Loss diagnosis and health management system and method for transformer

Abstract

The invention discloses a loss diagnosis and health management system and method for a transformer, and belongs to the technical field of intelligent transformers, and the system comprises a thermal fingerprint module which is used for recording a temperature drop curve of each part of the transformer after planned power failure, and obtaining a thermal inertia fingerprint of the transformer based on the fitting of multiple temperature drop curves; the theoretical temperature module is used for calculating the theoretical cold body temperature of each part of the transformer based on the real-time environment temperature and the thermal inertia fingerprint when the transformer runs; the net temperature rise module is used for acquiring the actual temperature value of each part of the transformer and subtracting the corresponding theoretical cold body temperature from the actual temperature value; weak thermal characteristic signals caused by insulation aging and loss abnormity of the transformer can be accurately extracted from strong background noise such as load fluctuation and environment temperature through technologies such as thermal inertia fingerprint construction and net temperature rise separation, early diagnosis of progressive health deterioration is realized, support is provided for intervening potential faults in advance, and the fault diagnosis accuracy is improved. And the passive situation of post-treatment of operation and maintenance is changed.

Description

Technical Field

[0001] This invention relates to the field of intelligent transformer technology, and more specifically, to a transformer loss diagnosis and health management system and method. Background Technology

[0002] As a key piece of equipment in the power system, the operational stability of transformers directly determines the safety and reliability of power transmission. Currently, power system maintenance commonly employs devices such as WS-ID™ temperature controllers and TCM monitors to monitor transformer operating status. These devices focus on real-time collection of temperature data from various parts of the transformer, triggering over-temperature alarms when temperatures exceed preset thresholds, and sending notifications to maintenance personnel via telephone after a fault occurs. This constitutes a standard technical solution for transformer temperature monitoring and basic alarms. However, existing monitoring technologies based on these devices can only achieve passive monitoring and threshold alarms of transformer operating status, lacking the ability to proactively assess the health status of the equipment. In actual maintenance scenarios, these technologies often only trigger alarms after a transformer fault has manifested, such as when internal components overheat and cause a rapid temperature rise, thus failing to identify the gradual health degradation process occurring inside the transformer in its early stages. Furthermore, they struggle to support predictive maintenance, leaving maintenance work in a passive, reactive state, unable to intervene in potential faults in advance. This not only increases the risk of sudden equipment outages but also significantly raises maintenance costs and losses due to faults.

[0003] The underlying technical reason for this limitation lies in a flawed understanding of the parameters representing the transformer's health status. This involves directly equating temperature, a comprehensive characteristic parameter, with the equipment's overall health. Transformer health issues, such as insulation aging, abnormal core losses, and excessive local winding losses, do not initially cause significant temperature changes. They only produce extremely slight additional temperature rises or minor distortions in localized heat distribution. These weak thermal characteristic signals are easily drowned out by strong background noise such as load fluctuations and ambient temperature changes during long-term transformer operation, making them difficult to effectively identify. Existing monitoring technologies rely on absolute temperature values ​​or simple thermometers. While existing alarm judgment methods exist, they lack the ability to separate and quantify weak thermal characteristic signals of degradation from complex background noise. In actual operating conditions where transformers operate for extended periods and load and ambient temperature change dynamically, maintenance personnel often set high temperature warning thresholds to pursue accurate early warnings and reduce false alarm rates. Although high thresholds can effectively avoid false alarms caused by load fluctuations and sudden changes in ambient temperature, they also treat the initial weak thermal characteristic signals that indicate major equipment failures as noise, resulting in a significant loss of sensitivity in early degradation diagnosis. This causes maintenance work to miss the best intervention opportunity, which may eventually lead to minor degradation gradually developing into a serious failure, severely affecting the stable operation of the power system. Summary of the Invention

[0004] To address the problems existing in the prior art, the present invention aims to provide a transformer loss diagnosis and health management system and method. This system can accurately extract weak thermal characteristic signals caused by transformer insulation aging and abnormal losses from strong background noise such as load fluctuations and ambient temperature through technologies such as thermal inertial fingerprint construction and net temperature rise separation. This enables early diagnosis of progressive health degradation, provides support for early intervention in potential faults, and changes the passive situation of post-event handling in operation and maintenance.

[0005] To solve the above problems, the present invention adopts the following technical solution: Firstly, a transformer loss diagnosis and health management system includes: The thermal fingerprint module is used to record the temperature drop curves of various parts of the transformer after a planned power outage, and obtains the thermal inertia fingerprint of the transformer based on the fitting of multiple temperature drop curves. The theoretical temperature module is used to calculate the theoretical cold body temperature of various parts of the transformer based on the real-time ambient temperature and thermal inertia fingerprint during transformer operation. The net temperature rise module is used to obtain the actual temperature values ​​of various parts of the transformer, and subtract the corresponding theoretical cold body temperature from the actual temperature value to obtain the net temperature rise of the heating element. The heat loss module is used to establish the relationship between the net heating element temperature rise and the load current, obtain the heat loss response curve, and calculate the curvature change of the heat loss response curve. The early warning judgment module is used to generate a health stress coefficient based on the curvature change and trigger an early warning based on the trend of the health stress coefficient change. The degradation location module is used to inject current pulses into the transformer windings after the warning is triggered, record the temperature rise response curves of various parts of the transformer to the current pulses, and determine the degradation location by analyzing the difference between the temperature rise response curves and the thermal inertia fingerprint.

[0006] Furthermore, the construction of thermal inertial fingerprints includes: For each planned power outage, the cooling process is analyzed by using the temperature drop curve reflected in the ambient temperature record to inversely deduce the dynamic heat dissipation coefficient set. Clustering and normalizing multiple sets of dynamic heat dissipation coefficients yields a baseline heat dissipation structure vector; Based on the baseline heat dissipation structure vector, a standard cooling expected curve cluster is generated through thermal network model simulation. Feature sets are extracted from the standard cooling expectation curve cluster to form a thermal inertial fingerprint.

[0007] Furthermore, the theoretical cold element temperature of each part of the transformer is calculated, including: Based on the real-time ambient temperature, the parameters of the reference heat dissipation structure vector in the thermal inertial fingerprint are remapped according to the temperature dependence of the material's thermal conductivity to generate a dynamic thermal conductivity matrix. Based on the dynamic thermal conductivity matrix and the initial state of the residual heat distribution of the transformer, the heat flow balance equation is constructed and solved to deduce the heat dissipation dynamics. The initial temperature and initial cooling rate features are extracted from the heat dissipation dynamics, matched with the standard cooling expectation curve cluster in the thermal inertial fingerprint, and the theoretical cold body temperature is output by interpolation calculation.

[0008] Furthermore, the construction of the loss thermal response curve includes: Based on the load current time series, the temperature rise sequence of the net heating element is compensated for hysteresis based on the thermal time constant to generate a quasi-steady-state temperature rise sequence and equivalent steady-state current value. Based on the equivalent steady-state current value, the quasi-steady-state temperature rise sequence is segmented to obtain multiple local linear relationship segments. Based on the physical model of copper loss and iron loss, the multiple local linear relationship segments are coupled to form the initial response curve. The confidence level is calculated based on the density and dispersion of data points in each current segment. The initial response curve is smoothed based on the confidence level and then weighted and fused with the historical loss thermal response curve to output the loss thermal response curve.

[0009] Furthermore, the change in curvature of the loss thermal response curve is calculated, including: By selecting a reference current point, the morphological deviation of the loss thermal response curve from the theoretical quadratic function basis within the preset current window of the reference current point is calculated to obtain the primary curvature characteristics. Based on the primary curvature characteristics, N detection segments are selected on the loss thermal response curve, and the rate of change of slope with current in each detection segment is extracted to form a multidimensional deformation feature vector. The multidimensional deformation feature vectors are mapped through a pre-constructed nonlinear feature interaction network to generate a comprehensive curvature degradation index. The trend test and persistence determination are performed on the comprehensive curvature degradation index sequence generated within a preset time window, and the curvature change is output.

[0010] Furthermore, the generation of the health stress coefficient includes: Calculate the standard score of the current curvature change based on the current curvature change sequence and the historical curvature change sequence; The standard score is mapped through a nonlinear transformation function based on the Arrhenius model to obtain the equivalent aging acceleration factor; The equivalent aging acceleration factor is multiplied by the differential of the current load rate and running time and then integrated to generate the cumulative health stress value as the health stress coefficient.

[0011] Furthermore, early warnings are triggered based on the changing trends of the health stress coefficient, including: Local linear fitting based on an adaptive sliding window is performed on the health stress coefficient sequence to extract the slope and confidence interval of the current trend; By combining the slope of the current trend, the confidence interval, and the absolute level of the health stress coefficient, a joint state assessment is conducted through a multi-level early warning rule table to generate a preliminary early warning level. Based on the initial warning level, the frequency of subsequent data collection and calculation will be dynamically adjusted, and when the preset level is reached, multi-cycle backtracking verification and consistency checks will be initiated to confirm the warning level. Based on the confirmed warning level, a tiered warning instruction is generated, which includes the warning level, potential fault mode, suggested response window, and maintenance priority.

[0012] Furthermore, the determination of the location of the deterioration includes: Extracting multi-time-constant spectra from temperature rise response curves; The abnormal time constant offset vector is generated by differentiating the multi-time constant spectrum with the health reference time constant spectrum stored in the thermal inertial fingerprint. The abnormal time constant offset vector is projected onto the preset fault location mapping matrix, and the confidence score of each candidate deterioration location is calculated. By integrating confidence scores, early warning levels, and historical diagnostic records of transformers, a location judgment conclusion and a recommended verification sequence are generated through a decision tree.

