Real-time monitoring method and system for DC bias magnetic field of transformer

Abstract

The present invention relates to the field of bias magnetic monitoring technology, and in particular to a real-time monitoring method and system for transformer DC bias magnetic monitoring. The method comprises the following steps: obtaining a transformer primary-side current signal and an iron core vibration signal; analyzing the zero-flux closed-loop characteristics of the transformer primary-side current signal and performing temperature drift elimination on the transformer primary-side current signal to generate a transformer primary-side optimization signal; performing magnetostrictive vibration noise separation on the iron core vibration signal to generate an iron core vibration separation signal; dynamically filtering the power frequency fundamental wave of the transformer primary-side optimization signal to extract a pure DC component; performing wavelet packet decomposition on the iron core vibration separation signal, and extracting the magnetostrictive characteristic spectrum of the decomposed iron core vibration separation signal to obtain the iron core magnetostrictive characteristic spectrum. The present invention improves the accuracy, real-time performance, and hierarchical protection response capability of transformer DC bias magnetic monitoring through multi-source signal fusion and material property modeling.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetic bias monitoring, and in particular to a real-time monitoring method and system for direct current magnetic bias of a transformer. Background Art

[0002] The DC bias problem of transformers has long existed in power systems. It is caused by the asymmetry of the core magnetic flux caused by the superposition of DC current in the system. It can easily lead to local saturation of the core, increased noise, intensified vibration, and abnormal temperature rise. In severe cases, it will cause damage to the transformer. Early bias monitoring methods were mostly based on excitation current analysis and waveform distortion judgment, but they were greatly affected by load changes and operating conditions, and lacked real-time performance and accuracy. Subsequently, indirect monitoring methods were developed that utilized multi-dimensional characteristic signals such as magnetic flux density, core temperature, and core vibration. Analysis and judgment were performed by setting thresholds or using techniques such as Fourier analysis and envelope demodulation. However, with the increasing number of power electronic devices connected, the system bias phenomenon has become more complex, which has placed higher demands on traditional algorithms. In recent years, advances in artificial intelligence, signal fusion, and sensor technology have prompted researchers to explore real-time monitoring methods based on new acquisition methods such as broadband Hall current sensors, vibration sensors, and fiber Bragg gratings. However, most traditional methods currently ignore the mechanical vibration signals caused by magnetostriction in the iron core, resulting in a slow or incomplete response to bias magnetization. At the same time, there is a lack of quantitative modeling of the "electromechanical" coupling characteristics, and it is impossible to accurately quantify the nonlinear relationship between the DC component and the iron core response. Summary of the Invention

[0003] Based on this, it is necessary to provide a real-time monitoring method and system for DC bias of a transformer to solve at least one of the above technical problems.

[0004] To achieve the above object, a real-time monitoring method for DC bias of a transformer is provided, the method comprising the following steps:

[0005] Step S1: obtaining a transformer primary-side current signal and a core vibration signal; analyzing the zero-flux closed-loop characteristic of the transformer primary-side current signal and performing temperature drift elimination on the transformer primary-side current signal to generate a transformer primary-side optimization signal; performing magnetostrictive vibration noise separation on the core vibration signal to generate a core vibration separation signal;

[0006] Step S2: Dynamically filter out the power frequency fundamental wave of the transformer primary side optimization signal to extract the pure DC component; perform wavelet packet decomposition on the core vibration separation signal, and extract the magnetostrictive characteristic spectrum of the decomposed core vibration separation signal to obtain the core magnetostrictive characteristic spectrum;

[0007] Step S3: constructing a DC component-vibration spectrum correlation model; using the pure DC component as the model input and the core magnetostrictive characteristic spectrum as the model output, updating the model coefficients of the DC component-vibration spectrum correlation model to obtain the bias magnetic sensitivity index;

[0008] Step S4: Obtain core material data; construct a material saturation curve based on the core material data, and calculate the real-time DC bias magnetic dynamic warning threshold through the bias magnetic sensitivity index and the material saturation curve; compare the DC bias magnetic dynamic warning threshold with the preset multi-level warning threshold curve to trigger the graded protection action and obtain the DC bias magnetic graded protection control data.

[0009] This method effectively extracts the magnetostrictive characteristics caused by DC bias by synergistically processing the transformer's primary-side current signal and core vibration signal, avoiding misjudgments caused by single-signal analysis and improving the accuracy and robustness of DC bias state identification. The current signal is analyzed for its zero-flux closed-loop characteristics and temperature drift interference is eliminated. The vibration signal is decomposed using wavelet packets and extracted using the magnetostrictive characteristic spectrum, achieving multi-level noise reduction and feature enhancement, significantly enhancing the system's signal recognition capabilities under complex operating conditions. By incorporating a bias sensitivity index and combining it with the core material's saturation characteristics to dynamically generate warning thresholds, this method is more targeted and timely than traditional static threshold setting methods, providing real-time reflection of system health. By comparing the dynamic warning thresholds with a multi-level warning threshold curve, a step-by-step response control mechanism is implemented, from mild to severe bias, effectively reducing the risks of core loss, noise, vibration, and lifespan reduction caused by bias accumulation. By constructing a saturation curve using core material data, this method possesses excellent material adaptability and is applicable to power transformers of different types and grades. The system has a clear structure and is easily integrated into existing transformer monitoring and protection systems, facilitating its application in intelligent power grid equipment operation and maintenance and fault warning systems. Therefore, through multi-source signal fusion and material property modeling, the present invention improves the accuracy, real-time performance, and tiered protection response capabilities of transformer DC bias monitoring.

[0010] Preferably, step S1 includes the following steps:

[0011] Step S11: collecting the transformer primary side current signal through a wide-band Hall effect sensor; collecting the core vibration signal through a high-frequency piezoelectric accelerometer;

[0012] Step S12: extracting the frequency band characteristics of the transformer primary side current signal to obtain the power frequency main component and the high-frequency harmonic component, and generating initial frequency band analysis data; identifying the zero flux reference segment of the initial frequency band analysis data, and generating zero flux reference segment data;

[0013] Step S13: performing temperature drift estimation on the zero flux reference segment data to generate temperature drift estimation data; performing thermal drift back calculation correction on the transformer primary side current signal data using the temperature drift estimation data to generate transformer primary side optimized signal data;

[0014] Step S14: performing spectrum structure analysis on the core vibration signal data to generate vibration spectrum structure data; extracting the magnetostrictive characteristic frequency band of the vibration spectrum structure data, and combining the vibration spectrum structure data and the magnetostrictive characteristic frequency band data for mutually exclusive decoupling to generate magnetostrictive vibration principal component data;

[0015] Step S15: Perform time domain signal restoration on the magnetostrictive vibration main component data to generate core vibration separation signal data.

[0016] The present invention utilizes a wideband Hall-effect sensor and a high-frequency piezoelectric accelerometer to collect current and vibration signals, respectively, covering a wide frequency range from DC to high frequencies. This ensures complete capture of bias magnetization and the magnetostrictive effect it induces, improving sampling accuracy and response speed. By extracting the frequency band characteristics of the primary-side current signal, it clearly distinguishes between the main power frequency component and high-frequency harmonic components, providing a clear data foundation for subsequent signal correction and bias magnetization identification, and enhancing the structured capabilities of signal processing. A zero-flux reference segment is automatically identified within the frequency band analysis data, which can be used to construct an environmentally adaptive reference benchmark without the need for external standards or manual settings, improving the system's automation and environmental adaptability. Temperature drift modeling and correction based on the zero-flux reference segment effectively restores the true bias magnetization characteristics, prevents temperature fluctuations from interfering with current signal identification, and improves the accuracy of the optimized current signal. Vibration spectrum structure analysis and magnetostrictive feature frequency band extraction, supplemented by a mutually exclusive decoupling method to isolate and eliminate non-magnetostrictive noise, accurately extract the main magnetostrictive component, and independently identify bias-induced vibration. Time-domain reconstruction of the extracted magnetically induced principal component signal maintains temporal consistency with the current signal, laying the foundation for subsequent construction of the bias magnetic model and feature fusion analysis, improving modeling consistency and time-series coordination. From frequency band division and temperature drift correction to vibration decoupling and time-domain restoration, the process is clearly structured and physically meaningful, facilitating its application across various types of transformers and other electromagnetic equipment.

[0017] Preferably, identifying the zero flux reference segment of the initial frequency band analysis data includes:

[0018] Perform a frequency domain scan on the frequency band analysis data. When the frequency component is stably distributed between 0 Hz and 10 kHz and the amplitude fluctuation range is less than ±3 μV, the frequency band is identified as the initial frequency band data.

[0019] In the initial frequency band data, detect the data segment containing the lowest magnetic flux disturbance interval. If all the following conditions are met, mark it as a candidate zero magnetic flux reference segment: the magnetic flux change rate is less than 1nWb / s, the maximum value of the signal derivative does not exceed 0.5μV / ms, and the continuous stabilization time is not less than 60 seconds;

[0020] A multi-point statistical stationarity analysis is performed on the candidate zero-flux reference segment to calculate its mean μ and standard deviation σ. When μ∈(–2μV, +2μV), σ≤0.8μV, and no abnormal frequency transition exceeding 50Hz occurs in the segment, the segment is confirmed to be the zero-flux reference segment.

[0021] This invention utilizes a multi-level judgment mechanism (frequency domain scanning → minimum disturbance detection → multi-point statistical analysis) to accurately extract true reference segments free of bias magnetic interference from large-scale frequency band data, ensuring a reliable and stable benchmark for subsequent temperature drift compensation. Strict physical constraints are imposed on candidate segments (such as flux change rate <1 nWb / s, signal derivative <0.5 μV / ms, and no sudden frequency jump >50 Hz), effectively eliminating segments affected by operating disturbances, electromagnetic interference, or short-term abnormal fluctuations, thereby improving overall signal robustness. Statistical stationarity analysis, combining the signal mean and standard deviation range, determines the zero-flux state. This reflects the current response characteristics under actual flux-stable conditions and provides high-reliability samples for temperature drift model construction. The recognition frequency range covers 0 Hz–10 kHz, adapting to both traditional power systems and new power electronics systems rich in high-frequency harmonics. It exhibits excellent adaptability and versatility for wide-band signals. Using a fully programmable logic recognition method, the system autonomously identifies and updates zero-flux reference segments within operational data, enabling online processing and real-time compensation without manual intervention, reducing system operation and maintenance costs. The extracted zero-flux reference segment data is stable, smooth, and free of mutations, and can be used as high-quality input for thermal drift backcalculation correction, significantly improving the accuracy and consistency of the optimized current signal.

