Power distribution transformer on-line detection and health assessment system

Abstract

This invention relates to an online detection and health assessment system for distribution transformers, belonging to the field of online monitoring and health assessment technology for power equipment. It includes an equipment perception layer, an edge computing layer, and a cloud-based intelligent platform layer. The equipment perception layer deploys multi-parameter fusion sensors and auxiliary equipment to construct a six-dimensional feature space encompassing vibration, magnetic flux, temperature, etc. The edge computing layer achieves real-time anomaly detection at the edge through a built-in AI chip and encrypts the data to the cloud via an edge gateway. The cloud-based intelligent platform layer integrates a magnetic flux-temperature coupled aging model, a PINN physical information neural network fusion model, and a comprehensive health index (CHI) calculation engine to perform in-depth assessment and remaining life prediction of the transformer, and matches differentiated operation and maintenance strategies based on four health levels. This system solves the problems of single monitoring parameters, incomplete aging models, and insufficient edge-cloud collaboration in existing technologies, achieving panoramic perception, quantitative assessment, and intelligent operation and maintenance of distribution transformer status.

Description

Technical Field

[0001] This invention relates to the field of online monitoring and health assessment technology for power equipment, and specifically to an online detection and health assessment system for distribution transformers. Background Technology

[0002] Distribution transformers are core equipment in power distribution networks. my country has approximately 20 million distribution transformers in operation, and their operational reliability directly impacts power quality and grid security. With the large-scale integration of distributed photovoltaic systems and electric vehicle charging facilities, the operating conditions of power distribution networks are becoming increasingly complex. Distribution transformers face multiple stresses, including overload, overvoltage, and harmonic pollution, leading to accelerated insulation aging and a significantly increased risk of failure. Statistics show that approximately 30% of transformer failures are directly related to insulation aging, and distribution transformer failures account for over 60% of all transformer failures, highlighting the urgent need for effective online detection and health assessment methods.

[0003] The existing online monitoring technology for distribution transformers has the following main shortcomings: (1) The monitoring parameters are singular. Most schemes only monitor temperature or oil chromatography, lacking the fusion perception of multiple physical quantities such as magnetic flux, vibration, and acoustic signature.

[0004] (2) The effect of magnetic flux density on aging has not been fully modeled. Magnetostriction and stray losses caused by high magnetic flux density will accelerate insulation fatigue, but the existing model only considers temperature as a single factor.

[0005] (3) Health assessment lacks quantitative standards. Existing technologies mostly use qualitative descriptions (such as "normal" and "abnormal"), and lack continuous quantitative health indices and confidence levels.

[0006] (4) Insufficient collaboration between the end side and the cloud, either relying solely on cloud computing leads to poor real-time performance, or relying solely on end-side computing leads to limited accuracy.

[0007] (5) The operation and maintenance decision-making lacks differentiated strategies and does not formulate targeted maintenance plans and load strategies based on the health status level. Summary of the Invention

[0008] This invention addresses the shortcomings of existing technologies by providing an online detection and health assessment system for distribution transformers. It solves the problems of single monitoring parameters, incomplete aging models, and insufficient end-to-cloud collaboration in existing technologies, and realizes panoramic perception, quantitative assessment, and intelligent operation and maintenance of distribution transformer status.

[0009] The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions: An online detection and health assessment system for distribution transformers includes: an equipment sensing layer, an edge computing layer, and a cloud-based intelligent platform layer.

[0010] The device's sensing layer deploys multi-parameter fusion sensors to achieve real-time five-dimensional sensing of vibration, magnetic flux, temperature, noise, and ultrasound. It is also equipped with an IoT smart meter, a partial discharge current sensor, and an infrared temperature sensor to construct a six-dimensional feature space.

[0011] The edge computing layer includes an AI chip and an edge gateway built into the multi-parameter fusion sensor. The AI ​​chip is used to realize real-time anomaly detection on the edge side, and the edge gateway is used to realize data aggregation, protocol conversion, and encrypted cloud upload.

[0012] The cloud-based intelligent platform layer includes a magnetic flux-temperature coupled aging model, a temperature-oil chromatography chemical aging model, a PINN physical information neural network fusion model, a comprehensive health index CHI calculation engine, a Weibull remaining life prediction model, and a differentiated operation and maintenance decision engine.

[0013] The multi-parameter fusion sensor includes three xSensus-NEO sensors, with the following deployment locations and monitoring functions: The first xSensus-NEO sensor is installed on the top of the fuel tank and is used to monitor vibration, magnetic flux, and temperature, with sampling frequencies of 10kHz, 1kHz, and 1Hz, respectively.