[0013] Furthermore, the following is also executed simultaneously during the construction of the thermal inertial fingerprint: When the transformer is in good condition, inject the same current pulse as during fault diagnosis into its windings and record the temperature rise response curves of each part. Extracting multi-time-constant spectra from temperature rise response curves; The multi-time constant spectrum is used as the health baseline time constant spectrum and stored in the thermal inertial fingerprint.

[0014] Secondly, the present invention also provides a method for transformer loss diagnosis and health management, specifically including the following steps: Step 1: Record the temperature drop curves of various parts of the transformer after the planned power outage, and obtain the thermal inertia fingerprint of the transformer based on the fitting of multiple temperature drop curves. Step 2: During transformer operation, calculate the theoretical cold body temperature of each part of the transformer based on the real-time ambient temperature and thermal inertia fingerprint. Step 3: Obtain the actual temperature values ​​of each part of the transformer, and subtract the corresponding theoretical cold body temperature from the actual temperature values ​​to obtain the net temperature rise of the heat-generating body. Step 4: Establish the correspondence between the net heating element temperature rise and the load current, obtain the loss thermal response curve, and calculate the curvature change of the loss thermal response curve. Step 5: Generate the health stress coefficient based on the curvature change, and trigger an early warning based on the trend of the health stress coefficient change. Step 6: After the warning is triggered, inject a current pulse into the transformer winding and record the temperature rise response curve of each part of the transformer to the current pulse. By analyzing the difference between the temperature rise response curve and the thermal inertia fingerprint, the deterioration area is determined.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This solution uses technologies such as thermal inertial fingerprint construction and net temperature rise separation to accurately extract weak thermal characteristic signals caused by transformer insulation aging and abnormal loss from strong background noise such as load fluctuation and ambient temperature, so as to realize early diagnosis of gradual health deterioration, provide support for early intervention of potential faults, and change the passive situation of post-event handling of operation and maintenance.

[0016] (2) This solution constructs the loss thermal response curve and analyzes the curvature change, and combines the comparison of multiple time constant spectra and the mapping of fault locations. It can not only quantify the degree of degradation, but also accurately locate the degradation area. This avoids the drawback of traditional monitoring that only knows the fault but cannot locate it. It allows the operation and maintenance work to shift from comprehensive investigation to precise handling, greatly improves the efficiency of fault handling, and reduces the operation and maintenance cost.

[0017] (3) This scheme balances the accuracy of early warning with the sensitivity of early diagnosis. Through the calculation of health stress coefficient and the evaluation of multi-level early warning rules, it avoids the missed reports caused by filtering weak deterioration signals with high thresholds and prevents false reports caused by low thresholds. It matches gradient disposal strategies for different deterioration levels to ensure the reliability of early warning and the rationality of operation and maintenance.

[0018] (4) This solution constructs a complete chain of deterioration identification, degree quantification, location positioning and early warning disposal, generates cumulative health stress data and trend analysis results, provides quantitative basis for predictive maintenance of transformers, can effectively delay the deterioration process and reduce the risk of sudden shutdown. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 This is a diagram showing the relationship between the various modules in a transformer loss diagnosis and health management system according to the present invention. Figure 2 This is a flowchart of a transformer loss diagnosis and health management method according to the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] Example 1: Please see Figure 1 A transformer loss diagnosis and health management system, comprising: The thermal fingerprint module is used to record the temperature drop curves of various parts of the transformer after a planned power outage. Based on fitting multiple temperature drop curves, the thermal inertia fingerprint of the transformer is obtained. The specific operation is as follows: The thermal fingerprint module captures the natural cooling process of the transformer after a planned power outage, records the temperature drop curves of various parts over time, and uses curve data from multiple power outage scenarios for fitting analysis to extract a thermal inertial fingerprint that is not affected by single operating conditions. The planned power outage scenario can eliminate the influence of load fluctuations on temperature changes, so that the temperature drop process is dominated only by the transformer's own thermal inertia and environmental heat dissipation conditions. Multiple data collections can cover different initial thermal states, ambient temperatures, and other scenarios, ensuring that the final thermal inertial fingerprint has comprehensiveness and stability. The thermal inertial fingerprint is essentially a comprehensive set of features reflecting the thermal conductivity and heat dissipation structure characteristics of various parts of the transformer.

[0023] The construction of thermal inertial fingerprints specifically includes the following steps: S11, for each planned power outage and subsequent cooling process, based on the temperature drop curve reflected in the ambient temperature record, the dynamic heat dissipation coefficient set is deduced in reverse. The specific operation is as follows: The temperature drop process of a transformer after a power outage is affected by both its own residual heat release and the heat dissipation from the environment. As an external variable, the ambient temperature directly changes the heat dissipation rate. Therefore, it is necessary to correct the temperature drop curve based on the real-time ambient temperature records during the cooling process. Based on the fundamental physical relationship between heat conduction and convection, the temperature drop rate is positively correlated with the heat dissipation coefficient, the difference between the ambient temperature and the transformer body temperature. By using the known ambient temperature sequence and the measured temperature drop curve, the dynamic heat dissipation coefficient set that changes over time can be obtained by reverse calculation. This dynamic heat dissipation coefficient set can reflect the changes in the heat dissipation capacity of various parts of the transformer under different temperature ranges and environmental conditions during the cooling process, avoiding the interference of ambient temperature fluctuations on the characterization of heat dissipation characteristics.

[0024] S12, clustering and normalizing multiple sets of dynamic heat dissipation coefficients to obtain the baseline heat dissipation structure vector, the specific operation is as follows: Multiple sets of dynamic heat dissipation coefficients correspond to the heat dissipation characteristics under different planned power outage scenarios. Affected by external factors such as initial temperature, ambient temperature fluctuations, and atmospheric flow conditions at each power outage, there are numerical differences between the coefficients, making them unsuitable for direct use in constructing a unified benchmark. Clustering processing employs conventional clustering algorithms in this field to group coefficients with similar characteristics into one class, effectively eliminating abnormal fluctuations caused by extreme environmental conditions or measurement errors, and retaining common heat dissipation characteristics in each group of coefficients to ensure the stability and reliability of subsequent analysis. Normalization processing uses linear or nonlinear transformation methods to unify the dimensions and numerical range of all dynamic heat dissipation coefficients, eliminating numerical deviations caused by differences in external conditions under different cooling scenarios, and making each set of dynamic heat dissipation coefficients horizontally comparable. After clustering and normalization processing, the obtained benchmark heat dissipation structure vector is no longer affected by external conditions in a single cooling process, accurately characterizing the heat dissipation structure characteristics of the transformer itself, and objectively reflecting the inherent distribution law of heat dissipation capacity of various parts of the transformer.

[0025] S13, based on the baseline heat dissipation structure vector, generates a standard expected cooling curve cluster through thermal network model simulation. The specific operation is as follows: The thermal network model is constructed based on the actual structural dimensions of the transformer and the inherent parameters such as the thermal conductivity of the material. The thermal network model is an existing model, and specific examples can be found in "Global Thermal Network Model of Oil-Immersed Transformer" published in the Journal of Metrology and "Simulation Analysis of Thermal Network Model of Dry Transformer Considering Fluid Dynamics" published in the Journal of Electrical Engineering. By abstracting the various parts of the transformer into interconnected nodes, the heat conduction path and heat dissipation capacity between the nodes are assigned values ​​through the benchmark heat dissipation structure vector, ensuring that the model can realistically replicate the inherent heat dissipation mechanism of the transformer. During the simulation, combined with the common initial thermal state of the transformer and the ambient temperature range, the power outage cooling process under different scenarios is simulated, generating multiple sets of corresponding temperature drop curves, i.e., the standard cooling expectation curve cluster. This curve cluster covers the typical cooling scenarios that may occur under normal transformer operation. Each curve corresponds to the theoretical cooling process under specific initial conditions and environmental parameters, which not only preserves the inherent thermal inertia characteristics of the transformer, but also reflects the influence of different external conditions on the cooling process.

[0026] S14, extract the feature set from the standard cooling expectation curve cluster to form a thermal inertial fingerprint. The specific operation is as follows: The standard cooling expectation curve cluster contains multiple theoretical cooling curves under different operating conditions. Each curve contains key information about the thermal inertia of the transformer. Redundant data needs to be removed through feature extraction to extract representative core features. The extraction process focuses on the inherent characteristic parameters of the curve, including but not limited to the initial cooling rate, characteristic temperature inflection point, curve decay trend, and steady-state temperature approach value. These parameters can accurately reflect the inherent thermal characteristics of the transformer, such as heat conduction efficiency and heat capacity distribution in various parts of the transformer. The characteristic parameters of all curves are integrated and screened to eliminate duplicate features and secondary interference features, forming a feature set that can comprehensively and uniquely characterize the thermal inertia characteristics of the transformer. This feature set is the thermal inertia fingerprint, which solidifies the inherent thermal characteristics of the transformer under normal conditions.