[0022] Preferably, in step S2, dynamically filtering out the power frequency fundamental wave of the transformer primary side optimization signal to extract the pure DC component includes:

[0023] Perform data window framing on the transformer primary side optimization signal data to generate framed signal data;

[0024] Performing short-time Fourier transform on the framed signal data to generate frame-level spectrum distribution data;

[0025] Extract the main peak frequency of the power frequency from the frame-level spectrum distribution data;

[0026] Build a real-time sine reference template based on the power frequency peak frequency data to generate power frequency template signal data;

[0027] Perform signal synchronization cancellation on the transformer primary side optimized signal data and the power frequency template signal data to generate signal data after power frequency removal;

[0028] The signal data after power frequency removal is low-pass filtered to generate pure DC component data on the primary side of the transformer.

[0029] This method extracts the main power frequency peak from the frame-level spectrum in real time to construct a real-time sinusoidal reference template, dynamically tracking and filtering out minor power frequency variations and avoiding filtering errors caused by power frequency drift. Using synchronous signal cancellation, the power frequency template signal is strictly aligned with the optimized current signal and then subtracted, effectively suppressing the power frequency fundamental and avoiding signal distortion caused by the fixed bandwidth design of traditional filters. Data windowing and short-time Fourier transforms enable local time-frequency analysis of the signal, improving the timeliness and accuracy of power frequency component detection and adapting to signal characteristic variations under non-steady-state conditions. After filtering out the power frequency component, low-pass filtering further smoothes the signal, removes high-frequency noise, and accurately isolates the pure DC component, providing high-quality input for subsequent DC bias magnetic feature extraction. The dynamic power frequency template construction combined with synchronous cancellation processing results in an efficient and stable process, suitable for real-time online monitoring systems and enhancing the bias magnetic detection system's adaptability to complex power grid environments. This method does not rely on fixed power frequency parameters or preset filter parameters, offering greater versatility and adaptability, making it suitable for varying power grid frequency fluctuations and a variety of operating conditions.

[0030] Preferably, in step S2, performing wavelet packet decomposition on the core vibration separation signal and extracting magnetostrictive characteristic spectrum of the decomposed core vibration separation signal includes:

[0031] Perform wavelet packet decomposition on the core vibration separation signal to generate multi-layer wavelet packet coefficient data;

[0032] Performing frequency band reconstruction on the multi-layer wavelet packet coefficient data to generate core vibration frequency band reconstruction signal data;

[0033] Perform energy normalization on the core vibration frequency band reconstructed signal data to generate frequency band energy distribution data;

[0034] identifying a frequency band having a magnetostrictive response characteristic based on the frequency band energy distribution data, and generating magnetostrictive frequency band candidate data;

[0035] The spectrum line refinement and amplitude envelope analysis are performed on the candidate data of the magnetostrictive frequency band to extract the main frequency and subharmonic frequency characteristics and generate the magnetostrictive characteristic spectrum data of the iron core.

[0036] This invention utilizes wavelet packet decomposition technology to perform multi-layer frequency band segmentation on vibration signals, achieving high-resolution time-frequency analysis, effectively separating signal components from different frequency bands, and improving the ability to capture complex vibration signal structures. Frequency band reconstruction reconstructs signals in specific frequency bands, highlighting magnetostriction-related frequency components and suppressing unrelated noise interference, thereby enhancing the significance and discernibility of the magnetostrictive response signal. Energy normalization is performed on the reconstructed signal to eliminate the impact of energy scale differences, resulting in a more balanced energy distribution across different frequency bands and facilitating the accurate identification of candidate magnetostrictive frequency bands. Based on the normalized energy distribution, frequency bands with magnetostrictive response characteristics are intelligently identified, reducing reliance on manually set thresholds and improving the automation and accuracy of frequency band screening. Spectral line refinement and envelope extraction of candidate frequency band data accurately capture the main frequency and its subharmonic characteristics, accurately reflecting the dynamic characteristics and amplitude variations of the magnetostrictive effect in the core. The extracted magnetostrictive characteristic spectrum data reflects subtle magnetostrictive variations within the core, providing highly sensitive feature input for bias magnetic monitoring and improving the timeliness and accuracy of fault warnings. Wavelet packet decomposition has good non-stationary signal processing capabilities, can adapt to the variable vibration signal characteristics during transformer operation, and enhance the robustness of the system under complex working conditions.

[0037] Preferably, constructing the DC component-vibration spectrum correlation model in step S3 includes:

[0038] Perform time domain window segmentation on the extracted DC component data of the primary side of the transformer to generate windowed DC feature sequence data;

[0039] Perform frequency band energy reorganization on the magnetostrictive characteristic spectrum data of the core to generate the main component characteristic data of the spectrum;

[0040] Calculate the spectral domain local abnormality index of the spectral principal component characteristic data;

[0041] The spectral domain local anomaly index is used to perform magnetic vibration pseudo-coupling stripping on the main component characteristic data of the spectrum, identify the pseudo-coupling band formed by the interference of power frequency harmonics and mechanical noise, and perform directional interference band shearing to generate the net spectral characteristic data after stripping;

[0042] Perform time index alignment on the windowed DC feature sequence data and the stripped net spectrum feature data to generate joint feature time series matrix data;

[0043] Perform cointegration analysis and partial least squares regression calculation on the joint characteristic time series matrix data, extract multidimensional correlation factors, and generate correlation coupling parameter set data;

[0044] A DC component-magnetically induced vibration spectrum correlation model is constructed based on the associated coupling parameter set data.

[0045] The present invention realizes the effective fusion of the characteristics of two different physical quantities, current and vibration, through time-domain windowing processing of the DC component and reorganization of the vibration spectrum energy, thereby enhancing the comprehensive perception capability of the bias magnetic state in multiple dimensions and multiple angles. The spectral domain local abnormality index is used to identify and remove the pseudo-coupling band formed by power frequency harmonics and mechanical noise interference, avoiding the misjudgment caused by spectrum aliasing in traditional methods, and significantly improving the purity and effectiveness of the model input data. Strict time index alignment is performed on the characteristic data from different sources to ensure that the multi-source characteristics of the model input are synchronous and comparable, providing a reliable basis for subsequent correlation analysis. Cointegration analysis eliminates the false correlation caused by non-stationarity, and partial least squares regression extracts multidimensional correlation factors, which not only ensures the robustness of the statistical model, but also improves the ability to analyze complex coupling relationships. The coupling characteristics of the DC component and the magneto-induced vibration spectrum are comprehensively characterized by multidimensional correlation factors, enriching the feature space of bias magnetic diagnosis and improving the discrimination ability and generalization performance of the model. The correlation model, constructed based on measured data and statistical analysis results, not only reflects the inherent physical connection between electromagnetic DC bias and core magneto-induced vibration, but also facilitates the design of subsequent dynamic early warning and protection strategies. Through sophisticated feature extraction and complex coupling modeling, the ability to identify early-stage bias anomalies is enhanced, effectively preventing missed detections and false alarms, and improving the reliability of transformer health monitoring.

[0046] Preferably, in step S3, the pure DC component is used as the model input, and the core magnetostrictive characteristic spectrum is used as the model output to update the model coefficients of the DC component-vibration spectrum correlation model, including:

[0047] The pure DC component is divided into time windows, representative DC time series segments are extracted, and DC input sample data is generated;

[0048] Extract the main frequency component and normalize the characteristics of the magnetostrictive characteristic spectrum of the iron core to generate normalized vibration output data;

[0049] Perform sample alignment and time-series synchronization on the DC input sample data and the normalized vibration output data to generate model-associated sample pairs;

[0050] Iteratively updating the least squares residual based on the model-related sample pairs adjusts the parameters of the DC component-magnetic vibration spectrum correlation model to generate updated model coefficient data;

[0051] The sensitivity change rate of the DC bias magnetization to the vibration spectrum is calculated based on the updated model coefficient data, and the bias magnetization sensitivity index is output.

[0052] The present invention iteratively updates model coefficients based on the most recently collected time window data, enabling real-time model response and adaptive adjustment to changes in the transformer's operating state, enhancing the model's timeliness and accuracy. Sample alignment and timing synchronization ensure temporal consistency between the DC input and vibration output, avoiding model errors caused by data asynchrony and improving the accuracy and stability of parameter estimation. A least-squares residual iterative update method effectively suppresses the influence of outliers and noise, enabling smooth and convergent model parameter adjustment and enhancing model robustness. The vibration spectrum is extracted and normalized to minimize the impact of data redundancy and scale differences on model training, improving the model's sensitivity to key magnetostrictive characteristics. By calculating the updated model coefficients, a sensitivity index reflecting the degree of influence of bias magnetization on vibration response is directly obtained. This index provides a quantitative basis for subsequent warning threshold setting and fault determination, enhancing the scientific nature and operability of the warning. The dynamic updating of model coefficients and the real-time calculation of the sensitivity index support intelligent monitoring of transformer health, enabling early detection and tiered protection of bias magnetization faults, and improving equipment operational safety and reliability. This update method takes into account both the data timing characteristics and the physical meaning of the spectrum, can maintain high performance under complex electromagnetic and mechanical vibration backgrounds, and is suitable for a variety of transformer types and working conditions.