[0014] The second xSensus-NEO sensor is installed on the high-voltage bushing side to monitor noise, ultrasound, and temperature, with an ultrasound sampling frequency of 48kHz.

[0015] The third xSensus-NEO sensor is installed on the low-pressure bushing side to monitor vibration and temperature.

[0016] The three sensors are connected to the edge gateway via an RS-485 bus.

[0017] Preferably, the Comprehensive Health Index (CHI) is calculated using a four-index fusion method, as shown in the formula: CHI=0.25×MHI+0.25×EHI+0.25×THI+0.25×IHI.

[0018] Among them, MHI is the Mechanical Health Index, calculated based on vibration and acoustic signature characteristics; EHI is the Electromagnetic Health Index, calculated based on magnetic flux and electrical characteristics; THI is the Thermal Health Index, calculated based on temperature and load characteristics; and IHI is the Insulation Health Index, calculated based on temperature history and oil chromatographic characteristics.

[0019] The CHI value ranges from 0 to 100, corresponding to four health levels: normal (CHI≥80), caution (60≤CHI<80), abnormal (40≤CHI<60), and dangerous (CHI<40).

[0020] The differentiated operation and maintenance decision engine automatically matches operation and maintenance strategies based on the four-level health status: Normal level (CHI≥80): Perform standard inspections every 3 months and operate at full capacity.

[0021] Attention level (60≤CHI<80): Shorten the inspection cycle to 1 month and limit peak load to less than 90%.

[0022] Abnormal level (40≤CHI<60): Encrypt inspection cycle to 1 week, conduct live testing once a month, and limit the load to less than 80%.

[0023] Hazard level (CHI<40): Perform daily inspections, immediately arrange power outage tests, switch to backup transformers, or limit the load to less than 50%.

[0024] Preferably, the electromagnetic health index EHI takes into account the magnetic flux density ratio and harmonic effects, and is calculated using the formula: EHI = 100 × (1 - D_flux / D_limit); where D_flux is the overexcitation cumulative damage index, D_flux = Σ[(B_rms / B_rated)^n-1] × Δt, B_rms is the effective value of magnetic flux density, B_rated is the rated magnetic flux density, n is the exponential factor with a value range of 4 to 6, Δt is the time interval, and D_limit is the damage limit threshold.

[0025] Preferably, the Insulation Health Index (IHI) is calculated based on the quantitative mapping relationship between oil chromatography data and the degree of polymerization (DP) of insulation, using the following formula: IHI=100×(DP_current-DP_failure) / (DP_initial-DP_failure).

[0026] Where DP_failure is the alarm threshold, with a value of 250; DP_initial is the initial degree of polymerization; and DP_current is the current degree of polymerization, mapped through the oil chromatography data mapping function DP = f(CO, furfural The value is estimated by tanδ (temperature history integral).

[0027] Preferably, the flux-temperature coupled aging model comprehensively considers the synergistic effects of the core flux density B and the winding hot spot temperature θ_hs on insulation aging, and the core equation is: dDP / dt =-A·exp(-Ea / kT)·B^α·DP^β.

[0028] Wherein, DP is the degree of insulating polymerization, A is the pre-exponential factor, Ea is the activation energy, k is the Boltzmann constant, T is the absolute temperature, α is the magnetic susceptibility index, ranging from 0.8 to 1.2, and β is the aging level, ranging from 0.9 to 1.1.

[0029] Preferably, the PINN physical information neural network fusion model embeds electromagnetic constraints, thermal constraints, aging constraints, and oil chromatography constraints into the loss function, and the composite loss function is: L_total = •L_data + •L_physics+ •L_boundary+ • L_reg; where L_data is the data loss, L_physics is the physical constraint loss, L_boundary is the boundary condition loss, and L_reg is the regularization loss; These are the corresponding weighting coefficients.

[0030] Preferably, the backbone network of the PINN physical information neural network fusion model adopts a 5-layer fully connected structure, with the node configuration as FC128→FC256→FC512→FC256→FC128. The activation function is Swish, combined with residual connections and LayerNorm regularization. The physical constraint layer encodes Maxwell's equations, the heat conduction equation, the Arrhenius-Coupled aging equation, and the gas generation rate equation as regularization terms.