[0027] In some embodiments of the present invention, a theoretical temperature module is also included, which is used to calculate the theoretical cold body temperature of various parts of the transformer based on the real-time ambient temperature and thermal inertia fingerprint during transformer operation. The specific operation is as follows: During transformer operation, the theoretical temperature module calculates the theoretical cold body temperature of each part. The calculation results directly determine the accuracy of the net heat-generating body temperature rise, thus affecting the reliability of loss diagnosis and health assessment. The theoretical cold body temperature is essentially the temperature state of the transformer under the current ambient temperature, excluding the influence of load heating, determined only by environmental heat dissipation and its own thermal inertia. It needs to rely on the thermal inertia fingerprint pre-built by the thermal fingerprint module and dynamically extrapolate it in combination with the real-time ambient temperature. When the theoretical temperature module is running, it needs to simultaneously receive real-time ambient temperature data and thermal inertia fingerprint data. The real-time ambient temperature is collected by temperature sensors deployed around the transformer in locations free from direct sunlight and heat source interference. The collection frequency is consistent with the temperature sensors to ensure data timeliness. By dynamically correcting the benchmark parameters in the thermal inertia fingerprint, a heat conduction model adapted to the current environmental conditions is constructed to solve for the theoretical cold body temperature of each part. This temperature reflects both the inherent thermal characteristics of the transformer and the influence of the real-time environment, providing a key basis for distinguishing between load heating and abnormal loss heating.

[0028] The calculation of the theoretical cold element temperature for each part of the transformer includes: S21, Based on the real-time ambient temperature, the reference heat dissipation structure vector in the thermal inertial fingerprint is remapped according to the temperature dependence of the material's thermal conductivity to generate a dynamic thermal conductivity matrix. The specific operation is as follows: The purpose of generating a dynamic thermal conductivity matrix through parameter remapping is to correct the influence of real-time ambient temperature on the thermal conductivity of materials, making the baseline heat dissipation structure vector adapt to the current environmental conditions and ensuring the accuracy of the heat conduction model. The baseline heat dissipation structure vector in the thermal inertial fingerprint is constructed based on the average state of multiple scenarios, and its corresponding material thermal conductivity parameters are adapted to a specific temperature range by default. However, the real-time ambient temperature during transformer operation will change the thermal conductivity of materials in various parts, thus affecting the heat dissipation characteristics. Therefore, the parameters need to be adjusted according to the temperature dependence of material thermal conductivity. The temperature dependence of thermal conductivity varies among different materials. The thermal conductivity of metal components decreases linearly with increasing temperature, while that of insulating materials decreases linearly. The thermal conductivity increases slightly with increasing temperature. It is necessary to obtain the thermal conductivity temperature coefficient of each part of the material in advance through material experiments or by consulting material handbooks, and establish a model of the correspondence between thermal conductivity and temperature. Based on this model, with the real-time ambient temperature as input, the parameters of the reference heat dissipation structure vector in the thermal inertia fingerprint are remapped element by element. The thermal conductivity parameters under the reference state are corrected to the actual thermal conductivity parameters corresponding to the current ambient temperature. After correction, the corrected parameters are arranged into a matrix according to the correspondence between the heat conduction path and heat dissipation node of each part of the transformer to generate a dynamic thermal conductivity matrix. Each element in the matrix corresponds to the thermal conductivity of a specific part and a specific heat dissipation path at the current ambient temperature.

[0029] S22, based on the dynamic thermal conductivity matrix and the initial state of the transformer's waste heat distribution, constructs and solves the heat flow balance equation, and derives the heat dissipation dynamics. The specific operation is as follows: The initial state of waste heat distribution during transformer operation needs to be determined by combining historical operating data and real-time monitoring data. Specifically, by using the actual temperature data, load change curves, and heat dissipation records of each part in the previous acquisition cycle, the inverse heat conduction problem solution method is used to back-calculate the current state of waste heat distribution at each part, including the total amount of waste heat, distribution density, and transfer trend, to ensure that the initial state is consistent with the actual operating conditions. The heat flow balance equation is constructed based on Fourier's law of heat conduction and the law of conservation of energy, using the dynamic thermal conductivity matrix as the core parameter, incorporating the heat capacity parameters of each part, the initial state of waste heat distribution, and the real-time ambient temperature boundary conditions. The left side of the equation represents the difference between the heat flow input and output at each part, and the right side represents... The thermal energy changes caused by the temperature changes of the parts themselves are considered. Given the complex structure of the transformer and the multi-path, multi-node coupling characteristics of heat flow transfer, the finite difference method is used to solve the heat flow balance equation. The solution time step is set to 1 minute, which is consistent with the temperature acquisition frequency. The temperature change, heat flow transfer rate and residual heat decay of each part are calculated step by step. During the solution process, the rationality of the calculation results needs to be verified in real time to ensure that the direction and rate of heat flow transfer in each part conforms to physical laws. The calculation error is controlled within ±0.5℃. Finally, dynamic heat dissipation data containing the temperature sequence, heat flow distribution sequence and heat dissipation rate sequence of each part over time are obtained, which completely replicates the heat dissipation process under the current environment and initial residual heat conditions.

[0030] To accurately characterize the heat flow balance relationship, a formula is introduced, specifically: dynamic thermal conductivity matrix. With temperature gradient vector The product of these, plus the residual heat distribution density vector , equal to the heat capacity matrix vector of temperature change The product of these factors, the formula is based on the physical laws of energy conservation and heat conduction. Combining dynamic thermal conductivity characteristics and waste heat changes, it fully reflects the dynamic relationship between heat flow and temperature in various parts. Through this formula, a quantitative relationship between thermal conductivity, temperature, waste heat and heat capacity can be accurately established, providing mathematical support for the deduction of heat dissipation dynamics.

[0031] Among them, the dynamic thermal conductivity matrix Its elemental value varies with real-time temperature The change corresponds to the thermal conductivity of the material in each part at the current temperature; Temperature gradient vector , characterizing the rate and direction of temperature change at various points in space; Waste heat distribution density vector This corresponds to the distribution of initial residual heat in each part; Heat capacity matrix It is calculated from the material density, specific heat capacity and volume of each part, and reflects the part's ability to store heat. Temperature change rate vector It represents the temperature change range of each part per unit time.

[0032] S23 extracts the initial temperature and initial cooling rate features from the heat dissipation dynamics, matches them with the standard cooling expectation curve cluster in the thermal inertia fingerprint, and outputs the theoretical cold body temperature through interpolation calculation. The specific operation is as follows: First, the initial temperature and initial cooling rate features are extracted from the heat dissipation dynamic data obtained in step S22. The initial temperature is selected from the temperature values ​​of various parts at the start of the heat dissipation dynamic simulation, which includes the combined effects of initial residual heat and ambient temperature. The initial cooling rate is obtained by calculating the slope of temperature change in the first 5 minutes after the simulation starts. A linear fitting method is used to eliminate random fluctuations and ensure that the features can accurately reflect the inherent characteristics of the initial stage of heat dissipation. After extraction, the two features are matched with the standard cooling expectation curve cluster in the thermal inertial fingerprint. The matching process uses a feature similarity algorithm to calculate the deviation value between the extracted features and the initial features corresponding to each standard curve in the curve cluster. The standard curve with the smallest deviation value is selected as the matching benchmark. The thresholds are set to ensure that the initial temperature deviation does not exceed ±1℃ and the initial cooling rate deviation does not exceed ±0.02℃ / min. If the feature corresponding to the real-time operating condition completely matches a single standard curve, the steady-state temperature corresponding to the standard curve is directly used as the theoretical cold body temperature. If the feature is between two or more standard curves, a quadratic interpolation method is used for calculation. The interpolation variables are the ambient temperature and the initial cooling rate. The operating condition parameters corresponding to the standard curve are used as interpolation nodes. The theoretical cold body temperature that is suitable for the current operating condition is obtained by constructing an interpolation function. After the calculation is completed, the consistency of the theoretical cold body temperature of each part is checked to ensure that the temperature difference of different parts in the same area conforms to the inherent heat distribution law of the transformer. After the check is passed, the final theoretical cold body temperature sequence is output.

[0033] In some embodiments of the present invention, a net temperature rise module is also included, which is used to obtain the actual temperature values ​​of various parts of the transformer, and subtract the corresponding theoretical cold body temperature from the actual temperature values ​​to obtain the net heating element temperature rise. The specific operation is as follows: When the net temperature rise module is running, it needs to acquire two types of key data simultaneously: one is the real-time actual temperature value of each part of the transformer, and the other is the theoretical cold body temperature value of the corresponding part output by the theoretical temperature module. Both types of data must strictly follow the timestamp alignment principle to ensure that each set of calculation objects corresponds to the temperature parameters of the same time and the same part, avoiding calculation errors caused by time differences. The acquisition of actual temperature values ​​relies on temperature sensors pre-deployed in the core part of the transformer. The sensors must be calibrated before use, and the calibration cycle should not exceed 6 months. The deployment location must cover key heat-generating and heat-dissipating components such as windings, iron core, tank walls and heat sinks, and avoid areas that are prone to measurement deviations, such as direct sunlight and local airflow dead corners. The acquisition frequency is consistent with the output frequency of the theoretical cold body temperature, usually set to once per minute, to ensure the timeliness and synchronization of the data. The acquired actual temperature values ​​must be pre-processed to remove abnormal jump data caused by sensor failure and electromagnetic interference. The pre-processing standard is that when the single temperature change exceeds ±5% of the temperature value of the previous cycle, it is judged as abnormal data and supplemented and corrected by linear interpolation to ensure the continuity and reliability of the actual temperature sequence.