[0053] Preferably, step S4 includes the following steps:

[0054] Step S41: obtaining core material data;

[0055] Step S42: extracting the hysteresis curve segment of the core material data to construct an initial material saturation curve; performing amplitude normalization processing on the bias magnetic sensitivity index data, and correlating and interpolating it with the material saturation curve data to calculate the real-time saturation response parameters and generate dynamic magnetic response fusion data;

[0056] Step S43: fitting the dynamic magnetic response fusion data to the magnetic bias danger interval, extracting the inflection point change rate, and generating a real-time DC magnetic bias dynamic warning threshold; performing interval comparison on the real-time DC magnetic bias dynamic warning threshold and a preset multi-level warning threshold curve, identifying the warning level interval, and generating a DC magnetic bias graded warning result, wherein the DC magnetic bias graded warning result includes a first-level warning, a second-level warning, and a third-level warning;

[0057] Step S44: matching the corresponding protection strategy template according to the DC bias magnetization graded warning result, and outputting DC bias magnetization graded protection control data, wherein the protection strategy template includes a light intervention strategy template, a moderate intervention strategy template, and a heavy intervention strategy template.

[0058] This method extracts the core material's hysteresis curve to construct a material saturation curve. Normalized bias sensitivity indexes are then interpolated with the saturation curve to generate real-time magnetic response fusion data. This transforms bias monitoring from a single signal indicator to a fusion of material physical properties, improving the physical accuracy and scientific validity of the early warning data. By fitting the hazard intervals and extracting the inflection point change rate from the dynamic magnetic response fusion data, the dangerous evolution of DC bias can be accurately characterized, and real-time warning thresholds can be dynamically generated, making the early warning mechanism more sensitive and consistent with actual operating conditions. By comparing the real-time warning thresholds with preset multi-level threshold curves, a clear distinction is made between primary, secondary, and tertiary warning levels, enabling hierarchical management of bias risk and facilitating differentiated intervention measures for different risk levels. Based on the graded warning results, mild, moderate, and severe intervention strategy templates are automatically matched, achieving intelligent and automated protection and control, improving system response speed and decision-making accuracy, and reducing the risk of human error. The saturation curves and dynamic thresholds constructed using core material property data adapt to the material differences of different transformer types and specifications, enhancing the versatility and scalability of the method. Through real-time monitoring and precise early warning, magnetic bias anomalies are detected early, enabling targeted protective measures to be implemented promptly. This effectively reduces core loss, noise, and vibration, extending equipment life, and ensuring the safe and stable operation of the power system. This hierarchical early warning and protection control system provides a scientific basis for decision-making in transformer operation and maintenance, promoting intelligent O&M management of power equipment, improving O&M efficiency, and reducing maintenance costs.

[0059] Preferably, step S44 includes the following steps:

[0060] Step S441: Matching the corresponding protection strategy template according to the DC bias magnetization graded warning result. When the DC bias magnetization warning level is level one, matching the light intervention strategy template and performing the following control operations to obtain level one DC bias magnetization protection control data, automatically adjusting the on-load tap-changing transformer tap to the next level closest to the magnetic balance position, maintaining the transformer operating voltage stable, and sending a status prompt message to the monitoring system without the need for immediate circuit breaker or switching to a standby transformer.

[0061] Step S442: When the DC bias magnetic warning level is level 2, the moderate intervention strategy template is matched and the following control operations are performed to obtain level 2 DC bias magnetic protection control data: the on-load tap changer is forced to shift toward the dual-gear position, and the magnetic field balance control algorithm is enabled to optimize the core magnetic flux density distribution. At the same time, the delayed locking logic is activated and a 30-minute observation period is entered.

[0062] Step S443: When the DC bias magnetic warning level is level 3, the heavy intervention strategy template is matched and the following control operations are performed to obtain level 3 DC bias magnetic protection control data: the main circuit breaker locking process is initiated to cut off the connection path between the main transformer and the power grid, and a control command is issued to start the standby transformer to enter the power supply state, and the protection action data and electrical parameters are synchronously recorded;

[0063] Step S444: Integrate the first-level DC bias protection control data, the second-level DC bias protection control data, and the third-level DC bias protection control data to obtain DC bias graded protection control data.

[0064] This invention divides DC bias magnetic warnings into three levels: primary, secondary, and tertiary, matching light, moderate, and severe intervention strategies, respectively. This ensures that protection measures align with the level of bias magnetic risk, avoids excessive or insufficient intervention, and improves system safety and cost-effectiveness. During a primary warning, the on-load tap changer is fine-tuned to a position close to the magnetic equilibrium position, achieving mild automatic adjustment of the bias magnetic field. This maintains voltage stability, reduces interference with normal operation, and prevents unnecessary tripping and outages. During a secondary warning, the tap changer is forced to shift between two positions and the magnetic field balance control algorithm is activated. This dynamically optimizes the core magnetic flux distribution. Combined with a delayed lockout observation period, this effectively suppresses the deterioration of bias magnetic field and improves the proactiveness and accuracy of bias magnetic control. During a third-level warning, the main circuit breaker is activated to shut off the faulty main transformer and rapidly switch to a backup transformer to ensure uninterrupted power supply, prevent widespread power outages caused by bias magnetic faults, and enhance the grid's risk resilience. Automatically matching protection strategies and executing corresponding operations throughout the entire process reduces manual intervention, improves response speed and accuracy, reduces the risk of human error, and enhances the overall intelligence of the protection system. During the three-level early warning phase, protection actions and electrical parameter data are recorded simultaneously, facilitating subsequent fault analysis, maintenance decision-making, and equipment health management, supporting the development of an intelligent operation and maintenance system. Effectively integrating protection and control data at different levels creates a complete, hierarchical protection and control output, facilitating centralized system management and monitoring, and improving system coordination, control, and response efficiency.

[0065] In this specification, a real-time monitoring system for DC bias of a transformer is provided, which is used to implement the above-mentioned real-time monitoring method for DC bias of a transformer. The real-time monitoring system for DC bias of a transformer includes:

[0066] The signal optimization module is used to obtain the transformer primary side current signal and the core vibration signal; analyze the zero flux closed-loop characteristics of the transformer primary side current signal and eliminate the temperature drift of the transformer primary side current signal to generate the transformer primary side optimization signal; perform magnetostrictive vibration noise separation on the core vibration signal to generate the core vibration separation signal;

[0067] The magnetostrictive characteristic analysis module is used to dynamically filter out the power frequency fundamental wave of the transformer primary-side optimization signal to extract the pure DC component; perform wavelet packet decomposition on the core vibration separation signal, and extract the magnetostrictive characteristic spectrum of the decomposed core vibration separation signal to obtain the core magnetostrictive characteristic spectrum;

[0068] The model association module is used to construct a DC component-vibration spectrum association model; the pure DC component is used as the model input, and the core magnetostrictive characteristic spectrum is used as the model output to update the model coefficient of the DC component-vibration spectrum association model to obtain the bias magnetic sensitivity index;

[0069] The monitoring and control module is used to obtain core material data; construct a material saturation curve based on the core material data, and calculate the real-time DC bias magnetic dynamic warning threshold through the bias magnetic sensitivity index and the material saturation curve; compare the DC bias magnetic dynamic warning threshold with the preset multi-level warning threshold curve, thereby triggering a graded protection action and obtaining DC bias magnetic graded protection control data.

[0070] The beneficial effects of the present invention are that the signal optimization module effectively restores the true trend of current changes by analyzing the zero-flux closed-loop characteristics of the current signal and eliminating temperature drift errors. At the same time, it separates the magnetostrictive noise in the core vibration signal, accurately retaining the key response components, and providing a high-quality data foundation for subsequent analysis. The magnetostrictive characteristic analysis module not only dynamically filters out the power frequency fundamental wave in the current and extracts the pure DC component, but also accurately extracts the magnetostrictive main frequency and subharmonic frequency related to bias magnetization in the core vibration through wavelet packet decomposition and spectrum energy analysis, thereby enhancing the sensitivity of identifying weak bias magnetization phenomena. The model association module establishes a coupled correlation model between the DC component in the current signal and the magnetostrictive spectrum, and updates the model parameters in real time based on sample data, giving the model adaptive learning capabilities, improving the system's real-time response to the transformer's bias magnetization state and modeling accuracy. Through the bias magnetization sensitivity index output by the model, the system can quantitatively evaluate the degree of bias magnetization's impact on core vibration, forming an indicator with physical significance, which facilitates digital monitoring, trend analysis, and alarm determination. The monitoring and control module uses the core material data to construct a saturation curve, and combines it with the real-time bias magnetic sensitivity index to calculate the dynamic warning threshold, so as to realize the setting of personalized protection strategies for transformers with different materials and structures, and improve the universality and professionalism of the system. By comparing the dynamic warning threshold with the preset multi-level threshold curve, the three-level warning and hierarchical control strategy are linked to effectively control the evolution process of bias magnetic risk and avoid premature retirement or sudden failure of equipment. The data between modules are closed-loop and efficiently collaborate, supporting online operation, real-time analysis, and automatic response, which significantly improves the intelligence and automation level of transformer bias magnetic status monitoring and protection, and reduces the pressure of manual operation and maintenance. Therefore, the present invention improves the accuracy, real-time performance and hierarchical protection response capability of transformer DC bias magnetic monitoring through multi-source signal fusion and material property modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic flow chart of the steps of a method for real-time monitoring of DC bias magnetic field of a transformer;

[0072] Figure 2 for Figure 1 Detailed implementation steps of step S1 in FIG.

[0073] Figure 3 for Figure 1 Detailed implementation steps of step S4 in FIG.

