physical measurement values ​​such as radial permanent deformation and pad displacement distance. The validation set construction process ensures that the data distribution is homologous to the training set but independent to avoid model evaluation bias.

[0073] The level matching calculation method includes: Input the validation set samples into the trained mechanical strength assessment model to obtain the degradation level prediction results output by the model. The comparative analysis process is as follows: Compare the model output level with the actual disintegration damage level sample by sample; The proportion of samples with completely consistent statistical levels to the total validation samples; When the ratio reaches the preset threshold (≥90%), the model is judged to meet the accuracy requirements.

[0074] Only samples with a complete grade match are considered in the match calculation, excluding partial matches. For example, if the model outputs a grade II value and the actual value is grade II, it is considered a match. However, if the model outputs a grade II value and the actual value is grade I or grade III, it is considered a mismatch. This strict standard ensures the reliability of the evaluation results.

[0075] Model calibration technology implementation includes: When the matching degree does not reach the threshold, the model calibration procedure is started: Analyze the characteristic distribution of mismatched samples and identify high-frequency misjudgment parameter combinations; Adjust the feature weight distribution of random forest or the time window length of LSTM; Add enhanced samples of similar working conditions to the training set; After retraining, use the same validation set for testing; The calibration cycle continues until the matching degree is achieved. The calibrated model is integrated into the online monitoring system, which receives the sensor data stream collected in real time in step 2. After feature extraction in step 3 and correlation analysis in step 4, the real-time assessment of the mechanical strength degradation level in step 5 is performed.

[0076] Specifically, the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention, wherein the real-time evaluation includes: Receive the real-time sensor data collected synchronously by the FPGA synchronization controller in step 2; After the real-time data is subjected to noise filtering, time alignment and feature extraction in step 3, it is input into the correlation analysis in step 4; Input the real-time characteristic parameters after preprocessing and correlation analysis into the mechanical strength evaluation model trained in step 5, and output the real-time mechanical strength degradation level; Depending on the output level, do the following: First mechanical strength degradation level: real-time data and evaluation results are recorded in the database to maintain regular inspection cycles; Second mechanical strength degradation level: generates early warning information to prompt shortening of local monitoring cycle; Third mechanical strength degradation level: Generate maintenance recommendations for push to the operation and maintenance system.

[0077] Real-time data stream processing mechanisms include: The online monitoring system continuously receives the raw sensor data stream transmitted by the FPGA synchronization controller in step 2. This data stream includes a synchronous sampling sequence of three channels: radial strain, axial acceleration, and pad pressure. Each frame of data carries a timestamp with microsecond precision. The data processing engine calls the signal processing pipeline defined in step 3: first, a sliding window mean filter is applied to eliminate electromagnetic interference noise, then the three signals are sliced ​​and aligned at 1ms intervals based on the timestamp, and finally a feature extraction algorithm is executed to generate feature vectors such as the maximum radial deformation, the axial main vibration frequency offset, and the pad pressure drop rate. The processed feature set is input into the correlation analysis module in step 4 to calculate the coupling coefficient and the correlation coefficient, forming a complete real-time evaluation input parameter set.

[0078] Model inference and grade output include: The feature parameter set generated in real time is fed into the mechanical strength assessment model trained and calibrated in step 5. The model performs inference operations based on a random forest or long short-term memory network architecture. Random forests use a multi-decision tree voting mechanism to determine the probability distribution of degradation levels, while LSTMs capture feature evolution trends through time series memory units. Discrete assessment results are output based on preset threshold logic: Level 1: The model determines that all input parameters are within the safety threshold range; Level 2: One or more parameters exceed the safety threshold but do not reach the critical failure value; Level 3: Key parameters exceed the critical point of material or structural failure; The output results are accompanied by confidence indicators and trigger parameter identifiers to form a structured evaluation report.

[0079] Execution of a graded response strategy includes: Activate differentiated operation and maintenance response protocols based on output levels: First-level response: The raw sensor data, characteristic parameters, and evaluation results are compressed and stored in a time series database, maintaining the preset regular inspection cycle (such as full data collection once every 24 hours), and the system maintains the baseline monitoring status.