[0034] The calculation of net heat generation temperature rise is based on a formula constructed from the relationship of temperature differences. This formula, starting from a physical perspective, directly locates the heat generation increment related to load and anomalies by stripping away the environmental and inherent thermal inertia influences represented by the theoretical cold body temperature. The formula is as follows: This formula is derived from the physical meaning of temperature. The theoretical cooling body temperature excludes the effect of load heating and only reflects the contribution of the environment and inherent thermal characteristics to the temperature. By subtracting the corresponding theoretical cooling body temperature from the actual temperature, the net temperature rise caused by load loss and potential abnormal heating can be accurately separated. This is the net temperature rise vector, corresponding to the net temperature rise value of each part of the transformer. The vector dimension is consistent with the number of temperature acquisition points. This is the actual temperature vector, which contains the real-time actual temperature values ​​of each part after preprocessing. The theoretical cold body temperature vector is taken from the temperature data of the corresponding time and location output by the theoretical temperature module. After the calculation is completed, the rationality of the net heat-generating element temperature rise sequence needs to be verified. The verification basis is the basic correlation law between the net temperature rise of each part of the transformer and the load current. That is, when the load current is stable, the net temperature rise should be kept within a fixed fluctuation range, and the fluctuation range should not exceed ±0.3℃. If it exceeds this range, it is judged as data abnormality. It is necessary to backtrack and check the accuracy of the actual temperature acquisition and the accuracy of the theoretical cold body temperature calculation until the verification is passed. The final output net heat-generating element temperature rise sequence accurately focuses on the temperature changes caused by load loss and abnormal heating.

[0035] In some embodiments of the present invention, the heat loss module is used to establish the correspondence between the net heating element temperature rise and the load current, obtain the heat loss response curve, and calculate the curvature change of the heat loss response curve. The specific operations are as follows: The loss thermal module is an analytical unit for transformer loss diagnosis and health management. Its task is to establish a quantitative correlation between the net heating element temperature rise and the load current. By constructing a loss thermal response curve and analyzing its curvature changes, it captures subtle signs of deterioration in transformer loss characteristics. During operation, the loss thermal module simultaneously receives the net heating element temperature rise sequence output by the net temperature rise module and the real-time load current time series of the transformer. The two types of data must be strictly time-stamp aligned to ensure that each set of correlated data corresponds to the same operating time. Through hysteresis compensation, piecewise fitting, and smoothing fusion, the loss thermal response curve is constructed. This curve essentially reflects the inherent correlation between load current changes and loss heating. Under normal conditions, the curve shape is stable. When the transformer experiences problems such as insulation aging, abnormal core or winding losses, the curve shape will undergo characteristic distortion. Based on the calculation and analysis of the curve curvature change, the implicit deterioration of loss characteristics can be transformed into quantifiable characteristic parameters.

[0036] The construction of the loss thermal response curve includes the following steps: S41, based on the load current time series, performs hysteresis compensation based on the thermal time constant on the net heat-generating element temperature rise series to generate a quasi-steady-state temperature rise series and equivalent steady-state current value. The specific operation is as follows: Firstly, by compensating for the lag in thermal time constant, the temperature rise lag effect caused by thermal inertia is eliminated, ensuring a precise match between the net heat source temperature rise sequence and the load current time sequence. This is because the temperature rise caused by transformer losses has an inherent lag; that is, after a change in load current, the net heat source temperature rise needs a certain amount of time to reach the corresponding steady state. This lag leads to deviations in the original data correlation, affecting the accuracy of curve construction. The thermal time constant needs to be determined based on characteristic parameters in the thermal inertia fingerprint, combined with the heat capacity and heat dissipation characteristics of various parts of the transformer, and obtained through thermal network model simulation or historical data fitting. The thermal time constant varies between different parts; the winding thermal time constant is typically 5 to 10 minutes, while the core and tank thermal time constant is 15 to 20 minutes.

[0037] The compensation process employs an exponential lag correction model. Using the load current time series as a benchmark, the net heating element temperature rise sequence is corrected for time offset point by point. The correction magnitude is determined by the thermal time constant of the corresponding part and the rate of current change. The faster the rate of current change, the greater the lag compensation. After compensation, stable periods with current fluctuations not exceeding ±3% are selected. The average net temperature rise within these periods is taken as the quasi-steady-state temperature rise sequence. Simultaneously, the arithmetic mean of the load current within these periods is calculated as the equivalent steady-state current value, ensuring that each set of equivalent steady-state current values ​​corresponds to a unique quasi-steady-state temperature rise, thus eliminating the interference of instantaneous current fluctuations on the correlation.

[0038] S42, based on the equivalent steady-state current value, the quasi-steady-state temperature rise sequence is segmented to obtain multiple local linear relationship segments. These segments are then coupled based on the physical model of copper loss and iron loss to form the initial response curve. The specific operation is as follows: By coupling segmented fitting with a physical model, an initial response curve that accurately reflects loss characteristics is constructed, taking into account the differences in loss patterns across different current ranges. The equivalent steady-state current value covers the entire operating range of the transformer, including no-load, low-load, rated load, and overload. It needs to be divided into multiple continuous segments based on current magnitude. The segmentation is determined by considering the loss characteristics, typically using no-load current, 30% rated current, 70% rated current, and rated current as dividing points, resulting in four segments. The current variation range within each segment is controlled within ±10% of the rated current. To preserve the monotonicity of loss characteristics within a given section, a least-squares method is used for linear fitting of the equivalent steady-state current value and quasi-steady-state temperature rise sequence for each section. This yields local linear relationship segments for each section, each corresponding to the proportional relationship between temperature rise and current within a specific current range. Since transformer losses consist of copper and iron losses, with iron losses dominating and remaining approximately constant in the no-load and low-load ranges, and copper losses dominating in the medium-to-high-load range as current increases squarely, a single linear relationship cannot cover the loss patterns across the entire range. Therefore, a coupling model based on physical models of copper and iron losses is necessary. The iron loss model uses a constant loss component, while the copper loss model uses a variable component proportional to the square of the current. By substituting each local linear relationship segment into the coupled model, fitting biases are corrected, ensuring smooth transitions in the initial response curves across each range and conforming to the physical laws governing loss composition. This fully reproduces the correlation between temperature rise and loss across the entire current range.

[0039] S43: Calculate the confidence level based on the density and dispersion of data points within each current segment. Smooth the initial response curve based on the confidence level and perform weighted fusion with the historical loss thermal response curve to output the loss thermal response curve. The specific operation is as follows: By using confidence-weighted averaging and historical data fusion, the stability and accuracy of the initial response curve are optimized, eliminating random data interference and ensuring that the curve reflects the inherent loss characteristics of the transformer. Each current segment is a continuous current interval pre-divided in step S42 based on the transformer's loss characteristics. The division is based on the transformer's no-load current, 30% of the rated current, 70% of the rated current, and the rated current as boundary points, dividing the entire operating current range into four segments: no-load segment, low-load segment, medium-load segment, and rated overload segment. The current variation range within each segment is controlled within a certain range. Within ±10% of the rated current, the division is based on the difference in the dominant loss type in different current ranges. In the no-load to low-load range, iron loss dominates and the loss value is approximately constant. In the medium-load to rated overload range, copper loss dominates and the loss value increases with the square of the current. This division method ensures the monotonicity and consistency of loss characteristics within each range, providing a standardized range basis for confidence calculation. Confidence calculation is based on the data point density and dispersion within each current range. Data point density is the number of equivalent steady-state current-temperature rise data pairs within that range; the more pairs, the higher the density. Dispersion... The standard deviation of the temperature rise data within each segment is used for characterization; a smaller standard deviation indicates lower dispersion. The confidence score calculation formula integrates density and dispersion into a value between 0 and 1 through normalization. Higher density and lower dispersion result in a confidence score closer to 1, and vice versa. The initial response curve is smoothed based on the confidence score. For segments with a confidence score higher than 0.8, the original fitting results are retained; for segments with a confidence score between 0.5 and 0.8, Gaussian filtering is used for smoothing, with the filter window size inversely proportional to the confidence score; for segments with a confidence score lower than 0.5, the data is discarded. The original data is discarded, and interpolation is performed based on the curve trends of adjacent high-confidence segments. Then, it is weighted and fused with the historical loss thermal response curves. The weights of the historical curves are dynamically allocated according to the time period: the weight of curves within the last 3 months is 0.6, the weight of curves within 3-12 months is 0.3, and the weight of curves more than 12 months is 0.1. The random deviation of the current curve is corrected by weighted averaging. After the fusion is completed, the smoothness of the curve is checked to ensure that the slope changes of adjacent segments are continuous and there are no sudden inflection points. After the check is passed, the final loss thermal response curve is output.

[0040] The calculation of the curvature change of the loss thermal response curve includes the following steps: S44. Select a reference current point and calculate the morphological deviation of the loss thermal response curve from the theoretical quadratic function basis within the preset current window at the reference current point to obtain the primary curvature characteristics. The specific operation is as follows: By comparing with the theoretical quadratic function basis, primary curvature features are extracted to capture subtle morphological changes in the loss thermal response curve. The selection of reference current points needs to cover the main operating range of the transformer to ensure a comprehensive reflection of the curve morphology. Typically, four reference points are selected: no-load current point, 50% rated current point, rated current point, and 110% rated current point, each corresponding to a key operating segment of the curve. The preset current window is centered on the reference current point, and the window width is dynamically adjusted according to the reference current magnitude. The window width is ±5% of the rated current in the low-load range and ±8% of the rated current in the medium-to-high-load range, ensuring sufficient data points within the window to support analysis while avoiding morphological interference caused by crossing different segments. The theoretical quadratic function basis is constructed based on the normal loss characteristics of the transformer. Copper loss is proportional to the square of the current, while iron loss is constant. Therefore, the loss thermal response curve under normal conditions approximately conforms to the quadratic function law. The theoretical basis function form is set as follows: Where a is the copper loss correlation coefficient and b is the iron loss constant component, the sum of squared residuals between each data point of the loss thermal response curve within the preset current window and the theoretical quadratic function basis is calculated, and the sum of squared residuals is used as the morphological deviation. The larger the deviation, the more significant the difference between the curve morphology and the normal state. This deviation is the primary curvature characteristic, and each reference current point corresponds to a primary curvature characteristic value.