[0074] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0075] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0076] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0077] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0078] To achieve this, please refer to Figures 1 to 3 A method for real-time monitoring of DC bias magnetic field of a transformer is provided, the method comprising the following steps:

[0079] Step S1: obtaining a transformer primary-side current signal and a core vibration signal; analyzing the zero-flux closed-loop characteristic of the transformer primary-side current signal and performing temperature drift elimination on the transformer primary-side current signal to generate a transformer primary-side optimization signal; performing magnetostrictive vibration noise separation on the core vibration signal to generate a core vibration separation signal;

[0080] Step S2: Dynamically filter out the power frequency fundamental wave of the transformer primary side optimization signal to extract the pure DC component; perform wavelet packet decomposition on the core vibration separation signal, and extract the magnetostrictive characteristic spectrum of the decomposed core vibration separation signal to obtain the core magnetostrictive characteristic spectrum;

[0081] Step S3: constructing a DC component-vibration spectrum correlation model; using the pure DC component as the model input and the core magnetostrictive characteristic spectrum as the model output, updating the model coefficients of the DC component-vibration spectrum correlation model to obtain the bias magnetic sensitivity index;

[0082] Step S4: Obtain core material data; construct a material saturation curve based on the core material data, and calculate the real-time DC bias magnetic dynamic warning threshold through the bias magnetic sensitivity index and the material saturation curve; compare the DC bias magnetic dynamic warning threshold with the preset multi-level warning threshold curve to trigger the graded protection action and obtain the DC bias magnetic graded protection control data.

[0083] In this embodiment of the present invention, the transformer primary-side current signal and core vibration signal are synchronously acquired. The current signal is acquired using a wide-bandwidth Hall effect current sensor with a frequency response ranging from DC to 10 kHz to ensure the capture of DC bias components and current harmonics. Because the sensor output is often affected by ambient temperature, the zero-flux closed-loop characteristics of the primary-side current signal are analyzed. A drift model is then established using temperature sensor data. A dynamic correction algorithm (such as temperature fitting residual regression) is then used to eliminate temperature drift and output an optimized primary-side signal. The core vibration signal is acquired using a high-sensitivity accelerometer. This signal contains a mixture of magnetostrictive vibration components and background mechanical noise. Using feature matching filtering and a bandpass isolation algorithm, background vibration interference is filtered out, and vibration components related to magnetostriction are extracted to generate a core vibration separation signal. The optimized primary-side current signal is then subjected to dynamic power frequency filtering. Adaptive notch filtering or harmonic prediction cancellation techniques are used to precisely filter out the 50 Hz power frequency component and its adjacent interference, extracting a pure DC component. This DC bias is a key indicator for determining magnetic saturation trends. At the same time, the core vibration separation signal is decomposed using wavelet packets to obtain a fine-grained spectral representation across multiple frequency bands. Energy threshold analysis and spectral packet energy ratios are used to extract the core's magnetostrictive characteristic spectrum, such as a specific energy concentration band ranging from hundreds of Hz to kHz. A DC component-vibration spectrum correlation model is constructed to quantify the impact of DC bias current on the core's vibration spectrum. This model can employ polynomial regression, neural networks, or recursive least squares modeling, taking as input a real-time DC component sequence and outputting the corresponding vibration characteristic spectrum. This model uses iterative coefficient optimization during continuous sampling to extract the response sensitivity between current and vibration, defined as the bias sensitivity index. This index characterizes the core's response to DC bias and serves as a key variable for subsequent protection threshold determination. Furthermore, core material parameter data, such as silicon steel grade, saturation magnetic induction (Bs), and magnetic permeability curve, are incorporated to establish a magnetic saturation characteristic curve model for the core material. Based on this model and the bias sensitivity index calculated in step S3, a real-time DC bias dynamic warning threshold is derived. Specifically, the critical bias current value before bias occurs is predicted under the current material state and operating conditions. This dynamic threshold is then compared with the system's pre-set multi-level warning threshold curve, such as mild, moderate, and severe bias levels. The level is determined and the corresponding DC bias graded protection control action, such as alarm, current limiting, bypass, or system shutdown, is triggered. The final graded protection control data is then output.

[0084] As an example of the present invention, refer to Figure 2 As shown, in this example, step S1 includes:

[0085] Step S11: collecting the transformer primary side current signal through a wide-band Hall effect sensor; collecting the core vibration signal through a high-frequency piezoelectric accelerometer;

[0086] Step S12: extracting the frequency band characteristics of the transformer primary side current signal to obtain the power frequency main component and the high-frequency harmonic component, and generating initial frequency band analysis data; identifying the zero flux reference segment of the initial frequency band analysis data, and generating zero flux reference segment data;

[0087] Step S13: performing temperature drift estimation on the zero flux reference segment data to generate temperature drift estimation data; performing thermal drift back calculation correction on the transformer primary side current signal data using the temperature drift estimation data to generate transformer primary side optimized signal data;

[0088] Step S14: performing spectrum structure analysis on the core vibration signal data to generate vibration spectrum structure data; extracting the magnetostrictive characteristic frequency band of the vibration spectrum structure data, and combining the vibration spectrum structure data and the magnetostrictive characteristic frequency band data for mutually exclusive decoupling to generate magnetostrictive vibration principal component data;

[0089] Step S15: Perform time domain signal restoration on the magnetostrictive vibration main component data to generate core vibration separation signal data.

[0090] In this embodiment of the present invention, a wideband Hall-effect sensor is used to sample the transformer's primary current. This sensor has a frequency response range of 0 Hz to 10 kHz, enabling it to capture both power frequency signals (e.g., 50 / 60 Hz) and extract their high-frequency harmonics and interference components. A sampling frequency of 25 kHz or higher is recommended to ensure high signal fidelity. Furthermore, a high-frequency piezoelectric accelerometer (e.g., operating in a 1 kHz–20 kHz frequency band and with a sensitivity greater than 100 mV / g) is attached to key locations of the transformer core (e.g., the yoke or core legs) to collect vibration signals, thereby capturing the structural vibration response caused by the magnetostrictive effect. The collected current signal is spectrally analyzed, extracting its frequency characteristics using techniques such as fast Fourier transform (FFT). The signal is then divided into a main power frequency component (centered at 50 Hz ± 5 Hz) and multiple high-frequency harmonic segments (e.g., 150 Hz, 250 Hz, and their multiples). This information is then used to generate initial frequency band analysis data. The zero-flux state of the current signal (i.e., the current reference state when the flux is at equilibrium) at different time intervals is further identified. By matching the characteristics of the frequency band equilibrium point, zero-flux reference segment data is calibrated. This reference segment serves as the basis for subsequent thermal drift estimation. Temperature drift estimation is performed based on the zero-flux reference segment data. Specifically, the current baseline variation trend at different time points within the reference segment is constructed. Combined with the transformer ambient temperature variation model, a temperature drift model is established using polynomial fitting or adaptive filtering methods, and temperature drift estimation data is calculated. This data is used to compensate the full-segment current signal. A thermal drift inverse modeling algorithm (such as sliding window inverse regression) is used to correct the amplitude drift caused by temperature. The optimized signal data for the transformer primary side is output to improve the magnetic response accuracy of the current signal. The spectral structure of the core vibration signal is analyzed. High-resolution spectral estimation (such as the Welch method or wavelet transform) is used to generate vibration spectral structure data. The characteristic frequency band related to magnetostriction (typically in the range of 1kHz to 3kHz) is extracted from this data and its magnetostrictive characteristic frequency band is identified. Subsequently, mutually exclusive decoupling algorithms (such as principal component analysis (PCA) or independent component analysis (ICA)) are used to isolate the magnetostrictive frequency band from mechanical or electrical noise components, generating clear magnetostrictive vibration principal component data. Inverse spectral transformation techniques (such as short-time inverse FFT or inverse wavelet transform) are then used to restore the magnetostrictive principal component in the time domain, converting it into a time-correlated vibration signal. This generates core vibration separation signal data, which serves as a key indicator for determining the degree of core bias vibration.

[0091] Preferably, identifying the zero flux reference segment of the initial frequency band analysis data includes:

[0092] Perform a frequency domain scan on the frequency band analysis data. When the frequency component is stably distributed between 0 Hz and 10 kHz and the amplitude fluctuation range is less than ±3 μV, the frequency band is identified as the initial frequency band data.

[0093] In the initial frequency band data, detect the data segment containing the lowest magnetic flux disturbance interval. If all the following conditions are met, mark it as a candidate zero magnetic flux reference segment: the magnetic flux change rate is less than 1nWb / s, the maximum value of the signal derivative does not exceed 0.5μV / ms, and the continuous stabilization time is not less than 60 seconds;

[0094] A multi-point statistical stationarity analysis is performed on the candidate zero-flux reference segment to calculate its mean μ and standard deviation σ. When μ∈(–2μV, +2μV), σ≤0.8μV, and no abnormal frequency transition exceeding 50Hz occurs in the segment, the segment is confirmed to be the zero-flux reference segment.

[0095] In an embodiment of the present invention, a full-frequency domain scan is performed on the frequency band analysis data. During the scan, if the frequency components of the signal are stably distributed between 0 Hz and 10 kHz for a continuous period of time (e.g., no less than 30 seconds), and the signal amplitude fluctuation range is maintained within ±3 μV, the signal segment is considered to be at the noise floor and not affected by significant operating disturbances. This segment is marked as the initial frequency band data. This step uses high-resolution short-time Fourier transform (STFT) or wavelet packet decomposition to evaluate dynamic frequency and amplitude changes. Within the initial frequency band data, the magnetic flux change characteristics implied by the signal are further analyzed. For each time window, the magnetic flux rate of change (in nWb / s) is calculated, and the maximum value of the signal derivative (reflecting the transient steepness of the signal) is extracted. If all of the following conditions are met: the magnetic flux rate of change is less than 1 nWb / s; the maximum value of the first-order derivative of the current signal does not exceed 0.5 μV / ms; and the continuous stability lasts for at least 60 seconds, the data segment is marked as a candidate zero-flux reference segment. This ensures that the selected data segment reflects the electromagnetic stability state, that is, the operating range where the transformer flux is closest to the equilibrium zero point. A multi-point statistical stationarity analysis is performed on the candidate zero-flux reference segment, calculating the sliding window mean and standard deviation within the selected segment. Specific conditions are as follows: the signal mean μ is within the interval: μ∈ (–2μV, +2μV); the signal standard deviation σ is less than or equal to σ≤0.8μV; and the signal spectrum within the segment is verified to be free of sudden frequency transitions (frequency abrupt changes exceeding ±50Hz). If the candidate segment meets all three of these statistical conditions, it is confirmed as a zero-flux reference segment.