[0080] Level 2 response: Automatically generates a standardized early warning message containing the abnormal parameter type (e.g., "axial frequency offset 7.5%"), location information (associated with sensor installation coordinates), and a monitoring strategy upgrade instruction. This instruction triggers the sampling rate of sensors in a specific area to increase to 1kHz and initiates high-frequency data recording mode.

[0081] Level 3 response: Calling the O&M management system interface to push a maintenance work order. The work order's structured fields include the damage location topology code (e.g., "3rd coil of high-voltage winding phase A"), failure mode classification (radial deformation / axial looseness / pad instability), and an index code to the recommended treatment library (e.g., "DZ and 203" correspond to the pad re-tightening process). The work order also activates the equipment's lockout protection logic chain.

[0082] Specifically, in the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention, the historical operating data is preprocessed in step 1 to form a library of reclosing conditions and actual damage samples, providing real damage labels for model training in step 5; The finite element simulation data is pre-processed in step 1 to identify high stress areas and theoretical deformation thresholds, providing a theoretical basis for sensor deployment in step 2; The laboratory simulation test data is preprocessed in step 1 to verify the consistency between the simulation data and the actual operation data.

[0083] The preprocessing and function implementation of historical operation data include: Historical operating data is derived from reclosing event archives and corresponding transformer disassembly inspection reports recorded by the power grid dispatching system for more than five years. Key information extraction is performed during the preprocessing phase: operating parameters such as the impulse current amplitude and reclosing phase angle are extracted from the event archives; physical damage indicators such as the radial permanent deformation of the winding and the displacement distance of the pad are identified from the disassembly report. By establishing a mapping relationship table between "operating parameters and damage indicators," a sample library with real damage labels is formed. This sample library provides the training labels required for supervised learning for the mechanical strength assessment model in step 5, enabling the model to learn the evolution of mechanical damage in actual operating scenarios. The sample library is stored in a relational database and supports sample retrieval and invocation by voltage level and winding type.

[0084] The processing and application paths of finite element simulation data include: Finite element simulation data is generated through multi-physics coupling modeling: a three-dimensional solid model of the transformer, including the core, windings, and insulating spacers, is established, and electromagnetic, structural, and thermal multi-field coupling boundary conditions are set. A cloud map of the winding electromagnetic force distribution is calculated under typical reclosing conditions (e.g., a 90° closing phase angle), outputting the coordinates of the radial stress peak, the axial vibration modal frequency spectrum, and the spacer contact pressure distribution threshold. Preprocessing extracts the coordinates of regions where stress exceeds 85% of the material's allowable stress, sensitive sections of vibration frequency deviation, and spacer locations where pressure falls below 20% of the design value. These parameters provide a spatial positioning basis for sensor deployment in step 2, guiding the optimal placement strategy for micro-foil strain gauges and flexible piezoresistive sensors.

[0085] The verification logic of the laboratory simulation test includes: The laboratory simulation test used a scaled model of a 220kV transformer and replicated the reclosing transient process using a controllable impulse current generator. During the preprocessing phase, data consistency verification was performed: time-frequency domain correlation analysis was performed between the deformation data collected during the test (radial strain waveform and axial vibration spectrum) and the finite element simulation results, and the error rate of key indicators (e.g., radial stress peak error ≤ 8%) was calculated. Furthermore, the spatial distribution of physical damage measured after the test (e.g., coil deformation location) was verified to match the damage patterns in the historical operation sample library (position overlap ≥ 90%). This verification process confirmed the physical consistency of the simulation data with the actual operation data, ensuring the engineering applicability of the sensor deployment strategy in step 2 and the correlation analysis model in step 4.

[0086] Specifically, in the method for evaluating the mechanical strength of transformer windings under reclosing conditions described in the present invention, the FPGA synchronization controller implements hardware clock synchronization through PTP; Combined with software timestamp correction, it eliminates the time offset caused by hardware characteristics differences between different sensors; Through the coordination of hardware synchronization and software correction, the time alignment used in step 3 accurately matches the timestamps of historical reclosing events to avoid correlation analysis deviations caused by time misalignment.

