Transformer partial discharge fault ai diagnosis method

Abstract

This invention proposes an AI-based diagnostic method for partial discharge faults in transformers, belonging to the field of AI diagnostics. The method includes the following steps: S1: Simultaneously acquiring multi-channel partial discharge raw signals through sensors deployed on the transformer body, and obtaining a time-series sequence containing discharge pulse waveforms, amplitude, and phase information after processing; S2: Constructing a physical feature template library, which includes air gap discharge templates, surface discharge templates, and floating potential discharge templates. All templates have built-in phase distribution intervals, amplitude statistical characteristics, discharge repetition rate variation laws, and pulse sequence time correlation parameters derived from the discharge physical mechanism; S3: Simultaneously inputting the time-series sequence into two parallel branches. The first branch extracts the actual physical feature vector based on the physical feature template library, and the second branch converts the time-series sequence into a phase-resolved map and extracts the depth feature vector through a convolutional neural network.

Description

Technical Field

[0001] This invention belongs to the field of AI diagnostics, and specifically relates to an AI diagnostic method for partial discharge faults in transformers. Background Technology

[0002] Currently, the core bottleneck in the engineering application of AI-based diagnostic methods for transformer partial discharge faults is the lack of physical interpretability of the diagnostic results. Most AI models are end-to-end black-box structures, only mapping signals to fault tags through multi-layer nonlinear transformations. The extracted mathematical features have no clear correlation with the physical mechanism of the discharge. The generation and development of partial discharge follow fixed physical laws; different discharge types correspond to specific physical parameters such as phase distribution, amplitude, and repetition rate. However, existing models only perform high-dimensional numerical calculations and cannot interpret feature weights into a physical description of the insulation state.

[0003] This deficiency leads to three problems on-site: First, the diagnostic conclusions lack traceable evidence and are not credible enough, making it difficult for maintenance personnel to judge the rationality of the reasoning; second, it cannot reflect the location of defects, the degree of deterioration and the development trend, resulting in an information gap between diagnosis and engineering decision-making; and third, the model only learns the statistical correlation of data, rather than the physical causal laws, and its generalization ability drops significantly when the operating conditions change.

[0004] The root cause lies in the disconnect between data-driven AI invention and design and the study of discharge physics mechanisms; physical knowledge has not been effectively integrated into the model. Existing shallow fusion methods cannot change the black-box nature of the model, resulting in low sample utilization efficiency and model rules that easily deviate from physical laws. Therefore, there is an urgent need to construct a diagnostic paradigm with physical interpretability to achieve a deep integration of statistical learning and physical laws. Summary of the Invention

[0005] This invention proposes an AI-based diagnostic method for transformer partial discharge faults, which solves the technical problems of existing AI diagnostic models for transformer partial discharge faults being black-box structures, lacking physical interpretability, having low reliability of diagnostic results, weak generalization ability, and being unable to support engineering operation and maintenance decisions.

[0006] The technical solution of the present invention is implemented as follows: an AI diagnosis method for partial discharge faults in transformers, the method comprising the following steps: S1: Multi-channel partial discharge raw signals are synchronously acquired by sensors deployed on the transformer body, and after processing, a time sequence containing discharge pulse waveform, amplitude and phase information is obtained; S2: Construct a physical feature template library, which includes air gap discharge templates, surface discharge templates, and floating potential discharge templates. All templates have built-in phase distribution range, amplitude statistical characteristics, discharge repetition rate variation law, and pulse sequence time correlation parameters derived from the discharge physical mechanism. S3: The time series is simultaneously input into two parallel branches. The first branch extracts the actual physical feature vector based on the physical feature template library. The second branch converts the time series into a phase-resolved map and extracts the deep feature vector through a convolutional neural network. S4: Calculate the physical similarity score by comparing the actual physical feature vector with the fault template, obtain the deep classification confidence score by passing the deep feature vector through a classifier, and determine the fault type by weighted fusion of the two. Also, extract the deviation term between the actual physical feature vector and the fault template with the highest similarity. S5: Generate diagnostic explanation text based on deviation items. The text includes a description of the physical mechanism corresponding to the fault type, an assessment of the degree of insulation degradation, and a prediction of the defect development trend. S6: Output the fault type, diagnostic explanation text, and deviation items, and update the weight coefficients used for weighted fusion and the parameters in the physical feature template library through incremental learning.