[0096] Preferably, in step S2, dynamically filtering out the power frequency fundamental wave of the transformer primary side optimization signal to extract the pure DC component includes:

[0097] Perform data window framing on the transformer primary side optimization signal data to generate framed signal data;

[0098] Performing short-time Fourier transform on the framed signal data to generate frame-level spectrum distribution data;

[0099] Extract the main peak frequency of the power frequency from the frame-level spectrum distribution data;

[0100] Build a real-time sine reference template based on the power frequency peak frequency data to generate power frequency template signal data;

[0101] Perform signal synchronization cancellation on the transformer primary side optimized signal data and the power frequency template signal data to generate signal data after power frequency removal;

[0102] The signal data after power frequency removal is low-pass filtered to generate pure DC component data on the primary side of the transformer.

[0103] In an embodiment of the present invention, a sliding window mechanism is used to perform time-domain framing on the optimized transformer primary-side current signal. The length of each frame is set to 1 to 3 times the power frequency period (i.e., 20 to 60 ms) based on the power frequency period (e.g., 50 Hz corresponds to 20 ms) to ensure a balance between frequency resolution and time response, generating framed signal data. The window overlap rate can be set to 50% to enhance continuity. A short-time Fourier transform (STFT) is performed on each frame signal to extract its spectral structure and generate corresponding frame-level spectral distribution data. In this spectrum, the main frequency component generally appears stably between 49.5 Hz and 50.5 Hz. The main frequency position and peak amplitude are detected to obtain the main power frequency peak frequency data. Based on the detected main power frequency peak frequency, a sinusoidal reference signal template with consistent amplitude, phase, and frequency is constructed in real time. The constructed template uses the following expression: S(t) = A⋅sin(2πft+ϕ); where A is the amplitude of the main power frequency peak, f is the main power frequency peak frequency (e.g., 49.8Hz), and ϕ is the phase offset of the current frame. Three parameters are dynamically adjusted using a synchronous phase-locked method to generate the power frequency template signal data. The original transformer primary-side optimized signal and the power frequency template signal are synchronized in time domain phase and amplitude, and then a point-to-point subtraction process is performed to dynamically eliminate the main frequency component, resulting in the power frequency-removed signal data. The core of this process is phase synchronization accuracy control to minimize phase cancellation residue. A low-pass filter with a cutoff frequency of less than 10Hz is then applied to the power frequency-removed signal data to remove remaining high-frequency harmonics and noise components, ultimately generating smooth and stable primary-side pure DC bias magnetic component data. Either an IIR Butterworth low-pass filter or an FIR smoothing filter can be used for this purpose, with the order and response time adjustable based on the real-time requirements of the project.

[0104] Preferably, in step S2, performing wavelet packet decomposition on the core vibration separation signal and extracting magnetostrictive characteristic spectrum of the decomposed core vibration separation signal includes:

[0105] Perform wavelet packet decomposition on the core vibration separation signal to generate multi-layer wavelet packet coefficient data;

[0106] Performing frequency band reconstruction on the multi-layer wavelet packet coefficient data to generate core vibration frequency band reconstruction signal data;

[0107] Perform energy normalization on the core vibration frequency band reconstructed signal data to generate frequency band energy distribution data;

[0108] identifying a frequency band having a magnetostrictive response characteristic based on the frequency band energy distribution data, and generating magnetostrictive frequency band candidate data;

[0109] The spectrum line refinement and amplitude envelope analysis are performed on the candidate data of the magnetostrictive frequency band to extract the main frequency and subharmonic frequency characteristics and generate the magnetostrictive characteristic spectrum data of the iron core.

[0110] In an embodiment of the present invention, a wavelet packet decomposition algorithm is used to perform multi-layer decomposition on the core vibration separation signal in the time domain. It is preferred to use a "sym8" or "db8" type wavelet function to perform 3-layer or 4-layer wavelet packet decomposition on the signal, thereby dividing the signal into multiple independent frequency band components and generating multi-layer wavelet packet coefficient data. Wavelet packet decomposition has a finer resolution for high-frequency components than traditional wavelet transform, and is more suitable for local feature analysis of non-stationary vibration signals. Based on the decomposed multi-layer wavelet packet coefficients, each frequency band is inversely reconstructed to obtain the distribution of the original signal in each frequency band, and generate core vibration frequency band reconstruction signal data. Each reconstructed signal represents a signal reconstruction component of a corresponding frequency band range, which is convenient for subsequent energy distribution analysis. Energy calculation is performed on the reconstructed signal of each frequency band, which is defined as follows: ;in For the The signal energy in each frequency band, To correspond to the reconstructed signal, All frequency band energies are normalized for the sample points to generate frequency band energy distribution data. Normalization helps identify frequency bands with relatively concentrated energy, which are associated with magnetostrictive coupling. Based on the normalized energy distribution, an energy threshold (e.g., a normalized value > 0.15 or a local peak) is set as a candidate magnetostrictive property criterion. Several frequency bands with abnormally enhanced energy are selected as candidate magnetostrictive frequency bands. High-resolution FFT analysis is performed on the candidate frequency band signals to extract spectral line details. Hilbert transform or envelope demodulation is then performed to calculate the amplitude envelope curve. Spectral line location is then used to extract the main frequency and subharmonic frequency components in the frequency domain. The envelope morphology is then combined to determine the correlation between the magnetostrictive source and the core magnetostrictive characteristic spectrum data. These spectrum data are generally concentrated in the frequency range of 600 Hz to 3000 Hz, reflecting the typical magneto-elastic coupling vibration characteristics of ferromagnetic materials.

[0111] Preferably, constructing the DC component-vibration spectrum correlation model in step S3 includes:

[0112] Perform time domain window segmentation on the extracted DC component data of the primary side of the transformer to generate windowed DC feature sequence data;

[0113] Perform frequency band energy reorganization on the magnetostrictive characteristic spectrum data of the core to generate the main component characteristic data of the spectrum;

[0114] Calculate the spectral domain local abnormality index of the spectral principal component characteristic data;

[0115] The spectral domain local anomaly index is used to perform magnetic vibration pseudo-coupling stripping on the main component characteristic data of the spectrum, identify the pseudo-coupling band formed by the interference of power frequency harmonics and mechanical noise, and perform directional interference band shearing to generate the net spectral characteristic data after stripping;

[0116] Perform time index alignment on the windowed DC feature sequence data and the stripped net spectrum feature data to generate joint feature time series matrix data;

[0117] Perform cointegration analysis and partial least squares regression calculation on the joint characteristic time series matrix data, extract multidimensional correlation factors, and generate correlation coupling parameter set data;

[0118] A DC component-magnetically induced vibration spectrum correlation model is constructed based on the associated coupling parameter set data.

[0119] In this embodiment of the present invention, the extracted pure DC component data from the transformer primary side is segmented into time windows. Preferably, a sliding window or a fixed-frame window (e.g., every 60 seconds or 1000 sampling points) is used for time series segmentation, thereby generating windowed DC feature sequence data with time resolution. This sequence describes the slow-changing trend and local fluctuation characteristics of the DC component during operation. Next, feature reconstruction is performed on the core magnetostrictive characteristic spectrum data previously extracted via wavelet packet decomposition. This involves performing principal component analysis (PCA) or independent component analysis (ICA) on the energy features of multiple selected frequency bands (e.g., 600 Hz to 3 kHz) to obtain dimensionality-reduced spectral principal component feature data. This step extracts representative frequency components from high-dimensional spectral information, improving subsequent modeling efficiency. Subsequently, to identify and eliminate spurious correlation interference caused by power grid harmonics or motor mechanical noise, a spectral domain local anomaly index, such as a Z-score, a sudden gradient index, or a high-order cumulant function, is calculated for the spectral principal components. Based on this anomaly indicator, the spectral data is subjected to pseudo-coupling identification and shearing, extracting anomalous disturbance segments in the frequency domain. Directed frequency band stripping is then performed to obtain net spectral signature data, which primarily contains true magnetically induced vibration components. After spectral cleanup, the processed spectral signature data is time-aligned with the windowed DC component signature data to ensure that both current and spectral characteristics are present within each identical time window. This generates a joint signature time series matrix, where each row represents the combined state of the DC current amplitude and spectral energy characteristics for a given time segment. Cointegration analysis and partial least squares regression (PLSR) are then performed to reveal the long-term stable relationship and dominant correlation factors between the two signals. Cointegration analysis verifies whether there is a synergistic trend between the two signals, while PLSR extracts explanatory multidimensional correlation factors from the high-dimensional input, outputting the resulting data as a set of associated coupling parameters. Finally, a resolvable DC component-magnetically induced vibration spectrum correlation model is constructed based on this parameter set. This model can be used to predict whether a specific spectral response corresponds to a DC bias trend and to assess whether the DC component anomaly has triggered a resonant response of the core structure, thereby providing multi-source signal coupling support for bias status warning.

[0120] Of particular importance is the use of the spectral domain local anomaly index to perform magnetic vibration pseudo-coupling stripping on the spectral main component characteristic data, identify the pseudo-coupling bands formed by interference between power frequency harmonics and mechanical noise, and perform directional interference band shearing. This also includes:

[0121] Perform local window sliding analysis on the main component feature data of the spectrum, extract the energy gradient change rate of each spectrum segment, and generate spectral domain local abnormality index data;

[0122] According to the spectral domain local anomaly index data, the pseudo-coupling section of the spectrum main component characteristic data is located, the abnormal superposition characteristics of the magnetic vibration frequency band are extracted, and the initial identification data of the magnetic vibration pseudo-coupling is generated;

[0123] Perform interference source classification detection on the initial identification data of magnetic vibration pseudo-coupling, distinguish interference induced by power frequency harmonics from interference induced by mechanical noise, and generate interference type annotation data;

[0124] Based on the interference type annotation data, the corresponding pseudo-coupling frequency band is subjected to directional shearing processing, and high interference energy segments are removed to obtain directional interference band shearing data;

[0125] The directional interference band shearing data and the original spectrum principal component characteristic data are spliced ​​and reconstructed to generate the net spectrum characteristic data after stripping.