[0041] S45. Based on the primary curvature characteristics, N detection segments are selected on the loss thermal response curve, and the rate of change of slope with current in each detection segment is extracted to form a multidimensional deformation feature vector. The specific operation is as follows: The detection segments are divided based on primary curvature features, and the slope change rate is extracted to construct a multi-dimensional deformation feature vector, comprehensively representing the overall deformation state of the curve. The selection of the N value needs to balance feature comprehensiveness and computational efficiency, and is usually set to 6 to 8 detection segments, covering the full current range from no-load to overload. The segment division adopts an adaptive method, with denser segment division in areas with larger primary curvature feature values ​​and wider segment division in areas with smaller feature values, ensuring that abnormal deformation areas are captured in a focused manner. The length of each detection segment is controlled within the range of 10% to 15% of the rated current, and adjacent segments have a 5% overlap of the rated current. To avoid missing features, the loss thermal response curve within each detection segment is piecewise linearly fitted to obtain the fitting slope of that segment. Then, the ratio of the difference in fitting slopes between adjacent detection segments to the corresponding current difference is calculated, which is the slope change rate of each detection segment. This parameter can accurately reflect the curvature of the curve at the segment junction. The slope change rates of all detection segments are arranged in ascending order of current to form a multi-dimensional deformation feature vector. The vector dimension is consistent with the number of detection segments N. This vector integrates the deformation information of the curve in different intervals and can comprehensively reflect the overall trend of curve shape change.

[0042] S46 maps the multidimensional deformation feature vector through a pre-constructed nonlinear feature interaction network to generate a comprehensive curvature degradation index. The specific operation is as follows: By using a pre-constructed nonlinear feature interaction network, multidimensional deformation feature vectors are mapped to a comprehensive curvature degradation index, achieving the fusion and quantification of multidimensional features for subsequent trend analysis. The nonlinear feature interaction network is trained and constructed based on historical transformer degradation data. The training dataset contains multidimensional deformation feature vectors and corresponding actual degradation levels under different degradation conditions, such as insulation aging, abnormal core loss, and local overheating of windings. Supervised learning is used for training. The network structure includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is consistent with the dimension of the multidimensional deformation feature vector. The hidden layer has two fully connected layers, with the number of nodes in each layer being 1.5 times the number of nodes in the input layer. The activation function is the ReLU function. The output layer is a single node, and the output range is mapped to the interval between 0 and 1, which is the comprehensive curvature degradation index. The closer the index is to 1, the more significant the curve deformation and the more severe the degradation. The closer it is to 0, the closer the curve shape is to the normal state. During the mapping process, the network integrates the interaction relationship of features in various dimensions through nonlinear transformation to avoid the one-sidedness of a single feature. At the same time, the network parameters are optimized by training with historical data to ensure that the output comprehensive curvature degradation index can accurately correspond to the actual degradation state and eliminate the risk of misjudgment caused by fluctuations in a single feature.

[0043] S47, Perform trend testing and persistence determination on the comprehensive curvature degradation index sequence generated within a preset time window, and output the curvature change amount. The specific operation is as follows: By analyzing trends and determining persistence, stable trends are extracted from the comprehensive curvature degradation index sequence, outputting curvature changes that reflect the actual degradation process. The size of the preset time window is set based on the gradual characteristics of transformer degradation, typically 24 hours, corresponding to 1440 comprehensive curvature degradation index data points collected minute by minute. The window sliding step is 1 hour to ensure the capture of short-term trend changes. Trend analysis uses linear regression to linearly fit the comprehensive curvature degradation index sequence within the time window to the time variable, obtaining the trend slope. A positive slope indicates a continuous increase in degradation. The slope is negative, indicating that the curve shape is recovering, while a slope close to zero indicates a stable state. Continuity is determined through analysis of variance (ANOVA). The variance of the exponential sequence within the time window is calculated. If the variance is less than a preset threshold (usually 0.001) and the absolute value of the trend slope is greater than 0.0005, the trend is considered continuous, excluding false trends caused by random fluctuations. Based on the continuous trend slope and the mean exponential value within the window, the curvature change is calculated. The curvature change is the product of the trend slope and the mean exponential value within the window, reflecting both the rate of trend change and the current level of degradation, thus accurately quantifying the degree of intensification or deceleration of the degradation process.

[0044] In some embodiments of the present invention, an early warning judgment module is further included, which is used to generate a health stress coefficient based on the curvature change and trigger an early warning based on the changing trend of the health stress coefficient. The specific operation is as follows: The early warning judgment module transforms the curvature change output by the loss heat module into a quantified health assessment indicator. By analyzing the trend of indicator changes, it triggers gradient early warnings, providing precise guidance for operation and maintenance. During operation, the early warning judgment module simultaneously receives the curvature change sequence and the transformer's real-time operating parameters. Relying on standardized feature conversion, trend analysis, and verification processes, it transforms the implicit signals of loss degradation into actionable early warning commands. This avoids false alarms caused by single-parameter judgments and accurately captures the development trend of gradual degradation. Its working logic runs through the entire process of health stress quantification, trend verification, and early warning generation, ensuring the reliability of early warnings through multi-dimensional evaluation.

[0045] The generation of the health stress coefficient specifically includes the following steps: S51, based on the current curvature change and the historical curvature change sequence, calculate the standard score of the current curvature change. The specific operation is as follows: By calculating the standard score of the current curvature change, the influence of historical data fluctuations and dimensions is eliminated, ensuring that the curvature change has a basis for horizontal comparability. The current curvature change needs to be correlated with historical curvature change sequences. The historical sequence selection must cover the normal operating cycle of the transformer, typically selecting curvature change data continuously collected within the last 30 days. Abnormal data caused by sudden load shocks or temporary sensor malfunctions are eliminated to ensure the sequence reflects the normal fluctuation range. The standard score is calculated based on the principle of statistical standardization, eliminating data distribution differences through the mean and standard deviation. The formula is: Standard score of current curvature change. Equal to the measured value of the current curvature change Subtract the arithmetic mean of the historical curvature change series Then divide by the standard deviation of the historical series. This formula is abstracted from the statistical characteristics of the normal distribution. It transforms the curvature change of any distribution into standard normal distribution data with a mean of 0 and a standard deviation of 1, which facilitates a unified assessment of the degree of deviation. After the calculation is completed, the standard score needs to be checked for reasonableness to ensure that it is within the range of [-3,3]. If it exceeds this range, it is judged as an extreme anomaly, which directly triggers the subsequent emergency verification process to avoid abnormal data interfering with subsequent analysis.

[0046] S52, the standard score is mapped through a nonlinear transformation function based on the Arrhenius model to obtain the equivalent aging acceleration factor. The specific operation is as follows: By employing a nonlinear transformation based on the Arrhenius model, the standard score is mapped to an equivalent aging acceleration factor, quantifying the accelerating effect of curvature change on the transformer aging process. The Arrhenius model, originally used to describe the effect of temperature on chemical reaction rates, is adapted here to suit the transformer degradation characteristics. Since the curvature change reflects the degree of loss degradation, a larger standard score indicates more significant degradation and a stronger aging acceleration effect. Therefore, a nonlinear transformation establishes the correlation between the two. The transformation function introduces a standard score correction term based on the Arrhenius model, and the formula is: Equivalent Aging Acceleration Factor. =Benchmark Aging Factor The formula exp(k×z), with the natural constant e as the base and the product of the degradation sensitivity coefficient k and the standard score z as the exponent, strengthens the influence of the standard score on the aging rate through the exponential function. This results in a gradual acceleration factor for slight degradation and a rapidly increasing acceleration factor for significant degradation, aligning with the nonlinear characteristics of transformer degradation. In the formula... The aging rate under healthy conditions is taken, with a default value of 1; k is determined based on the characteristics of the transformer material, with insulation materials ranging from 0.8 to 1.2 and core materials ranging from 0.6 to 0.9; z is the standard fraction calculated in step S51; during the conversion process, the real-time operating temperature of the transformer needs to be referenced simultaneously, and the sensitivity coefficient k is finely adjusted. For every 5°C that the temperature exceeds the rated operating temperature, the value of k is increased by 0.1 to ensure that the aging factor can adapt to the deterioration acceleration law under the synergistic effect of temperature.