[0007] Existing AI-based diagnostic methods for transformer partial discharge generally employ end-to-end deep neural network models, exhibiting a typical black-box reasoning structure. These models can only learn statistical features from data and complete label mapping; the extracted deep features have no clear correspondence with the physical mechanisms of partial discharge, resulting in a lack of interpretability in diagnostic results. Maintenance personnel cannot understand the basis of the model's decisions, making it difficult to trust and apply the diagnostic conclusions. Furthermore, existing methods can only output fault categories, failing to combine the physical mechanisms of discharge to provide information on defect location, insulation degradation degree, and development trends. This creates a significant information gap between diagnostic results and on-site engineering maintenance decisions. In addition, existing models rely on statistical correlations rather than physical causal laws, leading to a significant decrease in generalization ability when transformer operating conditions, structure, and materials change, resulting in frequent misjudgments and missed diagnoses. A few schemes that introduce physical features only achieve shallow fusion, failing to allow physical knowledge to dominate the reasoning logic and hindering continuous optimization and iteration of model and template parameters, making it difficult to adapt to the individual characteristics of different equipment.

[0008] This solution addresses the aforementioned issues by constructing a physical feature template library that integrates air gap discharge, surface discharge, and floating potential discharge. Mechanistic parameters are incorporated into the feature extraction process, and a dual-branch parallel structure of physical feature matching and deep network extraction is employed. This retains the high robustness of data-driven approaches while introducing traceable physical feature dimensions. Fault determination is achieved through weighted fusion of physical similarity scores and deep confidence scores. Simultaneously, feature deviation terms are extracted to generate explanatory text with physical mechanisms. This overcomes the core difficulties of black-box models being unexplainable and features being disconnected from mechanisms, resolving the technical bottleneck of deep integration of physical knowledge and AI-driven invention models. Furthermore, incremental learning enables dynamic updates of weights and template parameters, overcoming the problems of poor generalization and difficulty in adapting to actual field conditions with fixed models. It fills the technical gap where diagnostic results only have category labels without physical analysis, failing to support operational decisions. This truly achieves an organic combination of data-driven and physical mechanism-driven approaches, solving the long-standing technical challenges of poor interpretability, insufficient field applicability, and limited generalization ability in partial discharge AI-driven invention diagnosis.

[0009] As a preferred embodiment, step S1 uses three types of sensors—ultra-high frequency, ultrasonic, and high-frequency current—to synchronously acquire signals: the ultra-high frequency sensor is installed at the drain valve to couple electromagnetic wave signals, the ultrasonic sensor is attached to the outer shell to acquire acoustic emission signals, and the high-frequency current sensor extracts pulse current signals through the iron core grounding wire. The three types of sensors are triggered synchronously by the same source clock to ensure accurate timing alignment.

[0010] As a preferred implementation, when processing signals from the three types of sensors, background noise is first suppressed by wavelet threshold denoising, then pulse waveforms are extracted using adaptive threshold, the phase angle is calculated based on the pulse peak and the zero-crossing point of the power frequency, and finally the effective pulses of the channels are arranged in time sequence to form a multi-sensor fusion timing sequence.

[0011] As a preferred embodiment, the air gap discharge template constructed by the physical feature template library in step S2 obtains the nonlinear relationship between the discharge repetition rate and the applied voltage based on the equivalent circuit model, and defines the positive and negative half-cycle symmetrical phase window; the surface discharge template establishes a correlation model between pulse amplitude and discharge length based on the electric field distortion theory, and defines the phase lag offset interval; the floating potential discharge template obtains the pulse equal interval time correlation characteristics based on the analysis of the charging and discharging process.

[0012] In a preferred implementation, step S3 involves inputting the time series into two parallel branches. Specifically, the first branch directly calls the physical feature template library constructed in step S2 and extracts the actual physical feature vector through template matching. The second branch converts the time series into a two-dimensional phase-resolved map and inputs it into a convolutional neural network to automatically extract the depth feature vector.

[0013] In a preferred embodiment, the fault type determination in step S4 is achieved by calculating the similarity between the actual physical feature vector and the fault template in the physical feature template library to obtain a physical similarity score; inputting the deep feature vector into the classifier to obtain a deep classification confidence score; weighting the physical similarity score and the deep classification confidence score with preset weights to generate a comprehensive score, and determining the final fault type based on the category corresponding to the maximum value of the comprehensive score.

[0014] As a preferred implementation, the preset weights are determined through prior experiments and offline training. On a standard partial discharge sample set, the accuracy and recall of physical feature matching and deep feature classification are statistically analyzed, and initial weights are assigned based on the recognition reliability of the two types of features. At the same time, differentiated weights are set for air gap discharge, surface discharge, and floating potential discharge to ensure that the comprehensive score is more stable under typical fault modes and reduce the impact of single feature deviation on the final judgment result.