[0031] Preferably, the system employs a two-level health status assessment engine. The first level is real-time assessment at the edge, where an isolated forest anomaly detection algorithm and a lightweight CNN fault classification model are run on the AI ​​chip built into the multi-parameter fusion sensor, outputting the operating status and severity level with a response time of less than 5 seconds. The second level is cloud-based fusion assessment, where a PINN physical information neural network fusion model is run on the cloud intelligent platform layer for in-depth assessment, outputting the comprehensive health index (CHI) and remaining life prediction results. When an anomaly is detected at the edge, a data uploading mechanism is triggered, uploading high-frequency sampling data from the anomaly period to the cloud intelligent platform layer.

[0032] Preferably, the Weibull remaining lifetime prediction model adopts a two-parameter Weibull distribution function: R(t)=exp[-(t / η)^β]; where η is the characteristic lifetime, which is a function of the current degree of aggregation, temperature history integral, and load rate; β is the shape parameter, which depends on the aging acceleration factor and maintenance history; the model combines Monte Carlo stochastic simulation to consider parameter uncertainty, with N≥10000 simulations, and outputs the confidence interval of the remaining lifetime RUL.

[0033] Preferably, the six-dimensional feature space includes: V-vibration domain features, F-magnetic flux domain features, T-temperature domain features, A-acoustic domain features, E-electrical domain features, and H-history domain features.

[0034] V-Vibrational Domain Characteristics: including vibration RMS, peak value, kurtosis, 100Hz frequency band energy, and 50Hz frequency band energy.

[0035] F-Fluid domain characteristics include B_rms, flux harmonicity THD_B, saturation factor K_sat, and overexcitation accumulation D_flux.

[0036] T-temperature domain characteristics: including hot spot temperature θ_hs, temperature rise rate dT / dt, and cumulative damage index CDI.

[0037] A-Acoustic domain features include PRPD spectrum, TEO energy operator, and wavelet entropy.

[0038] E-Electrical Domain Characteristics: including load factor, current harmonic distortion rate (THD-I), active power (P), reactive power (Q), and power factor.

[0039] H-History domain characteristics: including parameter trends and cumulative damage.

[0040] The present invention can achieve the following effects: This invention provides an online monitoring and health assessment system for distribution transformers. Compared with existing technologies, it is based on a multi-source information fusion sensing and cloud-edge-device collaborative computing architecture. Through the deep integration of physical laws and artificial intelligence algorithms, it achieves real-time monitoring, accurate assessment, lifespan prediction, and intelligent operation and maintenance of distribution transformers. It has the following beneficial effects: I. Multi-dimensional panoramic perception enhances the comprehensiveness of condition monitoring. By deploying multi-parameter fusion sensors and various auxiliary sensing devices, a six-dimensional feature space is constructed, encompassing vibration, magnetic flux, temperature, noise, ultrasound, and electrical parameters. In particular, the introduction of magnetic flux and vibration monitoring fills the gaps in traditional single-parameter monitoring, enabling the detection of early mechanical and electromagnetic fault signs such as core loosening, winding deformation, and overexcitation.

[0041] II. Innovative Aging Coupled Model to Improve Lifetime Prediction Accuracy. A magnetic flux-temperature coupled aging model was established, innovatively incorporating magnetic flux density BB into the insulation aging equation (dDP / dt = -A•exp(-Ea / kT)•B^α•DP^βdDP / dt=-A•exp(-Ea / kT)•Bα•DPβ), quantitatively describing the synergistic accelerating effect of high magnetic flux density and high temperature on insulation aging, and solving the problem of lifetime prediction bias caused by traditional models that only consider the single factor of temperature.

[0042] III. Physics-AI Fusion Enhances the Reliability of Evaluation Results. The cloud-based PINN physical information neural network fusion model embeds physical laws such as Maxwell's equations, the heat conduction equation, and the Arrhenius-Coupled aging equation into the neural network's loss function. This hybrid modeling approach of "data-driven + mechanism-constrained" significantly improves the model's generalization ability and physical interpretability in scenarios with sparse or out-of-distribution data compared to purely data-driven AI models.

[0043] IV. Edge-Cloud Two-Tier Collaborative Architecture, Balancing Real-Time Performance and High Accuracy. A two-tier architecture of "real-time edge assessment + in-depth cloud assessment" has been constructed. On the edge, AI chips are used to achieve millisecond-level response (<5 seconds) anomaly detection and fault classification, ensuring real-time early warning; the cloud uses the PINN model for high-precision in-depth assessment and lifespan prediction. At the same time, an "anomaly-triggered cloud migration" mechanism is set up, which ensures both the processing efficiency of massive monitoring data and the in-depth analysis of key anomaly data.