[0126] In this embodiment of the present invention, a sliding analysis of the spectral principal component feature data is performed within a preset narrowband frequency domain window (e.g., a 5Hz bandwidth) to calculate the energy gradient change rate within each spectral segment. This energy gradient change rate is used to measure the nonlinear energy mutation behavior between adjacent frequency bands, thereby constructing a spectral domain local anomaly index. A larger value indicates more dramatic local frequency band changes and a higher probability of coupling anomalies. Based on this spectral domain local anomaly index, high-index segments are located within the spectral principal component feature data. By comparing the center frequency and width of the anomalous frequency band with the overlap of the typical magnetic resonance response interval, regions with frequency band overlap anomaly characteristics are identified, generating preliminary initial identification data for magnetic resonance pseudo-coupling, which serves as a candidate set of pseudo-coupling bands. Interference source classification analysis is performed on the frequency band signals within the initial identification data, including: detecting the energy proportion of 50Hz and its integer harmonic components to identify frequency bands induced by power frequency harmonics; analyzing the time-frequency stability and random pulse energy distribution characteristics to identify frequency bands induced by mechanical noise; and generating corresponding interference type annotation data to clearly label each frequency band. A directional shearing strategy is employed for the marked interference types: for power frequency harmonic interference, band rejection filtering and phase compensation are used to eliminate the main interference frequency and its subharmonics; for mechanical noise interference, frequency band suppression and energy compression are used to mitigate its impact. After shearing, the output is sheared data of the directional interference band. This sheared data is then combined with the original, uninterrupted spectrum principal component feature data for spectral segment splicing and continuity correction. Weighted interpolation and edge transition smoothing are then used to generate the net spectral feature data after stripping for subsequent magnetic bias modeling and analysis.

[0127] Of particular importance is the positioning of pseudo-coupling segments of the spectrum principal component feature data based on the spectral domain local anomaly index data, and the extraction of magnetic resonance frequency band anomaly superposition features, which also includes:

[0128] Perform threshold positioning processing on the spectrum main component feature data according to the spectrum domain local abnormality index data, extract the index range of the abnormal high fluctuation frequency band, and generate spectrum high abnormality index segment data;

[0129] Continuous segment extraction is performed on the spectrum high abnormal index segment data, and the spectrum segments where the frequency jump threshold continuously meets the conditions are identified to generate preliminary data of the pseudo-coupling segment;

[0130] The resonance overlap ratio between adjacent frequency points is calculated using the pseudo-coupling segment preliminary data of the original spectrum principal component characteristic data, and the segments with magnetic-resonance energy coupling trends in the spectrum are extracted to generate the resonance overlap segment data.

[0131] Combine the resonance overlap segment data with the spectrum high anomaly index segment data to perform frequency band intersection operations, extract the stable overlap superposition response area, and generate magnetic resonance frequency band anomaly superposition feature data;

[0132] The abnormal superposition characteristic data of the magnetic vibration frequency band are subjected to spectrum continuity detection and noise influence factor elimination, the effective coupling interval is retained, and finally the initial identification data of magnetic vibration pseudo-coupling is generated.

[0133] In an embodiment of the present invention, a threshold is applied to the spectral principal component feature data using spectral local anomaly index data. This threshold is determined based on historical statistics and empirical parameters and is set to a critical value where abnormal fluctuations significantly exceed the normal background level. For frequency points where the spectral local anomaly index exceeds this threshold, the corresponding frequency band index is extracted to form an abnormally high-fluctuation frequency band index range, which is output as spectral high-anomaly index segment data. Frequency jump continuity is detected on the spectral high-anomaly index segment data, and frequency bands that meet continuity requirements (e.g., continuous index length not less than a preset threshold and stable jump amplitude) are selected as preliminary pseudo-coupling segment candidate sets, forming pseudo-coupling segment preliminary selection data. Using this pseudo-coupling segment preliminary selection data, the resonance overlap ratio between adjacent frequency points is calculated for the original spectral principal component feature data. Specifically, the energy coupling trend between the magneto-induced vibration and the interfering frequency band is quantified by analyzing the energy correlation and phase synchronization between the frequency points. Spectral segments with significant coupling characteristics are identified, and resonance overlap segment data is generated. The frequency segment intersection operation is performed on the resonance overlap segment data and the spectral high anomaly index segment data to screen out frequency band regions that meet both high anomaly fluctuation and strong coupling characteristics. These regions represent the superposition response of the magnetic resonance signal and the interference signal, generating magnetic resonance frequency band anomaly superposition feature data. Spectral segment continuity detection is performed on the magnetic resonance frequency band anomaly superposition feature data to eliminate non-continuous or isolated frequency points, ensuring that the selected interval has physical continuity and analytical significance. At the same time, by setting a noise impact factor threshold, false anomalies caused by random noise in the frequency band are eliminated, retaining the valid coupling interval, and ultimately outputting the initial identification data of magnetic resonance pseudo-coupling, which serves as the input basis for subsequent interference stripping processing.

[0134] Preferably, in step S3, the pure DC component is used as the model input, and the core magnetostrictive characteristic spectrum is used as the model output to update the model coefficients of the DC component-vibration spectrum correlation model, including:

[0135] The pure DC component is divided into time windows, representative DC time series segments are extracted, and DC input sample data is generated;

[0136] Extract the main frequency component and normalize the characteristics of the magnetostrictive characteristic spectrum of the iron core to generate normalized vibration output data;

[0137] Perform sample alignment and time-series synchronization on the DC input sample data and the normalized vibration output data to generate model-associated sample pairs;

[0138] Iteratively updating the least squares residual based on the model-related sample pairs adjusts the parameters of the DC component-magnetic vibration spectrum correlation model to generate updated model coefficient data;

[0139] The sensitivity change rate of the DC bias magnetization to the vibration spectrum is calculated based on the updated model coefficient data, and the bias magnetization sensitivity index is output.

[0140] In an embodiment of the present invention, the extracted pure DC component data is segmented using a sliding time window. The window length can be set based on the transformer's operating sampling frequency, such as 10 seconds, 30 seconds, or 1000 sampling points, to extract DC time series segments with stable fluctuation patterns. These segments are then constructed into DC input sample data with uniform dimensions, serving as the model's input features. Subsequently, the dominant frequency component is extracted from the simultaneously acquired core magnetostrictive characteristic spectrum data. Frequency bands with high energy density or dominant spectral peaks (e.g., 500 Hz to 2 kHz) are preferably extracted, and their amplitudes are normalized to generate dimensionless, highly comparable normalized vibration output data. Normalization can be performed using maximum and minimum value normalization or Z-score standard deviation normalization. Next, alignment is performed using time tags, matching the DC input sample data with the normalized vibration output data by time index to form one-to-one model-associated sample pairs. This ensures that each input segment corresponds to an output spectrum vector, enabling supervised learning of the model. Based on these sample pairs, an iterative least squares residual update algorithm (e.g., Recursive Least Squares, RLS) is executed. This algorithm iteratively updates the original model parameters based on the error between the current model output and the actual vibration data, optimizing the model's predictive capabilities under the current state. By continuously inputting multiple sample pairs, continuous adjustments and convergence are achieved, ultimately generating updated model coefficient data, allowing the model parameters to adapt to changes in operating conditions. Finally, using the updated model coefficients, the incremental impact of each unit DC component input on the spectral response, namely the sensitivity change rate of the vibration spectrum, is calculated to further quantify the output's response to the input. This sensitivity change rate is defined as the bias sensitivity index, which is used to assess whether the current bias state excites significant vibration modes, thus providing a core evaluation metric for DC bias trend warning and vibration intervention.

[0141] As an example of the present invention, refer to Figure 3 As shown, in this example, step S4 includes:

[0142] Step S41: obtaining core material data;

[0143] Step S42: extracting the hysteresis curve segment of the core material data to construct an initial material saturation curve; performing amplitude normalization processing on the bias magnetic sensitivity index data, and correlating and interpolating it with the material saturation curve data to calculate the real-time saturation response parameters and generate dynamic magnetic response fusion data;

[0144] Step S43: fitting the dynamic magnetic response fusion data to the magnetic bias danger interval, extracting the inflection point change rate, and generating a real-time DC magnetic bias dynamic warning threshold; performing interval comparison on the real-time DC magnetic bias dynamic warning threshold and a preset multi-level warning threshold curve, identifying the warning level interval, and generating a DC magnetic bias graded warning result, wherein the DC magnetic bias graded warning result includes a first-level warning, a second-level warning, and a third-level warning;

[0145] Step S44: matching the corresponding protection strategy template according to the DC bias magnetization graded warning result, and outputting DC bias magnetization graded protection control data, wherein the protection strategy template includes a light intervention strategy template, a moderate intervention strategy template, and a heavy intervention strategy template.