[0015] As a preferred implementation, the diagnostic explanation text generated in step S5 specifically includes: matching the corresponding physical mechanism description according to the fault type determined in step S4; quantitatively evaluating the degree of insulation degradation of the equipment by combining the extracted actual physical feature vectors; and predicting the potential trend and risk level of the defect over time based on the deviation term.

[0016] After adopting the above technical solution, the beneficial effects of the present invention are as follows: This method can accurately acquire multi-dimensional discharge time series data including pulse, amplitude and phase by synchronous acquisition of multiple sensors and time series signal processing, providing high-quality basic information for subsequent feature extraction. It uses a physical feature template library and a deep network dual-branch parallel feature extraction method, which not only relies on the discharge physical mechanism to ensure the physical meaning and interpretability of the features, but also uses convolutional neural networks to mine deep hidden features, taking into account both diagnostic accuracy and logical rationality. By using a weighted fusion of physical similarity and deep confidence to determine the fault type, the reliability and anti-interference ability of the diagnostic results are significantly improved. At the same time, feature deviation items are extracted and diagnostic explanation text containing physical mechanisms, deterioration degree, and development trend is generated. This completely breaks the unexplainability of the black box model of traditional AI invention, allowing maintenance personnel to clearly understand the cause, severity, and development risk of the fault. This greatly improves the credibility and on-site acceptability of the diagnostic conclusions, effectively connects the diagnostic results with engineering maintenance decisions, and eliminates the gap between diagnostic information and maintenance needs. Through incremental learning to dynamically update the fusion weights and template parameters, it can continuously adapt to the structural characteristics, operating conditions, and noise environment of different transformers, continuously optimize the template matching accuracy and decision weights, significantly improve the model's generalization ability and long-term stability, and avoid the problem of traditional fixed models failing when operating conditions change. Furthermore, this method deeply integrates decades of accumulated knowledge of partial discharge physics with artificial intelligence models, greatly improving sample utilization efficiency, reducing reliance on massive amounts of labeled data, and lowering model training costs and deployment difficulty. At the same time, the explicit participation of physical features makes fault judgment more comprehensive, effectively distinguishing between real discharge and interference signals, reducing false positives and false negatives, and improving the accuracy of transformer insulation status assessment. This provides strong support for the safe and stable operation of transformers, preventive maintenance, and fault early warning. Overall, it realizes the upgrade of AI-based invention diagnosis from pure data-driven to mechanism and data fusion-driven, comprehensively solving the three core problems of interpretability, practicality, and generalization, and has extremely high engineering application value and technology promotion value. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall structure of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: As Figure 1 As shown, the AI ​​diagnostic method for partial discharge faults in transformers includes the following steps: S1: Multi-channel partial discharge raw signals are synchronously acquired by sensors deployed on the transformer body, and after processing, a time sequence containing discharge pulse waveform, amplitude and phase information is obtained; S2: Construct a physical feature template library, which includes air gap discharge templates, surface discharge templates, and floating potential discharge templates. All templates have built-in phase distribution range, amplitude statistical characteristics, discharge repetition rate variation law, and pulse sequence time correlation parameters derived from the discharge physical mechanism. S3: The time series is simultaneously input into two parallel branches. The first branch extracts the actual physical feature vector based on the physical feature template library. The second branch converts the time series into a phase-resolved map and extracts the deep feature vector through a convolutional neural network. S4: Calculate the physical similarity score by comparing the actual physical feature vector with the fault template, obtain the deep classification confidence score by passing the deep feature vector through a classifier, and determine the fault type by weighted fusion of the two. Also, extract the deviation term between the actual physical feature vector and the fault template with the highest similarity. S5: Generate diagnostic explanation text based on deviation items. The text includes a description of the physical mechanism corresponding to the fault type, an assessment of the degree of insulation degradation, and a prediction of the defect development trend. S6: Output the fault type, diagnostic explanation text, and deviation items, and update the weight coefficients used for weighted fusion and the parameters in the physical feature template library through incremental learning.

[0021] The AI-based diagnostic method for transformer partial discharge faults described in this application can be applied to online monitoring and intelligent inspection scenarios of main transformers in power grids. The entire system relies on substation condition monitoring devices and edge computing terminals. Based on the physical mechanism of partial discharge, it adopts a data-driven and mechanism-driven approach, achieving diagnosis through a closed-loop logic of multi-source acquisition, template matching, deep mining, decision fusion, interpretable reasoning, and incremental optimization. In this scenario, the system first executes the signal acquisition and preprocessing work of step S1. Ultra-high frequency, ultrasonic, and high-frequency current sensors are arranged on the operating transformer body according to specifications. Each sensor is synchronously triggered by a unified clock, simultaneously acquiring multi-channel partial discharge raw signals while the equipment is continuously running. After the signals are amplified and filtered by the front-end conditioning circuit, they are digitally processed by the edge computing unit to complete noise suppression, pulse identification, and phase calibration, ultimately obtaining a time sequence containing discharge pulse waveform, amplitude, and phase information, providing a unified and standardized data foundation for subsequent feature extraction and fault determination.