[0044] V. Quantifying Health Level Assessment for Precise O&M Decision-Making. A four-index fusion calculation method for the Comprehensive Health Index (CHI) is proposed (CHI = 0.25×MHI + 0.25×EHI + 0.25×THI + 0.25×IHI), quantifying health status into a value from 0 to 100. Based on this, a four-level early warning system (normal, attention, abnormal, and dangerous) is established, and differentiated O&M strategies (such as inspection cycle adjustment, load limiting, and spare parts management) are automatically matched. This transforms O&M from "periodic maintenance" to "condition-based maintenance" and "predictive maintenance," effectively reducing O&M costs and improving power supply reliability. Attached Figure Description

[0045] Figure 1 This is a diagram of the online detection and health assessment system for distribution transformers according to the present invention.

[0046] Figure 2 This is a flowchart of the two-level health status assessment engine (HSE) of the present invention.

[0047] Figure 3 This is a schematic diagram of the four-index fusion calculation model of the Comprehensive Health Index (CHI) of this invention.

[0048] Figure 4 This is the flowchart of the four-level early warning system and differentiated operation and maintenance decision-making of the present invention.

[0049] Figure 5 This is a flowchart of the multi-sensor data fusion and processing of the present invention. Detailed Implementation

[0050] The technical solution of the invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0051] Example 1: As Figure 1 As shown, the online detection and health assessment system for distribution transformers includes an equipment sensing layer, an edge computing layer, and a cloud-based intelligent platform layer. The equipment sensing layer deploys three xSensus-NEO multi-parameter fusion sensors on the 10kV S13-M-400 / 10 oil-immersed distribution transformer: xSensus-NEO #1 is installed on the top of the tank, monitoring vibration, magnetic flux, and temperature, with sampling frequencies of 10kHz / 1kHz / 1Hz, primarily monitoring core magnetic flux density, abnormal iron loss, and top oil temperature; xSensus-NEO #2 is installed on the high-voltage bushing side, monitoring noise, ultrasound, and temperature, with an ultrasound sampling frequency of 48kHz, primarily monitoring partial discharge and bushing insulation status; xSensus-NEO #3 is installed on the low-voltage bushing side, monitoring vibration and temperature, primarily monitoring winding loosening and lead wire vibration. In addition, an IoT smart meter (monitoring load current, voltage, power, and harmonics), a partial discharge current sensor HFCT (10MHz sampling), and two infrared temperature sensors (monitoring the temperature difference between the radiator inlet and outlet oil) are also deployed.

[0052] Example 2: Figure 2 As shown, the two-level Health Status Assessment Engine (HSE) comprises two layers: real-time edge assessment and cloud-integrated assessment. The first-level edge assessment runs on the xSensus-NEO's built-in AI chip (Quad-core ARM Cortex-A7), employing the Isolation Forest algorithm for multi-parameter anomaly detection and a lightweight CNN model for fault mode classification. It outputs the operating status (normal / attention / abnormal / dangerous) and severity level (1-4), with a response time of <5 seconds. When an anomaly is detected on the edge, a data upload mechanism is triggered, uploading high-frequency sampled data from the anomaly period to the cloud platform.

[0053] The second-level cloud-based fusion evaluation is deployed on a cloud-based intelligent platform, employing the PINN physical information neural network for deep evaluation. The PINN network backbone is a 5-layer fully connected structure (FC128→FC256→FC512→FC256→FC128), with the Swish activation function, coupled with residual connections and LayerNorm regularization. The physical constraint layer encodes four types of PDE constraints as regularization terms: electromagnetic constraints (Maxwell's equations), thermal constraints (heat conduction equations), aging constraints (Arrhenius-Coupled equations), and oil chromatography constraints (gas generation rate equations). The composite loss function is L_total = 1.0·L_data + 0.8·L_physics + 0.1·L_boundary + 1e-4·L_reg. Training uses the AdamW optimizer (learning rate 1e-4, cosine annealing), with 3000 training cycles and 512 batches. Validation was performed on measured data from a power company, with CHI predicting a MAPE of 8.7% and RUL predicting a MAPE of 15.3%.