[0047] S53, multiply the equivalent aging acceleration factor by the differential of the current load rate and running time and integrate it to generate the cumulative health stress value as the health stress coefficient. The specific operation is as follows: The cumulative health stress value is generated through integral calculation, transforming the instantaneous aging acceleration effect into a long-term cumulative health loss indicator, accurately characterizing the gradual deterioration degree of the transformer. The equivalent aging acceleration factor only reflects the instantaneous deterioration rate, while transformer health deterioration is a long-term cumulative process. Therefore, it needs to be combined with the current load rate and operating time to quantify the cumulative impact through integral calculation. The current load rate is obtained by collecting real-time transformer operating data, taking the average load rate within the integration period. The higher the load rate, the more obvious the deterioration acceleration effect, and its weight needs to be strengthened in the integration. The formula for integral calculation is: The cumulative health stress coefficient S is equal to the product of the equivalent aging acceleration factor A and the current average load rate r, starting from the integration start time. The integral up to the current integration time t; this formula is abstracted from the principle of energy accumulation, integrating the aging loss per unit time in the time dimension to obtain the cumulative health stress value. The aging loss per unit time is the product of the acceleration factor and the load rate. In the formula, the cumulative health stress coefficient S is the cumulative health stress value; the current average load rate r ranges from 0 to 1.1, corresponding to no load to 110% of the rated load; the differential element of the integration time is consistent with the data acquisition cycle, which is 1 minute. After the integration is completed, the current cumulative health stress value and the integration time are stored as the benchmark for the next integration. At the same time, invalid time such as load interruption and power outage during the integration period is eliminated to ensure that the integration result only reflects the deterioration accumulation under effective operating conditions.

[0048] The process of triggering an early warning based on the changing trend of the health stress coefficient includes the following steps: S54, perform local linear fitting based on an adaptive sliding window on the health stress coefficient sequence to extract the slope and confidence interval of the current trend. The specific operation is as follows: An adaptive sliding window fitting method is used to extract trend features from the health stress coefficient sequence, providing dynamic trend evidence for early warning assessment while balancing the timeliness and stability of trend capture. The size of the adaptive sliding window is not fixed and needs to be dynamically adjusted according to the fluctuation of the health stress coefficient. The larger the fluctuation amplitude, the smaller the window width to quickly capture sudden trends; the smaller the fluctuation amplitude, the larger the window width to filter out interference from occasional fluctuations. The window width is set to an adjustment range of 1 to 3 hours, corresponding to 60 to 180 health stress coefficient data points. The adjustment is based on the sum of squared residuals of the fitted curve in the previous window; a larger residual indicates greater fluctuation. The more severe the stress, the smaller the window width is (0.5 hours) and vice versa, until the set range boundary is reached. The fitting process uses local linear regression to fit the health stress coefficient sequence and time variable within the window, obtaining the slope of the fitted line. A positive slope indicates that the health stress continues to rise and the deterioration process intensifies; a negative slope indicates that the stress tends to stabilize or ease. At the same time, the confidence interval of the fitting result is calculated, with the confidence level set at 95%. The narrower the confidence interval, the more reliable the trend; the wider the interval, the greater the data fluctuation and the lower the trend reliability. The extracted trend slope and confidence interval together constitute the trend feature.

[0049] S55, combining the slope of the current trend, confidence interval, and absolute level of the health stress coefficient, performs a joint state assessment through a multi-level early warning rule table to generate a preliminary early warning level. The specific operation is as follows: A preliminary early warning level is generated through multi-dimensional joint assessment to avoid false alarms or missed alarms caused by a single indicator, ensuring the accuracy of the early warning. The joint assessment needs to integrate three indicators: trend slope, confidence interval, and absolute level of health stress coefficient. The trend slope reflects the rate of deterioration, the confidence interval reflects the reliability of the trend, and the absolute level reflects the current degree of deterioration accumulation. The three work together to form a complete assessment system. The assessment process relies on preset multi-level early warning rules. The rules are formulated based on historical transformer deterioration data and operation and maintenance experience, without the need for additional creative settings. When the absolute level of health stress coefficient is lower than the preset low threshold, and the trend slope is close to zero and the confidence interval is narrow, the warning level is triggered. The following conditions are considered for warning: no warning is given. When the absolute level is between the low and medium thresholds, and the trend slope is positive but the absolute value is small and the confidence interval is narrow, a Level 1 warning is given, which is slight degradation. When the absolute level is between the medium and high thresholds, or the trend slope is significantly positive and the confidence interval is narrow, a Level 2 warning is given, which is moderate degradation. When the absolute level exceeds the high threshold, or the trend slope rises sharply and the confidence interval is wide but continuously positive, a Level 3 warning is given, which is severe degradation. Each threshold needs to be set according to the transformer model, operating years and material characteristics. The high threshold for newly commissioned transformers can be set to 1.2 times that of old transformers to ensure that the threshold is adapted to the actual condition of the equipment.

[0050] S56, based on the initial warning level, dynamically adjust the frequency of subsequent data collection and calculation, and when the preset level is reached, initiate multi-cycle backtracking verification and consistency checks to confirm the warning. The specific operations are as follows: By dynamically adjusting the data collection frequency and conducting multi-cycle backtracking verification, the effectiveness of early warnings is confirmed, false warnings caused by occasional fluctuations are filtered out, and the response speed to actual degradation is improved. The subsequent data collection and calculation frequency is adjusted according to the initial warning level. When there is no warning, the original frequency of 1 minute / time is maintained; for a Level 1 warning, the frequency is increased to 30 seconds / time to enhance data capture density; for Level 2 and above warnings, the frequency is increased to 15 seconds / time to track the degradation trend in real time. When the initial warning level reaches Level 2 or above, multi-cycle backtracking verification is initiated, with the verification cycle set to 3 consecutive cycles. The duration of each cycle is correlated with the current collection frequency. The data collection frequency is 15 seconds, with each cycle lasting 5 minutes to ensure sufficient verification data. Backtracking verification compares the trend slope, confidence interval, and health stress coefficient changes over three consecutive cycles. If the trend remains consistent across the three cycles, showing a worsening trend, and the health stress coefficient continues to rise while the confidence interval steadily narrows, the warning is confirmed as valid. If the trend reverses, data fluctuates drastically, or the confidence interval widens significantly, it is determined to be a false warning. The initial warning is then revoked, and the original collection frequency is restored. Data from each cycle is recorded simultaneously during the verification process as a traceability basis for warning confirmation, ensuring that maintenance personnel can review the verification logic.

[0051] S57 generates a tiered warning command based on the confirmed warning level, including the warning level, potential fault mode, suggested response window, and maintenance priority. The specific operation is as follows: Based on the confirmed warning levels, tiered warning instructions are generated to provide clear and actionable execution guidelines for operation and maintenance work, achieving precise integration between warnings and operation and maintenance. Warning instructions must include four components to ensure completeness and operability. The warning level directly follows the confirmed level, clearly defining the severity of degradation. Potential fault modes are determined based on the characteristics of changes in the health stress coefficient and correlation with historical fault data. Level 1 warnings often correspond to minor insulation aging or small fluctuations in core loss; Level 2 warnings correspond to accelerated insulation aging or abnormal local winding losses; and Level 3 warnings correspond to serious faults such as winding overheating and accelerated insulation degradation. It is recommended that response windows be set according to the warning level: 72 hours for Level 1, 24 hours for Level 2, and 4 hours for Level 3, clearly defining the time requirements for operation and maintenance. Maintenance priorities correspond to warning levels: Level 3 warnings are the highest priority, requiring immediate on-site verification by operation and maintenance personnel; Level 2 warnings are medium priority, requiring priority handling within the response window; and Level 1 warnings are low priority, included in the routine operation and maintenance plan. Once generated, the command is simultaneously pushed to the operation and maintenance management platform, along with a health stress coefficient sequence, trend fitting curve, and verification data. This allows operation and maintenance personnel to quickly grasp the fault status and formulate targeted handling plans. At the same time, the command generation time, content, and push recipients are recorded, forming the initial node of the operation and maintenance closed-loop management.

[0052] In some embodiments of the present invention, a degradation location module is also included, which is used to inject a current pulse into the transformer winding after the warning is triggered, record the temperature rise response curve of each part of the transformer to the current pulse, and determine the degradation location by analyzing the difference between the temperature rise response curve and the thermal inertia fingerprint. The specific operation is as follows: After the early warning judgment module triggers an alert, the degradation location module pinpoints the specific location of degradation within the transformer, providing maintenance personnel with targeted handling information and overcoming the technical limitation of only providing early warnings without location information. Relying on pre-stored health baseline data from the thermal fingerprint module, the degradation location module injects current pulses with specific parameters into the transformer windings to capture the temperature rise response characteristics of various parts. This data is then compared and analyzed with the health baseline to quantify abnormal differences and map them to specific locations. The injection of current pulses and the acquisition of response curves must be precisely synchronized to ensure that temperature rise changes accurately reflect the inherent characteristics of the parts. Through multi-dimensional data fusion and decision analysis, the module achieves accurate determination of degradation locations and generates verification sequences, providing clear guidance for subsequent maintenance checks and promoting a shift in maintenance work from comprehensive investigation to precise location.

[0053] The determination of the location of the deterioration zone includes the following steps: S61, extract the multi-time constant spectrum from the temperature rise response curve. The specific operation is as follows: By extracting the multi-time constant spectrum of the temperature rise response curve, the differences in the thermal response characteristics of different parts to the current pulse are captured. After the current pulse is injected, the temperature rise response speed varies significantly among different parts of the transformer due to differences in heat capacity, thermal conductivity, and heat dissipation path. Parts with small heat capacity, such as windings, experience rapid temperature rise and short steady-state time, while parts with large heat capacity, such as the tank and core, experience slow temperature rise and long steady-state time. This difference can be quantified and characterized by time constants. The extraction process requires preprocessing the temperature rise response curve to remove electromagnetic interference spikes at the moment of pulse injection and retain the complete curve segment from pulse injection to temperature steady state. The condition determination criterion is that the temperature change within 5 consecutive minutes does not exceed ±0.1℃. The preprocessed curve is decomposed using a multi-exponential fitting method. The fitting function is a superposition of multiple exponential functions, with each exponential term corresponding to a time constant, which respectively characterizes the thermal response characteristics of different parts. The fitting process uses the least squares method for iterative optimization to ensure that the deviation between the fitted curve and the measured curve is controlled within ±0.2℃. After the fitting is completed, the time constants corresponding to each exponential term are extracted and sorted by value from smallest to largest to form a multi-time constant spectrum. Each time constant in the spectrum corresponds to the thermal response characteristics of a specific part of the transformer.