[0022] After signal processing, the system enters the dual-branch feature extraction process of steps S2 and S3. In step S2, the system calls a pre-built physical feature template library, which contains three typical templates: air gap discharge, surface discharge, and floating potential discharge. Each template contains phase distribution range, amplitude statistical characteristics, discharge repetition rate law, and pulse time correlation parameters derived from the discharge physical mechanism, serving as the benchmark for physical feature comparison. In step S3, the time series is simultaneously fed into two parallel branches. The first branch, based on the physical feature template library, performs matching and quantization calculations on the time series to extract actual physical feature vectors with clear physical meaning. The second branch first reconstructs the time series into a phase-resolved map, and then completes high-dimensional deep feature mining through a convolutional neural network to form a deep feature vector. The two branches process in parallel, taking into account both physical interpretability and data feature mining capabilities.

[0023] In the first branch of step S3, the matching and quantization calculation based on the physical feature template library is a process of aligning, retrieving features, and numerically mapping the multi-channel time-series discharge sequence output from S1 with the three standard templates constructed in S2: air gap discharge, surface discharge, and floating potential discharge. Its core purpose is to extract structured feature vectors with clear physical meaning that can be directly correlated with the discharge mechanism from the original time-series signal. This process first uses the physical feature template library as a benchmark framework, sequentially analyzing four core parameters in the time-series sequence: pulse amplitude, phase distribution, discharge repetition rate, and pulse time correlation. In the phase distribution matching stage, the system first divides the time-series sequence into continuous power frequency phase intervals according to the power frequency period, and then compares them with the preset mechanistic phase intervals of the three types of templates to determine the concentration range, distribution symmetry, and initial phase shift of the actual discharge pulse in the positive and negative half-cycles. The matching degree is then normalized and quantized to form phase feature components.

[0024] In the amplitude statistical feature matching stage, the system traverses all valid discharge pulses within the time sequence, statistically analyzes the mean, dispersion, and polarity distribution of pulse amplitudes, and performs fitting calculations with the typical amplitude variation patterns defined in the template. This quantifies the degree of closeness between the actual amplitude characteristics and the standard mechanism characteristics, generating amplitude feature components. In the discharge repetition rate feature calculation stage, the system uses phase as the abscissa to statistically analyze the frequency of pulse occurrence within a unit phase interval, plots the discharge repetition rate distribution curve, and performs correlation calculations with the variation trend derived from the equivalent circuit and electric field theory in the template to obtain the discharge repetition rate feature components.

[0025] In the pulse sequence time correlation matching stage, the system analyzes the distribution pattern of time intervals between consecutive pulses, determining whether they exhibit random, clustered, or equally spaced distribution characteristics. It then performs pattern matching with the time correlation characteristics of typical templates such as floating potential discharge to quantify the regularity of the pulse sequence and form time correlation feature components. After completing the individual matching and quantification of the above four core physical features, the system combines each component according to the weights defined by the template, ultimately forming an actual physical feature vector that is completely consistent with the dimensions of the physical feature template library.

[0026] Each term in this vector corresponds to an interpretable physical parameter, rather than an abstract numerical value. It can directly reflect the type and development level of the discharge defect, providing an interpretable, traceable, and comparable standardized input for template similarity calculation in the subsequent step S4. It also lays the mechanistic foundation for subsequent deviation term calculation and diagnostic explanation text generation.

[0027] The system then executes the decision fusion and deviation extraction operations of invention S4. On one hand, it calculates the similarity between the physical feature vector and each fault template to obtain a physical similarity score. On the other hand, it inputs the deep feature vector into a classifier to output a deep classification confidence score. The two results are then fused by weighting, and the final fault type is determined based on the highest score. Simultaneously with fault type determination, the system calculates and extracts the deviation term between the actual physical feature vector and the optimal matching template. This deviation term reflects the degree of difference between the actual discharge characteristics and the standard physical template, providing crucial quantitative evidence for generating interpretable diagnostic conclusions. This ensures that the diagnostic process no longer relies on simple data mapping but has traceable physical logic support.