[0146] In this embodiment of the present invention, in step S41, data on the core material used in the target transformer is first collected, including key parameters such as material type (e.g., cold-rolled grain-oriented silicon steel, amorphous alloy), magnetic permeability, coercivity, remanence, and maximum magnetic flux density. The BH characteristic curve segments within the hysteresis loop experimental data are specifically extracted to construct an initial material saturation curve reflecting the material's magnetic saturation trend, which serves as a baseline reference model for bias response analysis. In step S42, the bias susceptibility index data obtained in the previous stage (e.g., step S3) is amplitude normalized and mapped to the [0, 1] interval to unify the measurement scale. Subsequently, the initial material saturation curve is interpolated using the normalized bias susceptibility index as an input parameter. Using piecewise linear interpolation or spline interpolation techniques, real-time saturation response parameters (e.g., equivalent magnetic flux density, permeability change rate, etc.) under the bias state are calculated, generating dynamic magnetic response fusion data reflecting the material's current bias influence. In step S43, a critical section model for DC bias evolution is constructed based on the dynamic magnetic response fusion data. The dynamic curve is modeled using fitting methods (such as polynomial fitting or piecewise function fitting), and the rate of change of its inflection point, or the rate of change of the critical point of magnetic saturation, is extracted to measure the proximity of the bias state to the danger zone. This result is defined as the real-time DC bias dynamic warning threshold. This dynamic warning threshold is then compared with a preset multi-level warning threshold curve (e.g., the boundaries of Levels 1, 2, and 3, defined based on engineering experience) to identify the warning level interval within which the current state falls. This generates a final DC bias warning result, which includes three levels: Level 1 (mild bias), Level 2 (moderate bias), and Level 3 (severe bias). In step S44, a template is matched from a pre-defined protection strategy library based on the identified bias level. Specifically, a Level 1 warning corresponds to a mild intervention strategy template, such as fine-tuning reactive power compensation or increasing the carrier frequency; a Level 2 warning corresponds to a moderate intervention strategy template, such as activating a dynamic debiasing algorithm or loading bypass flux compensation; and a Level 3 warning activates a severe intervention strategy template, including forced power reduction, linked protection tripping, or an intelligent soft shutdown mechanism. The final output DC bias magnetic graded protection control data will serve as the decision-making basis of the control system and be used to intervene in the operating status of the transformer in real time to ensure the operating safety and life span of the equipment.

[0147] Preferably, step S44 includes the following steps:

[0148] Step S441: Matching the corresponding protection strategy template according to the DC bias magnetization graded warning result. When the DC bias magnetization warning level is level one, matching the light intervention strategy template and performing the following control operations to obtain level one DC bias magnetization protection control data, automatically adjusting the on-load tap-changing transformer tap to the next level closest to the magnetic balance position, maintaining the transformer operating voltage stable, and sending a status prompt message to the monitoring system without the need for immediate circuit breaker or switching to a standby transformer.

[0149] Step S442: When the DC bias magnetic warning level is level 2, the moderate intervention strategy template is matched and the following control operations are performed to obtain level 2 DC bias magnetic protection control data: the on-load tap changer is forced to shift toward the dual-gear position, and the magnetic field balance control algorithm is enabled to optimize the core magnetic flux density distribution. At the same time, the delayed locking logic is activated and a 30-minute observation period is entered.

[0150] Step S443: When the DC bias magnetic warning level is level 3, the heavy intervention strategy template is matched and the following control operations are performed to obtain level 3 DC bias magnetic protection control data: the main circuit breaker locking process is initiated to cut off the connection path between the main transformer and the power grid, and a control command is issued to start the standby transformer to enter the power supply state, and the protection action data and electrical parameters are synchronously recorded;

[0151] Step S444: Integrate the first-level DC bias protection control data, the second-level DC bias protection control data, and the third-level DC bias protection control data to obtain DC bias graded protection control data.

[0152] In this embodiment of the present invention, when the system detects a DC bias magnetic warning level of Level 1 (mild bias), it applies a preset mild intervention strategy template. Specifically, this involves automatically instructing the on-load tap-changer (OLTC) to fine-tune the tap position, prioritizing adjustment to the next higher position closest to the current magnetic flux imbalance, thereby mitigating the magnetic flux deviation trend. This operation ensures that the core magnetic state is improved without interrupting power supply. Simultaneously, the system sends a status message to the monitoring center, informing operations and maintenance personnel that a mild bias intervention is in effect. This eliminates the need for circuit breaker disconnection or switching to a backup transformer, enabling low-impact online intervention. If the DC bias magnetic warning level is Level 2 (moderate bias), the corresponding moderate intervention strategy template is applied. At this point, the control system issues a forced adjustment command, causing the OLTC to adjust using a "dual-shift" method, shifting the tap changer by one shift in both the positive and negative directions to find a more optimal balance point, thereby increasing the adjustment range. At the same time, a magnetic field balancing control algorithm is activated to analyze and optimize the local magnetic flux density distribution in the core in real time, maximizing the balance of flux paths and slowing the rate at which certain magnetic paths enter local saturation zones. Furthermore, to prevent over-protection measures from impacting operational continuity, the system activates delayed blocking logic and enters a 30-minute observation period. During this period, the system tracks and re-evaluates the magnetic bias trend and load dynamics in real time to determine whether a subsequent response level escalation is necessary. If the warning level is upgraded to Level 3 (severe magnetic bias), the system immediately applies the severe intervention strategy template. The control system executes the following high-level protection actions: it initiates the main circuit breaker lockout process, disconnecting the main transformer from the main grid to prevent further faults caused by severe core saturation or thermal damage. Simultaneously, it sends a start command to the backup transformer system, rapidly switching to the backup power source to maintain power supply stability. The system also automatically and synchronously records the triggering data for the current protection action, including magnetic bias level, tap status, current and voltage parameters, and magnetic field status, and generates a comprehensive event log for subsequent analysis and fault attribution. The system integrates all levels of control response data, including primary, secondary, and tertiary DC bias protection data, to generate the final "DC bias graded protection control data." This data serves as decision-making input for the equipment control center to execute protection strategies. It also serves as a backup for bias operation behavior data on the intelligent monitoring platform for further training optimization or historical trend analysis.

[0153] In this specification, a real-time monitoring system for DC bias of a transformer is provided, which is used to implement the above-mentioned real-time monitoring method for DC bias of a transformer. The real-time monitoring system for DC bias of a transformer includes:

[0154] The signal optimization module is used to obtain the transformer primary side current signal and the core vibration signal; analyze the zero flux closed-loop characteristics of the transformer primary side current signal and eliminate the temperature drift of the transformer primary side current signal to generate the transformer primary side optimization signal; perform magnetostrictive vibration noise separation on the core vibration signal to generate the core vibration separation signal;

[0155] The magnetostrictive characteristic analysis module is used to dynamically filter out the power frequency fundamental wave of the transformer primary-side optimization signal to extract the pure DC component; perform wavelet packet decomposition on the core vibration separation signal, and extract the magnetostrictive characteristic spectrum of the decomposed core vibration separation signal to obtain the core magnetostrictive characteristic spectrum;

[0156] The model association module is used to construct a DC component-vibration spectrum association model; the pure DC component is used as the model input, and the core magnetostrictive characteristic spectrum is used as the model output to update the model coefficient of the DC component-vibration spectrum association model to obtain the bias magnetic sensitivity index;

[0157] The monitoring and control module is used to obtain core material data; construct a material saturation curve based on the core material data, and calculate the real-time DC bias magnetic dynamic warning threshold through the bias magnetic sensitivity index and the material saturation curve; compare the DC bias magnetic dynamic warning threshold with the preset multi-level warning threshold curve, thereby triggering a graded protection action and obtaining DC bias magnetic graded protection control data.

[0158] The beneficial effects of the present invention are that the signal optimization module effectively restores the true trend of current changes by analyzing the zero-flux closed-loop characteristics of the current signal and eliminating temperature drift errors. At the same time, it separates the magnetostrictive noise in the core vibration signal, accurately retaining the key response components, and providing a high-quality data foundation for subsequent analysis. The magnetostrictive characteristic analysis module not only dynamically filters out the power frequency fundamental wave in the current and extracts the pure DC component, but also accurately extracts the magnetostrictive main frequency and subharmonic frequency related to bias magnetization in the core vibration through wavelet packet decomposition and spectrum energy analysis, thereby enhancing the sensitivity of identifying weak bias magnetization phenomena. The model association module establishes a coupled correlation model between the DC component in the current signal and the magnetostrictive spectrum, and updates the model parameters in real time based on sample data, giving the model adaptive learning capabilities, improving the system's real-time response to the transformer's bias magnetization state and modeling accuracy. Through the bias magnetization sensitivity index output by the model, the system can quantitatively evaluate the degree of bias magnetization's impact on core vibration, forming an indicator with physical significance, which facilitates digital monitoring, trend analysis, and alarm determination. The monitoring and control module uses the core material data to construct a saturation curve, and combines it with the real-time bias magnetic sensitivity index to calculate the dynamic warning threshold, so as to realize the setting of personalized protection strategies for transformers with different materials and structures, and improve the universality and professionalism of the system. By comparing the dynamic warning threshold with the preset multi-level threshold curve, the three-level warning and hierarchical control strategy are linked to effectively control the evolution process of bias magnetic risk and avoid premature retirement or sudden failure of equipment. The data between modules are closed-loop and efficiently collaborate, supporting online operation, real-time analysis, and automatic response, which significantly improves the intelligence and automation level of transformer bias magnetic status monitoring and protection, and reduces the pressure of manual operation and maintenance. Therefore, the present invention improves the accuracy, real-time performance and hierarchical protection response capability of transformer DC bias magnetic monitoring through multi-source signal fusion and material property modeling.

[0159] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0160] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A real-time monitoring method for DC bias of a transformer, characterized in that: The following steps are involved: Step S1: obtaining a transformer primary-side current signal and a core vibration signal; analyzing the zero-flux closed-loop characteristic of the transformer primary-side current signal and performing temperature drift elimination on the transformer primary-side current signal to generate a transformer primary-side optimization signal; performing magnetostrictive vibration noise separation on the core vibration signal to generate a core vibration separation signal; Step S2: Dynamically filter out the power frequency fundamental wave of the transformer primary side optimization signal to extract the pure DC component; perform wavelet packet decomposition on the core vibration separation signal, and extract the magnetostrictive characteristic spectrum of the decomposed core vibration separation signal to obtain the core magnetostrictive characteristic spectrum; Step S3: constructing a DC component-vibration spectrum correlation model; using the pure DC component as the model input and the core magnetostrictive characteristic spectrum as the model output, updating the model coefficients of the DC component-vibration spectrum correlation model to obtain the bias magnetic sensitivity index; Step S4: Obtain core material data; construct a material saturation curve based on the core material data, and calculate the real-time DC bias magnetic dynamic warning threshold through the bias magnetic sensitivity index and the material saturation curve; compare the DC bias magnetic dynamic warning threshold with the preset multi-level warning threshold curve to trigger the graded protection action and obtain the DC bias magnetic graded protection control data.