[0087] The hardware clock synchronization mechanism includes: The FPGA synchronization controller integrates an IEEE 1588 Precision Time Protocol (PTP) module, receiving time synchronization signals from a master clock source via a fiber optic network. Hardware clock synchronization logic is deployed within the controller: During each PTP synchronization cycle, the controller calculates the path delay and clock offset between the slave and master clocks and dynamically adjusts the local clock counter. This process achieves a unified time base with microsecond-level accuracy at the physical layer, providing a consistent hardware time reference source for all connected sensor data channels. The synchronization signal is transmitted to each acquisition card via a dedicated clock distribution circuit, eliminating the time jitter generated by traditional network communication protocols.

[0088] Software timestamp correction methods include: To address the inherent delay differences in different sensor signal processing links, software correction logic is implemented in the FPGA firmware layer. The specific implementation includes: During the system initialization phase, a standard step test signal is injected into each sensor channel; Measure the actual response delay time from signal input to FPGA acquisition buffer; Dynamically compensate the measured delay value in the timestamp marking link; For example, micro-foil strain gauge signals experience a fixed delay due to the bridge balancing circuit. The correction logic subtracts this delay from their timestamps. This compensation mechanism eliminates timing deviations caused by sensor hardware characteristics, ensuring precise temporal alignment of transient responses of different physical quantities.

[0089] The specific implementation process of the present invention is based on the demand for mechanical strength testing of power transformers under reclosing conditions, especially for the technical bottleneck of synchronous collection of multi-directional deformation data and quantitative analysis of synergistic influence. During implementation, a multi-source data foundation is first constructed: reclosing event records of more than five years and corresponding transformer disassembly detection reports are extracted from the power grid dispatching system to form a historical operation database; the winding stress distribution under different reclosing phase angles is simulated through the electromagnetic, structural and thermal multi-physics field coupling model to generate radial stress peak coordinates, axial vibration sensitive frequency bands and pad pressure thresholds; a 220kV transformer scale model is used in conjunction with a controllable impulse current generator to carry out laboratory simulation tests, record deformation data and verify the actual damage location through disassembly. After the three types of data are pre-processed by deduplication, timestamp unification and dimension standardization, a high-stress area coordinate set (to guide sensor layout), a mapping table of easily damaged locations (associated with historical fault points) and a damage sample feature library (including working condition parameters and physical damage labels) are output.

[0090] During the sensor deployment phase, precise placement was performed based on preprocessing results. Micro-foil strain gauges were bonded to the coil surfaces where radial stress, as determined by finite element simulation, exceeded 85% of the material's allowable stress to monitor transient radial strain. MEMS accelerometers were installed in the gaps between pads where axial vibration frequency deviations of 5% or more were reported in historical disassembly reports to capture triaxial acceleration signals. Flexible piezoresistive sensors were embedded in areas where simulations indicated pad contact pressures were 20% below the design value to measure pressure fluctuations. All sensor signals were connected to an FPGA-based synchronization controller, which achieved hardware clock synchronization using the IEEE 1588 PTP protocol. Compensation for sensor response delays (e.g., 0.2ms for strain gauge bridge delay) was injected, and the output was a synchronized data stream with microsecond-corrected timestamps.

[0091] The signal processing process consists of three steps: A sliding mean filter with a 20ms window width is applied to the raw data to suppress high-frequency noise caused by core magnetostriction; the three signals are aligned in 1ms window slices based on the FPGA timestamp and linearly interpolated to match the historical event time base; within the aligned window, the radial maximum deformation, deformation change rate, axial main vibration frequency (FFT analysis between 0 and 200Hz), vibration amplitude, pad pressure fluctuation range, and pressure drop rate (least squares fitting slope) are extracted. The feature extraction results are input into the correlation analysis module: the Pearson correlation coefficient is used to calculate the correlation between the radial deformation change rate and the axial vibration amplitude, and the radial-axial coupling coefficient is output; partial least squares regression is used to analyze the correlation between the axial main vibration frequency and the pad pressure fluctuation range, and the axial-pad correlation coefficient is output; a three-dimensional covariance matrix of radial deformation, axial main frequency offset, and pad pressure drop rate is constructed, and the dominant degradation factor is identified through eigenvalue decomposition (e.g., the dimension weight corresponding to the largest eigenvalue is greater than 0.7).