[0054] Example 3: As Figure 3 As shown, the Comprehensive Health Index (CHI) is calculated using a four-index weighted fusion method. The Mechanical Health Index (MHI) is based on vibration and acoustic signature features: extracting time-domain features such as vibration RMS, peak value, kurtosis, and waveform factor, as well as energy in the 100Hz (core magnetostriction) and 50Hz (winding fastening) frequency bands; acoustic signature features include the ultrasonic PRPD spectrum, TEO energy operator, and wavelet packet energy entropy. MHI = 100×(1 - D_mech / D_limit), where D_mech is the cumulative mechanical damage index.

[0055] The Electromagnetic Health Index (EHI) is based on magnetic flux and electrical characteristics. Key indicators include the magnetic flux density ratio B_rms / B_rated (>1.05 indicates overexcitation), the magnetic flux harmonic rate THD_B, and the saturation factor K_sat=B_peak / B_rms (>√2 indicates saturation). EHI=100×(1-D_flux / D_limit), where D_flux=Σ[(B_rms / B_rated)^n-1]×Δt (n=4~6).

[0056] The Thermal Health Index (THI) is based on temperature and load characteristics: the hot spot temperature is estimated using IEC 60076-7 as θ_hs = θ_o + H×g, the relative aging rate is L = 2^((θ_hs-98) / 6), and THI = 100×(1-LOLP / 100%).

[0057] Insulation health index (IHI) is based on temperature history and oil chromatographic characteristics: a quantitative mapping between oil chromatogram and degree of polymerization (DP) is established: DP = f(CO, CO2, furfural, H2O, tanδ), IHI = 100 × (DP_current - 250) / (DP_initial - 250).

[0058] Example 4: Figure 4 As shown, the four-level early warning system automatically determines the health level based on the CHI value and matches differentiated operation and maintenance strategies. Normal Level (CHI≥80, Green): All parameters are normal; standard inspections (3-month cycle), routine tests (3-year cycle), full-capacity operation, and regular spare parts inventory are implemented. Caution Level (Yellow, 60≤CHI<80): A single parameter exceeds the caution value or the trend deteriorates; response time is within 7 days; shorten the inspection cycle (1 month), conduct tests ahead of schedule (1 year), limit peak load to <90%, and increase critical spare parts inventory. Abnormal Level (Orange, 40≤CHI<60): RUL<90 days or multiple parameters are abnormal; response time is within 24 hours; intensive inspections (1 week), live-line testing once a month, limit rated load to <80%, and prepare major spare parts. Danger Level (Red, CHI<40): RUL<30 days or critical parameters exceed limits; response time is <5 minutes; daily inspections, immediate power outage tests, switching to a backup transformer or limiting load to <50%, and emergency procurement or mobilization of reserve spare parts are implemented.

[0059] Example 5: Figure 5 As shown, the multi-sensor data fusion processing flow includes six stages. Data Acquisition: Six types of sensors continuously collect data at their respective sampling frequencies: vibration 10kHz, magnetic flux 1kHz, ultrasound 48kHz, temperature 1Hz, electrical 1Hz, and partial discharge 10MHz. Data Aggregation: The edge gateway collects data from each sensor via the Modbus RTU / TCP protocol, performing protocol conversion, time alignment, and data cleaning. Feature Extraction: A six-dimensional feature vector VFTAEH is constructed, extracting 5-10 key features from each dimension. Edge-side Assessment: Anomalies are detected using an isolated forest algorithm, and fault patterns are classified using a lightweight CNN, outputting the local status. Cloud-based Assessment: The PINN model performs deep coupled calculations, outputting CHI, RUL, and maintenance strategies. Strategy Output: Differentiated operation and maintenance strategies are matched based on the four-level health level, automatically generating maintenance work orders and load scheduling suggestions.

[0060] Example 6: If a 10kV substation has 30 S13 series distribution transformers, the system described in this invention will be deployed. After 12 months of operation and verification: (1) Fault warning: 72 hours in advance, two transformer windings were warned of overheating. After power outage inspection, it was confirmed that the low voltage winding had local overheating. Timely handling was carried out to avoid the fault from expanding.

[0061] (2) Health assessment: The accuracy of the CHI assessment was MAPE=8.7%, which was highly consistent with the results of the power outage test.

[0062] (3) Operation and maintenance optimization: Adjust the inspection cycle according to the health status, reduce the annual inspection workload by 35%, and increase the inspection density of abnormal equipment by 300%.

[0063] (4) Economic benefits: Avoiding two power outage losses (approximately RMB 1.2 million), optimizing maintenance strategies to save RMB 220,000 per year, system investment of RMB 2.79 million, and an estimated payback period of 1.5 years.