[0028] In step S5 of the invention, the system uses the aforementioned deviation items as the core basis, combined with the determined fault type, to generate a diagnostic explanation text containing professional content. The text not only provides a description of the physical mechanism corresponding to the fault, but also quantitatively assesses the degree of insulation degradation based on the magnitude of the deviation, and infers the defect development trend and operational risks based on pulse characteristics and phase distribution patterns. This expands the output from a single category label to a complete engineering analysis report, enabling maintenance personnel to clearly understand the cause, severity, and potential risks of the fault. This effectively breaks down the information barrier between the AI-generated diagnostic results and on-site engineering decisions, significantly improving the credibility and practicality of the diagnostic conclusions.

[0029] Finally, in step S6 of the invention, the system synchronously uploads the fault type, diagnostic explanation text, and deviation items to the monitoring backend and remote operation and maintenance platform, providing intuitive support for on-site maintenance and scheduling management. Simultaneously, the system uses the diagnostic data as incremental learning samples to optimize the weighted fusion coefficients and dynamically corrects various parameters in the physical feature template library. This allows the templates and decision rules to gradually adapt to the transformer's structural characteristics, operating conditions, and on-site environment, forming a complete closed loop from signal acquisition, processing, feature extraction, fault diagnosis, interpretation output to model optimization. Through this process, the method can sustainably maintain high reliability, high interpretability, and strong on-site adaptability in practical engineering scenarios, providing professional and intelligent technical support for transformer insulation condition assessment and safe and stable operation.

[0030] Example 2: Before introducing the multi-sensor synchronous acquisition scheme in step S1, traditional implementation processes often relied on single-channel signals or non-strictly synchronized multi-channel data. This resulted in misalignment of different physical quantities (electromagnetic, ultrasonic, and current) on the time axis, making effective fusion difficult. With this implementation method, the implementation scenario fundamentally changes. At the field deployment level, the system no longer installs only a single sensor, but instead deploys three types of sensors—ultra-high frequency, ultrasonic, and high-frequency current—in a standardized manner, and performs hard synchronization through a unified clock source. This directly changes the underlying logic of data acquisition, transforming it from serial acquisition to parallel synchronous acquisition. In terms of workflow, a crucial pre-processing step of timing alignment is added to the signal processing stage. The system first ensures that the three types of sensors synchronously acquire raw signals under the same millisecond-level clock trigger. Then, before wavelet denoising and feature extraction, it performs time axis calibration and phase unification of the multi-channel signals. This optimization completely solves the problem of misalignment of different signals in the time dimension, enabling the subsequently extracted pulse waveform, amplitude and phase information to truly reflect the multidimensional characteristics of the same physical discharge event. This provides a high-quality data base with spatiotemporal synchronization for the dual-branch feature fusion in step S3, greatly improving the accuracy and reliability of the subsequent diagnostic process.

[0031] Example 3: After introducing the signal processing scheme described above, which includes wavelet threshold denoising, adaptive threshold extraction, phase angle calculation, and timing arrangement, the core change in the implementation scenario and workflow lies in the dual improvement of data quality and feature granularity. The original signal processing workflow may be relatively coarse, directly performing threshold segmentation or simple filtering, which can easily lead to background noise drowning out weak discharge pulses or introducing spurious features. After implementing this scheme, the signal processing workflow is reshaped into a multi-stage, refined purification process. First, the system uses a wavelet threshold denoising algorithm, an advanced technique that effectively suppresses Gaussian noise while preserving signal edges (pulse features), unlike traditional mean filtering. Next, an adaptive thresholding method is used to extract pulses. This method dynamically adjusts the threshold based on local signal features, rather than using a fixed threshold, thus enabling more accurate identification of real discharge pulses in complex electromagnetic environments. Subsequently, the phase angle is calibrated by calculating the relative position of the pulse peak and the power frequency zero-crossing point, a crucial step in converting the time-domain signal into frequency-domain features usable for physical analysis. Finally, the effective pulses from each channel are arranged in time sequence to form a standardized multi-sensor fusion timing sequence. The introduction of this entire process ensures that the data entering subsequent steps is no longer the mixed raw signal, but high-quality data that has been rigorously purified, has clear features, and is time-ordered. This fundamentally solves the problem of misjudgment caused by noise interference and feature ambiguity, and provides clear and clean input for physical template matching in step S2 and deep feature learning in step S3.

[0032] Example 4: After introducing the three types of templates constructed based on the physical mechanism of discharge in step S2, the core change in the implementation scenario and workflow lies in the transformation of the feature extraction paradigm from purely data-driven to mechanism- and data-integrated driven. In the implementation scenario, the system is no longer a simple black box model, but an interpretable system with built-in domain expert knowledge. The construction of the physical feature template library is no longer based on empirical parameters derived from massive data statistics, but is rooted in fundamental theories such as electromagnetics and insulation engineering. The air gap discharge template is based on the equivalent circuit model, the surface discharge template is based on the electric field distortion theory, and the floating potential discharge template is based on the analysis of the charging and discharging process. This makes the workflow undergo a qualitative leap in step S2. The process is no longer directly extracting features from the signal, but first establishing a priori knowledge base containing clear physical meanings and constraints. In the first branch of the subsequent step S3, the matching calculation is no longer a simple numerical similarity comparison, but a conformity verification based on physical laws. For example, when matching the surface discharge template, the system not only compares the amplitude, but also verifies whether there is a hysteresis shift in the phase that conforms to the electric field distortion theory.