[0092] The mechanical strength assessment model is trained using a random forest architecture: the input layer receives the five aforementioned basic features and two collaborative parameters. Using Level I and III labels from a historical damage sample library as supervisory data, a grid search is used to optimize the tree depth (typically 10 and 15) and the minimum number of split samples. Verification is performed using 20% ​​of the samples not used in training from laboratory simulations. Calibration is completed when the model output level matches the actual disassembly damage level by ≥90%. During online evaluation, real-time feature parameters are input into the calibration model output degradation level: Level I (all parameters within safety thresholds) triggers regular data recording; Level II (single parameter exceeds the limit) generates an alert with location information and increases the local sampling rate to 1kHz; Level III (radial deformation > 2 times the yield threshold / axial frequency offset > 10% / pad pressure drop > 1.0N / s) triggers a maintenance work order to be sent to the operation and maintenance system. The work order includes a topological code for the damage location (e.g., "low-voltage winding L3 phase, wire 5") and a remediation process index.

[0093] This implementation process uses multi-source data to collaboratively guide optimized sensor deployment, hardware synchronization and software calibration to ensure data timing accuracy, multi-dimensional feature correlation analysis to reveal deformation coupling mechanisms, and machine learning models to quantitatively assess cumulative damage. This fully addresses the lack of assessment of the collaborative impact of multi-directional deformations, as highlighted in previous research. Technical parameters and defined features form a supporting framework for each step. For example, the sensor deployment basis is based on the peak stress coordinates output by finite element simulations, Pearson correlation coefficient calculations enable correlation analysis, and the random forest model architecture and three-level degradation level outputs reflect the core assessment logic.

[0094] The present invention systematically solves the problem of synchronous collection and correlation analysis of multi-directional deformation data under the impact of reclosing through multi-source data fusion and multi-dimensional analysis framework. First, three types of basic data are obtained and pre-processed through step 1: historical operation data provides a sample library of "reclosing conditions and actual damage" (including event records and disassembly reports for more than 5 years), finite element simulation data clarifies the high-stress areas of the winding (such as the middle section of the low-voltage winding with the largest radial stress) and the theoretical deformation threshold, and laboratory simulation test data verifies the consistency between simulation and reality (such as the disassembly of the scaled model to verify the deformation position matching rate). After deduplication and unified format, the three types of data respectively form high-stress area data (simulation results) to guide sensor deployment, vulnerable location data (historical records) and real damage labels (historical samples) for model training, providing multi-dimensional support for subsequent data collection and analysis.

[0095] Secondly, steps 2 and 3 enable the simultaneous collection and feature extraction of multi-directional deformation data. Based on the high-stress area data preprocessed in step 1, micro-foil strain gauges (monitoring radial tensile / compressive strain), MEMS accelerometers (monitoring axial vibration acceleration amplitude and frequency), and flexible piezoresistive sensors (monitoring pad pressure changes) are deployed in key areas of the winding. Microsecond-level timestamp alignment (hardware clock synchronization + software correction) is achieved through an FPGA synchronization controller, enabling the acquisition of radial, axial, and pad data on the same time basis. The collected raw data undergoes sliding window mean filtering (to remove high-frequency interference), 1ms time window alignment (to match historical event timestamps), and feature extraction (such as maximum radial deformation, axial main vibration frequency offset, and pad pressure drop rate) in step 3. This generates quantitative parameters reflecting multi-directional deformation characteristics, providing input for correlation analysis.

[0096] Finally, steps 4 and 5 enable correlation analysis of multi-directional deformation and overall mechanical strength assessment. Step 4 uses the extracted multi-dimensional deformation features as input. The Pearson correlation coefficient is used to calculate the correlation between radial and axial deformation (output coupling coefficient). Partial least squares regression is used to analyze the correlation between axial vibration and pad loosening (output correlation coefficient). The multivariate covariance matrix is ​​used to quantify synergistic effects (identify the dominant degradation factors), thereby converting independent deformation data into synergistic effect parameters. Step 5 uses these parameters and historical damage samples as a training set to train a mechanical strength assessment model using a random forest or LSTM algorithm. The model outputs degradation levels (Level I: Normal, Level II: Mild Degradation, Level III: Severe Degradation). Once embedded in the online monitoring system, the model receives sensor data in real time and, after preprocessing and correlation analysis, inputs it into the model. The model then outputs the real-time winding mechanical strength status, accurately assessing changes in overall mechanical strength due to accumulated deformation.