[0054] S62, the multi-time constant spectrum is differentially analyzed with the healthy baseline time constant spectrum stored in the thermal inertial fingerprint to generate an abnormal time constant offset vector. The specific operation is as follows: By calculating the difference between the multi-time constant spectrum and the healthy reference time constant spectrum, an abnormal time constant offset vector is generated to quantify the degree of deviation between the current thermal response characteristics and the healthy state. The healthy reference time constant spectrum is stored in the thermal inertial fingerprint and is a multi-time constant spectrum extracted after injecting a current pulse with the same parameters when the transformer is in a healthy state. It has a reference benchmark that is completely consistent with the current detection, ensuring the effectiveness of the difference comparison. The difference calculation adopts a corresponding subtraction method element by element. Taking the healthy reference time constant spectrum as the benchmark, the time constant of the corresponding position in the current multi-time constant spectrum is subtracted from the benchmark value to obtain the offset of each time constant. In order to avoid false offsets caused by small fluctuations, the offset needs to be threshold filtered. The offset threshold is set to ±5% of the corresponding value of the healthy reference time constant. Offsets exceeding this threshold are retained, and those not exceeding it are zeroed out to eliminate the interference of measurement error and small environmental fluctuations. The filtered offsets are combined in the order of time constant to form an abnormal time constant offset vector. The vector dimension is consistent with the number of time constants in the multi-time constant spectrum, and each element accurately corresponds to the degree of abnormality of the thermal response characteristics of a specific part.

[0055] S63, Project the abnormal time constant offset vector onto the preset fault location mapping matrix, and calculate the confidence score of each candidate deteriorated location. The specific operation is as follows: By projecting the abnormal offset vector onto the fault location mapping matrix, the confidence score of each candidate location is calculated, thus mapping abnormal features to specific locations. The fault location mapping matrix is ​​a standardized matrix pre-constructed based on transformer structural characteristics and historical fault data. The matrix rows correspond to the elements of the abnormal time constant offset vector, i.e., the offset of each time constant, while the columns correspond to each candidate deteriorated location of the transformer, such as windings, core, tank, and insulation layer. Matrix elements are weight coefficients, representing the degree of correlation between the corresponding time constant offset and the deterioration of that location. The higher the correlation, the larger the absolute value of the weight coefficient. The weight coefficients are obtained through training and optimization using a large number of historical fault cases and have a clear physical correspondence. The projection calculation uses a weighted inner product operation of vector and matrix, and the formula is the confidence score vector of each candidate location. =Abnormal time constant offset vector Multiply by the fault location mapping matrix This formula is abstracted from the feature mapping logic, transforming multidimensional abnormal offset features into confidence scores for each part, realizing the part-based association of abnormal features; after the calculation is completed, the score vector is normalized, mapping the location confidence scores of each part to the interval between 0 and 1, and the higher the score, the greater the possibility of deterioration in that part.

[0056] S64 integrates confidence scores, warning levels, and historical diagnostic records of transformers, and generates location judgment conclusions and recommended verification sequences through a decision tree. The specific operation is as follows: By integrating multi-dimensional information and relying on decision trees to generate location judgment conclusions and recommended verification sequences, the reliability of the location results and the operability of operation and maintenance are ensured. The three integrated information items each have their own emphasis: the confidence score directly reflects the probability of location degradation, the warning level characterizes the severity of degradation, and historical diagnostic records provide reference for similar faults. The collaboration of the three can avoid misjudgments caused by relying on a single piece of information. The decision tree is pre-trained and constructed based on historical transformer fault data. The root node is the warning level, the branch nodes are the confidence score interval and the frequency of historical faults, respectively, and the leaf nodes are the final location judgment conclusions. Each branch has a clearly defined judgment threshold, eliminating the need for additional creative settings. For example, under a Level 3 warning, if the confidence score is higher than 0.8 and the historical failure frequency of this part is ≥3 times, it is directly judged as deteriorated. When the confidence score is between 0.5 and 0.8, the judgment is made in combination with the scores of adjacent parts. After the location judgment conclusion is generated, it is sorted from high to low confidence score to generate a recommended verification sequence. Each sequence node corresponds to a candidate part and a matching verification method. Parts with high confidence scores are given priority to use precise detection methods, such as winding local resistance test and core insulation resistance test. Parts with low confidence scores are used to use conventional detection methods, such as infrared thermal imaging scanning. This provides maintenance personnel with a clear verification sequence and technical guidance, and improves the efficiency of fault handling.

[0057] Additionally, it should be noted that during the construction of the thermal inertial fingerprint, the following steps are performed simultaneously: In the transformer's healthy state, inject the same current pulses as during fault diagnosis into its windings, and record the temperature rise response curves of each part; extract the multi-time constant spectrum from the temperature rise response curves; and store the multi-time constant spectrum as the health baseline time constant spectrum in the thermal inertial fingerprint. The detailed operations are as follows: When a transformer is in a clearly healthy state, the health status determination needs to be combined with the factory inspection report, recent operation and maintenance records, and real-time operating parameters to ensure that the transformer has no problems such as insulation aging, abnormal losses, or local overheating. It is usually carried out within 3 months after the new operation or after a comprehensive overhaul and stable operation for 1 month. At this time, the thermal characteristics of each part of the transformer are in an inherent stable state, and the interference of previous deterioration can be eliminated. The entire synchronous process is carried out in parallel with the standard cooling expectation curve cluster generation stage of thermal inertial fingerprint construction. Finally, the health reference time constant spectrum is incorporated into the thermal inertial fingerprint dataset to form a complete reference system that includes heat dissipation characteristics, cooling laws, and thermal response characteristics.

[0058] The injection of current pulses must strictly match the parameter settings of the fault diagnosis stage to ensure that the excitation conditions of the two tests are consistent and to avoid distortion of thermal response characteristics caused by parameter differences. The pulse parameters, including amplitude, frequency, duration, and injection method, must be solidified and stored in the system in advance as a unified standard for subsequent fault diagnosis. The amplitude setting must take into account both the detectability of temperature rise and the safety of the equipment, and is usually taken as 5% to 10% of the transformer's rated current. This can generate a distinguishable temperature rise response in each part without causing local overheating of the winding or insulation damage due to excessive current. The frequency is set to power frequency or low frequency pulse, which is completely consistent with the frequency used in fault diagnosis. Generally, a 50Hz power frequency pulse is selected to match the inherent electrical characteristics of the transformer winding. The duration is controlled at 3 to 5 seconds to ensure that the pulse energy is sufficient to excite the thermal response of each part, while avoiding the accumulation of residual heat that affects subsequent curve acquisition. The injection method adopts phase-by-phase injection, injecting current pulses into each phase winding of the transformer in sequence. During the injection process, the winding voltage, current, and insulation status are monitored simultaneously to ensure that the injection process is stable and without abnormalities, and without risks such as arcing or insulation breakdown.

[0059] The recording of the temperature rise response curve must be precisely synchronized with the current pulse injection. This relies on temperature sensors already deployed throughout the transformer to achieve full-dimensional data acquisition. The sensor deployment locations must be identical to the temperature acquisition locations used in constructing the thermal inertial fingerprint, covering key nodes such as windings, core, tank walls, heat sinks, and insulation layers. This ensures the response curve comprehensively reflects the thermal characteristics of each component. The acquisition frequency must be higher than the conventional temperature acquisition frequency, set to 100Hz, to capture the complete dynamic process from the moment of pulse injection to the steady-state of the temperature rise, avoiding the omission of high-frequency thermal response characteristics. The acquisition duration starts from the pulse injection and continues until the temperature of each component returns to its initial pre-injection temperature and stabilizes for at least 5 minutes, ensuring a complete record of the entire cycle of temperature rise, peak value, and natural cooling. During the recording process, auxiliary parameters such as ambient temperature and atmospheric humidity must be collected simultaneously to correct for minor interference from environmental factors on the temperature rise response. The environmental parameter acquisition frequency must be consistent with the temperature acquisition frequency to ensure accurate timestamp alignment.

[0060] The extraction method for the multi-time constant spectrum is completely consistent with that used in the degradation localization stage, ensuring the consistency of characteristics between the baseline spectrum and subsequent detection spectra, and providing a unified basis for difference comparison. Before extraction, the acquired temperature rise response curves are preprocessed to remove electromagnetic interference spikes at the moment of pulse injection and accidental fluctuations in sensor data. Peak removal uses a moving average filtering method with a filter window width of 5 data points to eliminate interference without distorting the true temperature rise trend. After preprocessing, a multi-exponential fitting method is used to decompose the curves. The fitting function is a superposition of multiple exponential terms, each corresponding to a time constant, representing the thermal response rate of different parts. The fitting process uses the least squares iterative optimization method, and the iteration terminates when the sum of squared residuals between the fitted curve and the measured curve is less than a preset threshold, which is set to 0.01℃. 2 To ensure fitting accuracy, after fitting, the time constants corresponding to each exponential term are extracted and sorted by value from smallest to largest to form a multi-time constant spectrum. Each time constant in the spectrum corresponds one-to-one with the heat capacity, thermal conductivity and heat dissipation path of a specific part.