[0033] Building a physical feature template library does not require continuous acquisition of existing sensor data. Its core construction logic is based on theoretical derivation and mechanism modeling of the physical mechanism of partial discharge, rather than relying on real-time or continuous sensor data fitting. The core feature parameters of the template library (such as the phase range and amplitude law of various discharges) are all derived from professional theories such as the equivalent circuit model of air gap discharge, the theory of electric field distortion along the surface discharge, and the dynamic equation of charge and discharge of floating potential. After standard test conditions, they are calibrated to converge to a standardized parameter range, completing the basic construction. Subsequently, parameter fine-tuning is only performed in the incremental learning stage of step S6, combined with field diagnostic data (including data processed in S1), rather than continuously acquiring sensor data during the construction stage.

[0034] The processed data in step S1 (including the time series of discharge pulse waveforms, amplitude, and phase information) mainly plays a role in calibration, verification, and adaptation optimization during the construction of the physical feature template library. Specifically, it is manifested in the following ways: 1. After the basic parameters of the template library are initially determined through mechanism derivation, the rationality and engineering adaptability of the derivation parameters are verified using the standardized time series processed in S1. For example, the boundary parameters of the phase window and amplitude interval in the template are calibrated by using the actual phase distribution and amplitude range in the time series to ensure that the template parameters match the actual discharge characteristics on site; 2. For transformers of different models and operating conditions, the parameter ranges in the template library are adjusted in a targeted manner using the data processed in S1 to avoid deviations between theoretically derived parameters and actual operating conditions on site; 3. After the template library is constructed, the data processed in S1 serves as the input basis for subsequent template matching (step S3) and similarity calculation (step S4), which verifies the effectiveness of the template library in reverse and provides measured data support for the optimization and improvement of the initial parameters of the template library. This ensures that the template library can accurately capture the physical characteristics of actual discharge on site, rather than directly participating in the mechanism derivation process of the core parameters of the template.

[0035] The physical feature template library is composed of four core elements: three types of basic fault templates, a unified and standardized feature parameter system, a parameter range storage structure, and an adaptation and update interface. The three types of basic templates are: internal air gap discharge template, insulation surface discharge template, and floating potential discharge template, which are the core components of the template library. The unified and standardized feature parameter system is a feature dimension standard shared by all templates, ensuring that the feature dimensions of different templates are aligned and can be compared horizontally. The parameter range storage structure does not store each feature as a fixed value, but rather solidifies it in the form of a reasonable value range, adapting to different transformer models and different operating conditions. The adaptation and update interface reserves an incremental learning access port to support subsequent dynamic parameter correction and range fine-tuning.

[0036] The core characteristic parameters and mechanism derivation of the air gap discharge template in the physical characteristic template library are as follows: An equivalent circuit model of air gap discharge inside solid insulation is constructed, treating the insulating air gap as a capacitor and resistor connected in series, and then connected in parallel with the insulating dielectric capacitor, using the external power frequency AC voltage as the excitation source; based on gas breakdown theory, the critical condition for self-sustaining discharge of the air gap is derived; combined with the periodic variation law of the power frequency voltage, the relationship between the air gap electric field strength and the voltage phase is derived, determining the phase distribution range of the air gap discharge; by analyzing the transient process of the equivalent circuit, the variation law of the charging and discharging voltage of the air gap capacitor before and after discharge is obtained, and the statistical characteristics of the discharge pulse amplitude are derived; by combining the air gap breakdown and arc extinction conditions, the correlation between the discharge repetition rate and the external voltage is derived; combined with the physical characteristics of random ionization breakdown of air gap discharge, the distribution characteristics of the pulse sequence time interval are derived; the various characteristic parameters obtained above are preliminarily regularized to form the air gap discharge template.