2. The real-time monitoring method of transformer DC bias according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: collecting the transformer primary side current signal through a wide-band Hall effect sensor; collecting the core vibration signal through a high-frequency piezoelectric accelerometer; Step S12: extracting the frequency band characteristics of the transformer primary side current signal to obtain the power frequency main component and the high-frequency harmonic component, and generating initial frequency band analysis data; identifying the zero flux reference segment of the initial frequency band analysis data, and generating zero flux reference segment data; Step S13: performing temperature drift estimation on the zero flux reference segment data to generate temperature drift estimation data; performing thermal drift back calculation correction on the transformer primary side current signal data using the temperature drift estimation data to generate transformer primary side optimized signal data; Step S14: performing spectrum structure analysis on the core vibration signal data to generate vibration spectrum structure data; extracting the magnetostrictive characteristic frequency band of the vibration spectrum structure data, and combining the vibration spectrum structure data and the magnetostrictive characteristic frequency band data for mutually exclusive decoupling to generate magnetostrictive vibration principal component data; Step S15: Perform time domain signal restoration on the magnetostrictive vibration main component data to generate core vibration separation signal data.

3. The real-time monitoring method of transformer DC bias according to claim 2, characterized in that: Identifying the zero flux reference segment of the initial frequency band resolution data includes: Perform a frequency domain scan on the frequency band analysis data. When the frequency component is stably distributed between 0 Hz and 10 kHz and the amplitude fluctuation range is less than ±3 μV, the frequency band is identified as the initial frequency band data. In the initial frequency band data, detect the data segment containing the lowest magnetic flux disturbance interval. If all the following conditions are met, mark it as a candidate zero magnetic flux reference segment: the magnetic flux change rate is less than 1nWb / s, the maximum value of the signal derivative does not exceed 0.5μV / ms, and the continuous stabilization time is not less than 60 seconds; A multi-point statistical stationarity analysis is performed on the candidate zero-flux reference segment to calculate its mean μ and standard deviation σ. When μ∈(–2μV, +2μV), σ≤0.8μV, and no abnormal frequency transition exceeding 50Hz occurs in the segment, the segment is confirmed to be the zero-flux reference segment.

4. The real-time monitoring method of transformer DC bias according to claim 1, characterized in that: Dynamically filtering out the power frequency fundamental wave of the transformer primary side optimization signal to extract the pure DC component in step S2 includes: Perform data window framing on the transformer primary side optimization signal data to generate framed signal data; Performing short-time Fourier transform on the framed signal data to generate frame-level spectrum distribution data; Extract the main peak frequency of the power frequency from the frame-level spectrum distribution data; Build a real-time sine reference template based on the power frequency peak frequency data to generate power frequency template signal data; Perform signal synchronization cancellation on the transformer primary side optimized signal data and the power frequency template signal data to generate signal data after power frequency removal; The signal data after power frequency removal is low-pass filtered to generate pure DC component data on the primary side of the transformer.

5. The real-time monitoring method of transformer DC bias according to claim 1, characterized in that: In step S2, the wavelet packet decomposition of the core vibration separation signal and the magnetostrictive characteristic spectrum extraction of the decomposed core vibration separation signal include: Perform wavelet packet decomposition on the core vibration separation signal to generate multi-layer wavelet packet coefficient data; Performing frequency band reconstruction on the multi-layer wavelet packet coefficient data to generate core vibration frequency band reconstruction signal data; Perform energy normalization on the core vibration frequency band reconstructed signal data to generate frequency band energy distribution data; identifying a frequency band having a magnetostrictive response characteristic based on the frequency band energy distribution data, and generating magnetostrictive frequency band candidate data; The spectrum line refinement and amplitude envelope analysis are performed on the candidate data of the magnetostrictive frequency band to extract the main frequency and subharmonic frequency characteristics and generate the magnetostrictive characteristic spectrum data of the iron core.

6. The real-time monitoring method of transformer DC bias according to claim 1, characterized in that: Constructing the DC component-vibration spectrum correlation model in step S3 includes: Perform time domain window segmentation on the extracted DC component data of the primary side of the transformer to generate windowed DC feature sequence data; Perform frequency band energy reorganization on the magnetostrictive characteristic spectrum data of the core to generate the main component characteristic data of the spectrum; Calculate the spectral domain local abnormality index of the spectral principal component characteristic data; The spectral domain local anomaly index is used to perform magnetic vibration pseudo-coupling stripping on the main component characteristic data of the spectrum, identify the pseudo-coupling band formed by the interference of power frequency harmonics and mechanical noise, and perform directional interference band shearing to generate the net spectral characteristic data after stripping; Perform time index alignment on the windowed DC feature sequence data and the stripped net spectrum feature data to generate joint feature time series matrix data; Perform cointegration analysis and partial least squares regression calculation on the joint characteristic time series matrix data, extract multidimensional correlation factors, and generate correlation coupling parameter set data; A DC component-magnetically induced vibration spectrum correlation model is constructed based on the associated coupling parameter set data.

7. The real-time monitoring method of transformer DC bias according to claim 1, characterized in that: In step S3, the pure DC component is used as the model input, and the core magnetostrictive characteristic spectrum is used as the model output to update the model coefficients of the DC component-vibration spectrum correlation model, including: The pure DC component is divided into time windows, representative DC time series segments are extracted, and DC input sample data is generated; Extract the main frequency component and normalize the characteristics of the magnetostrictive characteristic spectrum of the iron core to generate normalized vibration output data; Perform sample alignment and time-series synchronization on the DC input sample data and the normalized vibration output data to generate model-associated sample pairs; Iteratively updating the least squares residual based on the model-related sample pairs adjusts the parameters of the DC component-magnetic vibration spectrum correlation model to generate updated model coefficient data; The sensitivity change rate of the DC bias magnetization to the vibration spectrum is calculated based on the updated model coefficient data, and the bias magnetization sensitivity index is output.

8. The real-time monitoring method of transformer DC bias according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: obtaining core material data; Step S42: extracting the hysteresis curve segment of the core material data to construct an initial material saturation curve; performing amplitude normalization processing on the bias magnetic sensitivity index data, and correlating and interpolating it with the material saturation curve data to calculate the real-time saturation response parameters and generate dynamic magnetic response fusion data; Step S43: fitting the dynamic magnetic response fusion data to the magnetic bias danger interval, extracting the inflection point change rate, and generating a real-time DC magnetic bias dynamic warning threshold; performing interval comparison on the real-time DC magnetic bias dynamic warning threshold and a preset multi-level warning threshold curve, identifying the warning level interval, and generating a DC magnetic bias graded warning result, wherein the DC magnetic bias graded warning result includes a first-level warning, a second-level warning, and a third-level warning; Step S44: matching the corresponding protection strategy template according to the DC bias magnetization graded warning result, and outputting DC bias magnetization graded protection control data, wherein the protection strategy template includes a light intervention strategy template, a moderate intervention strategy template, and a heavy intervention strategy template.

9. The real-time monitoring method of transformer DC bias according to claim 8, characterized in that: Step S44 includes the following steps: Step S441: Matching the corresponding protection strategy template according to the DC bias magnetization graded warning result. When the DC bias magnetization warning level is level one, matching the light intervention strategy template and performing the following control operations to obtain level one DC bias magnetization protection control data, automatically adjusting the on-load tap-changing transformer tap to the next level closest to the magnetic balance position, maintaining the transformer operating voltage stable, and sending a status prompt message to the monitoring system without the need for immediate circuit breaker or switching to a standby transformer. Step S442: When the DC bias magnetic warning level is level 2, the moderate intervention strategy template is matched and the following control operations are performed to obtain level 2 DC bias magnetic protection control data: the on-load tap changer is forced to shift toward the dual-gear position, and the magnetic field balance control algorithm is enabled to optimize the core magnetic flux density distribution. At the same time, the delayed locking logic is activated and a 30-minute observation period is entered. Step S443: When the DC bias magnetic warning level is level 3, the heavy intervention strategy template is matched and the following control operations are performed to obtain level 3 DC bias magnetic protection control data: the main circuit breaker locking process is initiated to cut off the connection path between the main transformer and the power grid, and a control command is issued to start the standby transformer to enter the power supply state, and the protection action data and electrical parameters are synchronously recorded; Step S444: Integrate the first-level DC bias protection control data, the second-level DC bias protection control data, and the third-level DC bias protection control data to obtain DC bias graded protection control data.

10. A real-time monitoring system for DC bias of transformer, characterized in that: The method for real-time monitoring of DC bias of a transformer according to claim 1 is used to implement the real-time monitoring system of DC bias of a transformer, the real-time monitoring system comprising: The signal optimization module is used to obtain the transformer primary side current signal and the core vibration signal; analyze the zero flux closed-loop characteristics of the transformer primary side current signal and eliminate the temperature drift of the transformer primary side current signal to generate the transformer primary side optimization signal; perform magnetostrictive vibration noise separation on the core vibration signal to generate the core vibration separation signal; The magnetostrictive characteristic analysis module is used to dynamically filter out the power frequency fundamental wave of the transformer primary-side optimization signal to extract the pure DC component; perform wavelet packet decomposition on the core vibration separation signal, and extract the magnetostrictive characteristic spectrum of the decomposed core vibration separation signal to obtain the core magnetostrictive characteristic spectrum; The model association module is used to construct a DC component-vibration spectrum association model; the pure DC component is used as the model input, and the core magnetostrictive characteristic spectrum is used as the model output to update the model coefficient of the DC component-vibration spectrum association model to obtain the bias magnetic sensitivity index; The monitoring and control module is used to obtain core material data; construct a material saturation curve based on the core material data, and calculate the real-time DC bias magnetic dynamic warning threshold through the bias magnetic sensitivity index and the material saturation curve; compare the DC bias magnetic dynamic warning threshold with the preset multi-level warning threshold curve, thereby triggering a graded protection action and obtaining DC bias magnetic graded protection control data.

Images