Claims

1. A method for evaluating the mechanical strength of transformer windings under reclosing conditions, characterized in that: include: Step 1: Obtain basic data including historical operation data, finite element simulation data, and laboratory simulation test data. After deduplication and formatting preprocessing, generate high-stress area data and vulnerable location data for clarifying sensor deployment locations, as well as historical damage samples and simulation and test verification data for model training; Step 2: Based on the preprocessed high-stress area data and vulnerable location data, micro-foil strain gauges, MEMS accelerometers, and flexible piezoresistive sensors are deployed in the winding area. The original time series data of radial strain, axial acceleration, and pad pressure are synchronously collected through an FPGA synchronization controller. Step 3: The collected original time series data is subjected to sliding window mean filtering, 1ms time window alignment based on synchronized timestamps, and feature extraction. Step 4: Based on the extracted multi-dimensional deformation features, such as the maximum radial deformation, the main axial vibration frequency, and the pad pressure fluctuation range, the Pearson correlation coefficient is used to analyze the correlation between the radial deformation and the axial vibration amplitude. The partial least squares regression is used to analyze the correlation between the main axial vibration frequency and the pad pressure fluctuation to identify the parameters that affect the multi-directional deformation synergy. In step 5, the identified multi-directional deformation synergistic influence parameters and the actual damage samples in the historical operation data are input into the training model to train an evaluation model that outputs the mechanical strength degradation level. Based on the deformation characteristics and synergistic influence parameters collected in real time, the evaluation model outputs the real-time winding mechanical strength evaluation results.

2. The method for evaluating the mechanical strength of transformer windings under reclosing conditions according to claim 1, characterized in that: Also includes: The historical operation data is the transformer reclosing event record and corresponding disassembly inspection report for more than 5 years, which is used to provide the reclosing operating conditions and actual damage sample library required for model training; The finite element simulation data simulates the electromagnetic force distribution of the winding under different reclosing conditions through a multi-physics field coupling model, outputting the radial maximum stress position, axial vibration frequency, and pad contact pressure threshold to identify high-stress areas for sensor deployment; The laboratory simulation test data simulates the reclosing condition through a 220kV transformer scale model and a controllable impulse current generator, records the winding deformation data and disassembles the transformer to verify the actual damage, which is used to verify the consistency of the simulation data with the actual operation data.

3. The method for evaluating the mechanical strength of transformer windings under reclosing conditions according to claim 2, wherein: Sensors include micro-foil strain gauges, MEMS accelerometers, and flexible piezoresistive sensors; The micro foil strain gauge is deployed in the area with the maximum radial stress determined by the finite element simulation data in step 1, and is used to monitor the radial tensile or compressive strain value in real time; The MEMS acceleration sensor is deployed at the wire-piece pad gap where the axial vibration frequency deviation is significant as recorded in the historical disassembly detection report in step 1, to monitor the acceleration amplitude and frequency of the axial vibration; The flexible piezoresistive sensor is deployed in the area with the lowest pad contact pressure threshold determined by the finite element simulation data in step 1, and indirectly reflects the looseness of the pad through pressure changes; The sensors achieve microsecond-level timestamp alignment through an FPGA synchronization controller, which is used to collect radial strain, axial acceleration and pad pressure data under the same time reference.

4. The method for evaluating the mechanical strength of transformer windings under reclosing conditions according to claim 3, wherein: Step 1 includes: Noise filtering uses sliding window mean filtering to filter out high-frequency electromagnetic interference noise and retain low-frequency deformation characteristics related to reclosing impact; Time alignment is based on the synchronization timestamp generated by the FPGA synchronization controller in step 2. The radial strain, axial acceleration, and pad pressure data are divided into time windows of 1ms and matched with the reclosing event timestamps of the historical operation data in step 1 to form a time-aligned multi-dimensional data set. Feature extraction includes: Extract the maximum deformation and deformation change rate within each time window; Perform FFT transformation on the acceleration signal to extract the main vibration frequency and vibration amplitude; Extract the fluctuation range and pressure drop rate of the pressure signal; The extracted multi-dimensional deformation features are used as input parameters for the association analysis in step 4.