[0037] Construction of the surface discharge template: A simulation model of the surface discharge electric field is established, considering the difference in dielectric constant between the insulating medium and air and the edge effect, and the distribution law of the tangential electric field and the normal electric field on the insulating surface is solved; based on the electric field distortion theory, the hysteresis characteristics of the starting phase of the surface discharge relative to the zero-crossing point of the power frequency voltage are derived, and the phase shift characteristic interval is determined; combined with the surface discharge creepage propagation mechanism, the correlation between the discharge length and the surface electric field strength and the dielectric withstand voltage is derived, and the correlation law between the discharge pulse amplitude and the discharge length is further derived; based on the physical characteristics of multiple creepages and gradual development of surface discharge, the statistical distribution characteristics of the pulse amplitude and the variation law of the discharge repetition rate with the applied voltage are derived; combined with the surface charge accumulation and discharge characteristics, the cluster distribution characteristics of the pulse sequence are derived; the characteristic parameters obtained above are regularized to form the surface discharge template.

[0038] Construction of the floating potential discharge template: Based on the charging and discharging dynamics of the floating electrode, a model of charge induction, accumulation, and discharge of the floating metal electrode in a power frequency alternating electric field is established; according to the law of charge conservation, the relationship between the floating electrode potential and the voltage phase is derived, and the phase distribution range of the floating potential discharge is determined; the instantaneous discharge process of the floating electrode charge is analyzed, and the steady-state characteristics of the discharge pulse amplitude are derived; the calculation method of the recharging time after the floating electrode discharge is derived, and the equidistant distribution characteristics of the pulse sequence are determined; combined with the constant charging and discharging period, the steady-state law of the discharge repetition rate is derived; the characteristic parameters obtained above are regularized to form the floating potential discharge template.

[0039] Example 5: After introducing the dual-branch parallel processing scheme in step S3, the core change in the implementation scenario and workflow lies in the combination of breadth and depth of feature extraction. The original implementation process typically used a single branch, relying either on manually designed physical features or end-to-end deep features, neither of which could be simultaneously achieved. This scheme simultaneously inputs the time series into two parallel branches, constructing a workflow where the physical feature branch and the deep feature branch work collaboratively. In terms of workflow, step S3 is split into two synchronous but independent processing paths. The first branch utilizes the physical template library constructed in step S2 to perform explicit and interpretable physical feature matching, ensuring the logical basis and anti-interference capability of the diagnosis. The second branch transforms the time series into a two-dimensional phase-resolved map and uses a convolutional neural network for implicit, high-dimensional deep feature mining, capturing complex patterns that are difficult for the human eye to perceive. This parallel processing enables the system in the implementation scenario to simultaneously utilize known physical laws and learned data patterns, achieving complementary advantages.

[0040] Example 6: After introducing the weighted fusion scheme based on physical similarity and deep confidence, the core change in the implementation scenario and workflow lies in the shift of the decision-making mechanism in the fault determination stage from a single vote to a weighted comprehensive decision. The original implementation process might directly take the intersection or union of the conclusions of two branches, or simply prioritize the result of one branch, which easily leads to the bias of a single feature dominating the final judgment. After implementing this scheme, the workflow in step S4 adds a crucial fusion and decision-making step. The system first calculates two independent decision criteria: physical similarity score and deep classification confidence, and then performs a weighted summation using preset weights to generate a comprehensive score. This mechanism allows the contribution of physical features and deep features to be dynamically adjusted under different fault types. For example, in some typical fault modes, the matching degree of physical features may be higher, and the system will automatically assign them a higher weight; while in complex working conditions, the generalization ability of deep features may be stronger, so the weight is adjusted. This weighted fusion strategy makes the system decision-making process in the implementation scenario more scientific and flexible, effectively avoiding misjudgments caused by single feature bias, and significantly improving the accuracy and stability of the final fault type determination.