[0061] The health baseline time constant spectrum needs to be stored in the thermal inertial fingerprint in a standardized format, and stored in association with the previously constructed baseline heat dissipation structure vector and standard cooling expected curve cluster to form a complete thermal characteristic baseline dataset. When storing, metadata such as construction time, environmental parameters, current pulse parameters, and health status judgment basis should be attached to facilitate the traceability of baseline validity during subsequent operation and maintenance. When the transformer undergoes major overhaul, winding replacement, or other structural modifications, this synchronization process needs to be re-executed to update the health baseline time constant spectrum to ensure that the baseline matches the current actual state of the equipment. The storage format adopts a matrix form, with matrix rows corresponding to each time constant and columns labeled with the corresponding feature association parts and extraction parameters. This provides a standardized data interface for differential calculation and vector projection in the subsequent degradation location stage, enabling efficient access and accurate comparison of baseline data.

[0062] Example 2: Please see Figure 2 Based on Example 1, this example provides a method for transformer loss diagnosis and health management, specifically including the following steps: Step 1: Record the temperature drop curves of various parts of the transformer after the planned power outage, and obtain the thermal inertia fingerprint of the transformer based on the fitting of multiple temperature drop curves. Step 2: During transformer operation, calculate the theoretical cold body temperature of each part of the transformer based on the real-time ambient temperature and thermal inertia fingerprint. Step 3: Obtain the actual temperature values ​​of each part of the transformer, and subtract the corresponding theoretical cold body temperature from the actual temperature values ​​to obtain the net temperature rise of the heat-generating body. Step 4: Establish the correspondence between the net heating element temperature rise and the load current, obtain the loss thermal response curve, and calculate the curvature change of the loss thermal response curve. Step 5: Generate the health stress coefficient based on the curvature change, and trigger an early warning based on the trend of the health stress coefficient change. Step 6: After the warning is triggered, inject a current pulse into the transformer winding and record the temperature rise response curve of each part of the transformer to the current pulse. By analyzing the difference between the temperature rise response curve and the thermal inertia fingerprint, the deterioration area is determined.

[0063] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A transformer loss diagnosis and health management system, characterized in that, include: The thermal fingerprint module is used to record the temperature drop curves of various parts of the transformer after a planned power outage, and obtains the thermal inertia fingerprint of the transformer based on the fitting of multiple temperature drop curves. The theoretical temperature module is used to calculate the theoretical cold body temperature of various parts of the transformer based on the real-time ambient temperature and thermal inertia fingerprint during transformer operation. The net temperature rise module is used to obtain the actual temperature values ​​of various parts of the transformer, and subtract the corresponding theoretical cold body temperature from the actual temperature value to obtain the net temperature rise of the heating element. The heat loss module is used to establish the relationship between the net heating element temperature rise and the load current, obtain the heat loss response curve, and calculate the curvature change of the heat loss response curve. The early warning judgment module is used to generate a health stress coefficient based on the curvature change and trigger an early warning based on the trend of the health stress coefficient change. The degradation location module is used to inject current pulses into the transformer windings after the warning is triggered, record the temperature rise response curves of various parts of the transformer to the current pulses, and determine the degradation location by analyzing the difference between the temperature rise response curves and the thermal inertia fingerprint.

2. The transformer loss diagnosis and health management system according to claim 1, characterized in that, The construction of thermal inertial fingerprints includes: For each planned power outage, the cooling process is analyzed by using the temperature drop curve reflected in the ambient temperature record to inversely deduce the dynamic heat dissipation coefficient set. Clustering and normalizing multiple sets of dynamic heat dissipation coefficients yields a baseline heat dissipation structure vector; Based on the baseline heat dissipation structure vector, a standard cooling expected curve cluster is generated through thermal network model simulation. Feature sets are extracted from the standard cooling expectation curve cluster to form a thermal inertial fingerprint.

3. The transformer loss diagnosis and health management system according to claim 2, characterized in that, Calculate the theoretical cold element temperature of each part of the transformer, including: Based on the real-time ambient temperature, the parameters of the reference heat dissipation structure vector in the thermal inertial fingerprint are remapped according to the temperature dependence of the material's thermal conductivity to generate a dynamic thermal conductivity matrix. Based on the dynamic thermal conductivity matrix and the initial state of the residual heat distribution of the transformer, the heat flow balance equation is constructed and solved to deduce the heat dissipation dynamics. The initial temperature and initial cooling rate features are extracted from the heat dissipation dynamics, matched with the standard cooling expectation curve cluster in the thermal inertial fingerprint, and the theoretical cold body temperature is output by interpolation calculation.

4. The transformer loss diagnosis and health management system according to claim 3, characterized in that, The construction of the loss thermal response curve includes: Based on the load current time series, the temperature rise sequence of the net heating element is compensated for hysteresis based on the thermal time constant to generate a quasi-steady-state temperature rise sequence and equivalent steady-state current value. The equivalent steady-state current value is used to fit the quasi-steady-state temperature rise sequence into segments to obtain multiple local linear relationship segments. Based on the physical model of copper loss and iron loss, the multiple local linear relationship segments are coupled to form the initial response curve. The confidence level is calculated based on the density and dispersion of data points in each current segment. The initial response curve is smoothed based on the confidence level and then weighted and fused with the historical loss thermal response curve to output the loss thermal response curve.

5. The transformer loss diagnosis and health management system according to claim 4, characterized in that, Calculate the change in curvature of the loss thermal response curve, including: By selecting a reference current point, the morphological deviation of the loss thermal response curve from the theoretical quadratic function basis within the preset current window of the reference current point is calculated to obtain the primary curvature characteristics. Based on the primary curvature characteristics, N detection segments are selected on the loss thermal response curve, and the rate of change of slope with current in each detection segment is extracted to form a multidimensional deformation feature vector. The multidimensional deformation feature vectors are mapped through a pre-constructed nonlinear feature interaction network to generate a comprehensive curvature degradation index. The trend test and persistence determination are performed on the comprehensive curvature degradation index sequence generated within a preset time window, and the curvature change is output.

6. The transformer loss diagnosis and health management system according to claim 5, characterized in that, The generation of the health stress coefficient includes: Calculate the standard score of the current curvature change based on the current curvature change sequence and the historical curvature change sequence; The standard score is mapped through a nonlinear transformation function based on the Arrhenius model to obtain the equivalent aging acceleration factor; The equivalent aging acceleration factor is multiplied by the differential of the current load rate and running time and then integrated to generate the cumulative health stress value as the health stress coefficient.

7. The transformer loss diagnosis and health management system according to claim 6, characterized in that, Early warnings are triggered based on the changing trend of the health stress coefficient, including: Local linear fitting based on an adaptive sliding window is performed on the health stress coefficient sequence to extract the slope and confidence interval of the current trend; By combining the slope of the current trend, the confidence interval, and the absolute level of the health stress coefficient, a joint state assessment is conducted through a multi-level early warning rule table to generate a preliminary early warning level. The frequency of subsequent data collection and calculation will be dynamically adjusted based on the initial warning level, and when the preset level is reached, multi-cycle backtracking verification and consistency checks will be initiated to confirm the warning level. Based on the confirmed warning level, a tiered warning instruction is generated, which includes the warning level, potential fault mode, suggested response window, and maintenance priority.

8. A transformer loss diagnosis and health management system according to claim 7, characterized in that, The determination of the location of the deterioration includes: Extracting multi-time-constant spectra from temperature rise response curves; The abnormal time constant offset vector is generated by differentiating the multi-time constant spectrum with the health reference time constant spectrum stored in the thermal inertial fingerprint. The abnormal time constant offset vector is projected onto the preset fault location mapping matrix, and the confidence score of each candidate deterioration location is calculated. By integrating confidence scores, early warning levels, and historical diagnostic records of transformers, a location judgment conclusion and a recommended verification sequence are generated through a decision tree.

9. A transformer loss diagnosis and health management system according to claim 2, characterized in that, The following is also executed simultaneously during the construction of thermal inertial fingerprints: When the transformer is in good condition, inject the same current pulse as during fault diagnosis into its windings and record the temperature rise response curves of each part. Extracting multi-time-constant spectra from temperature rise response curves; The multi-time constant spectrum is used as the health baseline time constant spectrum and stored in the thermal inertial fingerprint.

10. A method for transformer loss diagnosis and health management, applied to a transformer loss diagnosis and health management system according to any one of claims 1-9, characterized in that, Includes the following steps: Step 1: Record the temperature drop curves of various parts of the transformer after the planned power outage, and obtain the thermal inertia fingerprint of the transformer based on the fitting of multiple temperature drop curves. Step 2: During transformer operation, calculate the theoretical cold body temperature of each part of the transformer based on the real-time ambient temperature and thermal inertia fingerprint. Step 3: Obtain the actual temperature values ​​of each part of the transformer, and subtract the corresponding theoretical cold body temperature from the actual temperature values ​​to obtain the net temperature rise of the heat-generating body. Step 4: Establish the correspondence between the net heating element temperature rise and the load current, obtain the loss thermal response curve, and calculate the curvature change of the loss thermal response curve. Step 5: Generate the health stress coefficient based on the curvature change, and trigger an early warning based on the trend of the health stress coefficient change. Step 6: After the warning is triggered, inject a current pulse into the transformer winding and record the temperature rise response curve of each part of the transformer to the current pulse. By analyzing the difference between the temperature rise response curve and the thermal inertia fingerprint, the deterioration area is determined.

Images