5. The method for evaluating the mechanical strength of transformer windings under reclosing conditions according to claim 4, characterized in that: Also includes: The Pearson correlation coefficient is used to calculate the correlation between the radial deformation change rate and the axial vibration amplitude, and the radial and axial coupling coefficients are output; Partial least squares regression is used to analyze the correlation between the pad pressure fluctuation range and the axial main vibration frequency, and the axial and pad correlation coefficient is output; By establishing a multivariate covariance matrix, the covariance of radial, axial and pad deformation characteristics is calculated to identify the dominant degradation factors; The coupling coefficient, correlation coefficient and dominant degradation factor output by the correlation analysis are used as input parameters for model training in step 5.

6. The method for evaluating the mechanical strength of transformer windings under reclosing conditions according to claim 5, characterized in that: The input parameters of the mechanical strength evaluation model include: the extracted maximum radial deformation, the axial main vibration frequency offset, and the pad pressure drop rate; The multi-directional deformation coupling strength output by the correlation analysis includes the radial and axial coupling coefficients, and the axial and pad correlation coefficients; The output parameter of the mechanical strength assessment model is the mechanical strength degradation level, where: First mechanical strength degradation level: maximum radial deformation ≤ preset threshold, axial main vibration frequency offset ≤ 5%, pad pressure drop rate ≤ 0.5 N / s, and multi-directional deformation coupling strength ≤ 0.3; Second mechanical strength degradation level: Any parameter exceeds the threshold of the first mechanical strength degradation level but does not reach the third mechanical strength degradation level standard; The third mechanical strength degradation level: the maximum radial deformation is greater than 2 times the preset threshold, or the axial main vibration frequency offset is greater than 10%, or the pad pressure drop rate is greater than 1.0N / s; The mechanical strength assessment model adopts random forest algorithm or long short-term memory network training, takes actual damage samples of historical operation data as labels, optimizes tree depth or time step parameters through cross-validation, and captures the nonlinear relationship between multi-dimensional features and degradation level.

7. The method for evaluating the mechanical strength of transformer windings under reclosing conditions according to claim 6, wherein: The mechanical strength assessment model was validated using 20% ​​of the samples from the laboratory simulation test data that were not used in the training. The degradation level output by the mechanical strength assessment model was compared with the actual disassembly damage level. When the matching degree is ≥90%, the matching degree is the proportion of samples whose output grade of the mechanical strength assessment model is consistent with the actual grade, and the model calibration is completed; The calibrated mechanical strength assessment model is used for real-time assessment in step 5.

8. The method for evaluating the mechanical strength of transformer windings under reclosing conditions according to claim 7, characterized in that: Real-time assessments include: Receive the real-time sensor data collected synchronously by the FPGA synchronization controller in step 2; After the real-time data is subjected to noise filtering, time alignment, and feature extraction in step 3, it is input into the correlation analysis in step 4; Input the real-time characteristic parameters after preprocessing and correlation analysis into the mechanical strength evaluation model trained in step 5, and output the real-time mechanical strength degradation level; Real-time mechanical strength degradation levels include: First mechanical strength degradation level: real-time data and evaluation results are recorded in the database to maintain regular inspection cycles; Second mechanical strength degradation level: generates early warning information to prompt shortening of local monitoring cycle; Third mechanical strength degradation level: Generate maintenance recommendations for push to the operation and maintenance system.

9. The method for evaluating the mechanical strength of transformer windings under reclosing conditions according to claim 8, characterized in that: Also includes: History The operating data is pre-processed in step 1 to form a sample library of reclosing conditions and actual damage, providing real damage labels for model training in step 5; Finite element simulation data is pre-processed in step 1 to identify high stress areas and theoretical deformation thresholds, providing a theoretical basis for sensor deployment in step 2. The laboratory simulation test data is preprocessed in step 1 to verify the consistency between the simulation data and the actual operation data.

10. The method for evaluating the mechanical strength of transformer windings under reclosing conditions according to claim 9, characterized in that: Also includes: The FPGA synchronization controller implements hardware clock synchronization through PTP; Combined with software timestamp correction, it eliminates the time offset caused by hardware characteristics differences between different sensors; Through the coordination of hardware synchronization and software correction, the time alignment used in step 3 accurately matches the timestamps of historical reclosing events to avoid correlation analysis deviations caused by time misalignment.

Images