[0041] Example 7: After introducing the differentiated weight initialization and generation scheme for explanatory text containing mechanisms, degradation, and trends, the core change in the implementation scenario and workflow lies in the system's upgrade from passive diagnosis to proactive intelligence and intelligent explanation. Regarding weight settings, the workflow changes from fixed weights to refined weight initialization. Through prior experiments and offline training, the system statistically analyzes the recognition reliability of physical and depth features for different fault types (air gap, surface, suspension) to set differentiated initial weights. This enables the system to have optimized decision-making capabilities for different fault modes from the initial deployment stage. In terms of explanatory text generation, a crucial natural language generation and knowledge mapping step is added to the workflow. The system no longer simply outputs a fault code but automatically generates structured explanatory text containing physical mechanism descriptions, quantitative assessments of insulation degradation, and predictions of defect development trends, based on deviation items and fault types. This change significantly improves the system's usability in the implementation scenario, making the fault type of the diagnostic results more intuitively displayed, effectively eliminating the cognitive gap between maintenance personnel and the AI ​​model, and greatly improving the credibility of diagnostic conclusions and on-site adoption rates.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based diagnostic method for partial discharge faults in transformers, characterized in that, The method includes the following steps: S1: Multi-channel partial discharge raw signals are synchronously acquired by sensors deployed on the transformer body, and after processing, a time sequence containing discharge pulse waveform, amplitude and phase information is obtained; S2: Construct a physical feature template library, which includes air gap discharge templates, surface discharge templates, and floating potential discharge templates. All templates have built-in phase distribution range, amplitude statistical characteristics, discharge repetition rate variation law, and pulse sequence time correlation parameters derived from the discharge physical mechanism. S3: The time series is simultaneously input into two parallel branches. The first branch extracts the actual physical feature vector based on the physical feature template library. The second branch converts the time series into a phase-resolved map and extracts the deep feature vector through a convolutional neural network. S4: Calculate the physical similarity score by comparing the actual physical feature vector with the fault template, obtain the deep classification confidence score by passing the deep feature vector through a classifier, and determine the fault type by weighted fusion of the two. Also, extract the deviation term between the actual physical feature vector and the fault template with the highest similarity. S5: Generate diagnostic explanation text based on deviation items. The text includes a description of the physical mechanism corresponding to the fault type, an assessment of the degree of insulation degradation, and a prediction of the defect development trend. S6: Output the fault type, diagnostic explanation text, and deviation items, and update the weight coefficients used for weighted fusion and the parameters in the physical feature template library through incremental learning.

2. The AI ​​diagnostic method for transformer partial discharge faults as described in claim 1, characterized in that: The multi-channel partial discharge raw signal acquired in step S1 is simultaneously acquired by three types of sensors: ultra-high frequency, ultrasonic, and high-frequency current. The ultra-high frequency sensor is installed at the drain valve to couple electromagnetic wave signals, the ultrasonic sensor is attached to the outer shell to acquire acoustic emission signals, and the high-frequency current sensor extracts pulse current signals through the iron core grounding wire. The three types of sensors are triggered synchronously by the same source clock to ensure accurate timing alignment.

3. The AI ​​diagnostic method for transformer partial discharge faults as described in claim 2, characterized in that: When processing signals from the three types of sensors, background noise is first suppressed by wavelet threshold denoising, then pulse waveforms are extracted by adaptive threshold, the phase angle is calculated based on the pulse peak and the zero-crossing point of the power frequency, and finally the effective pulses of the channels are arranged in time sequence to form a multi-sensor fusion time sequence.

4. The AI ​​diagnostic method for transformer partial discharge faults as described in claim 1, characterized in that: In step S2, the air gap discharge template constructed from the physical feature template library obtains the nonlinear relationship between the discharge repetition rate and the applied voltage based on the equivalent circuit model, and defines the positive and negative half-cycle symmetrical phase window; the surface discharge template establishes a correlation model between pulse amplitude and discharge length based on the electric field distortion theory, and defines the phase lag offset interval; the floating potential discharge template obtains the pulse equal interval time correlation characteristics based on the analysis of the charging and discharging process.

5. The AI ​​diagnostic method for transformer partial discharge faults as described in claim 1, characterized in that: In step S3, the time series is input into two parallel branches. Specifically, the first branch directly calls the physical feature template library constructed in step S2 and extracts the actual physical feature vector through template matching; the second branch converts the time series into a two-dimensional phase-resolved map and inputs it into a convolutional neural network to automatically extract the depth feature vector.

6. The AI ​​diagnostic method for partial discharge faults in transformers as described in claim 1, characterized in that: In step S4, the fault type is determined by calculating the similarity between the actual physical feature vector and the fault template in the physical feature template library to obtain a physical similarity score; the deep feature vector is input into the classifier to obtain a deep classification confidence score; the physical similarity score and the deep classification confidence score are weighted and summed using preset weights to generate a comprehensive score, and the category corresponding to the maximum value of the comprehensive score is determined as the final fault type.

7. The AI ​​diagnostic method for transformer partial discharge faults as described in claim 6, characterized in that: The preset weights are determined through prior experiments and offline training. On the standard partial discharge sample set, the accuracy and recall of physical feature matching and deep feature classification are statistically analyzed respectively, and the initial weights are assigned based on the recognition reliability of the two types of features. At the same time, differentiated weights are set for air gap discharge, surface discharge, and floating potential discharge to ensure that the comprehensive score is more stable under typical fault modes and reduce the impact of single feature deviation on the final judgment result.

8. The AI ​​diagnostic method for partial discharge faults in transformers as described in claim 1, characterized in that: The diagnostic explanation text generated in step S5 specifically includes: matching the corresponding physical mechanism description according to the fault type determined in step S4; quantitatively assessing the degree of insulation degradation of the equipment by combining the extracted actual physical feature vectors; and predicting the potential trend and risk level of the defect over time based on the deviation term.

Images