[0064] The detailed working principles of each level and core module of this system are as follows: I. Sensing Layer Principle: Construction and Mapping of Multi-Dimensional Feature Space. The system begins with the digital mapping of the transformer's physical state. Unlike traditional single-parameter monitoring, this system deploys multiple types of sensors to capture the transformer's "vital signs" from different physical dimensions: Mechanical and electromagnetic sensing (V / F domain): High-frequency acquisition of vibration and magnetic flux signals is achieved by using multi-parameter fusion sensors installed at the oil tank and casing.

[0065] Principle: The magnetostrictive effect of the iron core will produce vibrations at a specific frequency (such as 100Hz); loosening or deformation of the winding will cause an increase in the 50Hz component in the vibration spectrum. By extracting the vibration RMS, kurtosis, and magnetic flux harmonics (THD_B), the system can infer the iron core tightness and the mechanical state of the winding.

[0066] Acoustics and partial discharge sensing (A domain): utilizing ultrasonic waves and high-frequency current sensors (HFCT).

[0067] Principle: Partial discharges occurring inside insulation generate ultrasonic waves and high-frequency pulsed currents. Through PRPD (phase-resolved pulsed discharge) spectrum analysis, the system can identify the type and severity of insulation defects.

[0068] Thermal and electrical sensing (T / E domain): via IoT meters and infrared sensors.

[0069] Principle: Based on the IEC 60076-7 standard, the winding hot spot temperature (θ_hs) is calculated using load current, ambient temperature and oil temperature data; at the same time, voltage and current harmonics (THD-I) are monitored to assess the thermal shock of electrical stress on the equipment.

[0070] Insulation chemical sensing (I-domain basis): via oil chromatography data.

[0071] Principle: As insulating paper ages, it decomposes and produces specific gases (such as CO, etc.). (furfural). By detecting the concentration of these dissolved gases and combining them with temperature history, the degree of polymerization (DP) of the insulating paper can be inferred, i.e., the remaining amount of insulation life.

[0072] II. Edge Computing Layer Principles: Real-time Filtering and Rapid Response at the Edge. To address the issue of high latency in cloud-based analytics, the edge layer acts as a "sentinel," performing initial processing at the data source: Anomaly Detection Principle (Isolation Forest): The AI ​​chip runs the Isolation Forest algorithm, which isolates data points by randomly cutting the feature space. Because anomalous data (such as sudden vibration changes caused by a short circuit) is usually easy to isolate and has a short path, the algorithm can quickly identify abnormal states that deviate from the normal operating mode.

[0073] Lightweight classification principle (lightweight CNN): Once an anomaly is detected, the lightweight convolutional neural network (CNN) immediately extracts and classifies the waveform features of the abnormal data, determines the type of fault (such as mechanical loosening, overheating, etc.), and issues a local warning within 5 seconds.

[0074] Data triggering mechanism: High-frequency data upload will only be initiated when the end side determines that it is abnormal or when in-depth analysis is required according to the policy, which greatly reduces bandwidth pressure and cloud computing load.

[0075] III. Cloud-based Intelligent Platform Layer Principles: In-depth Evaluation of Physical-AI Fusion. The cloud serves as the "brain" of the system, utilizing high-performance computing resources for complex mechanistic model analysis and deep learning inference.

[0076] (1) The PINN physical information neural network fusion principle is the core innovative principle of this system. Traditional pure data-driven AI models lack physical interpretability, while traditional mechanistic models struggle to handle complex nonlinear relationships. PINN combines the two in the following ways: Network learning: Neural networks learn the mapping relationship between inputs (vibration, temperature, etc.) and outputs (health index) through training data (historical sensor data).

[0077] Physical constraints (L_physics in the Loss Function): During training, the network output must satisfy physical laws. For example: Electromagnetic constraints: must conform to Maxwell's equations.

[0078] Thermal constraint: The temperature distribution must conform to the heat conduction equation.

[0079] Aging constraints: The insulation aging rate must conform to the Arrhenius equation and the flux-temperature coupling equation.

[0080] Results: Even in the case of missing data, the model can still output logical evaluation results based on physical laws, greatly improving the reliability of the results.

[0081] (2) Magnetic flux-temperature coupled aging principle: Traditional insulation life assessment only considers temperature. This invention introduces magnetic flux density BB as an accelerating factor: Synergistic effect: When the transformer is over-excited (magnetic flux density BB increases), the iron core heats up and magnetostrictive vibration intensifies, which mechanically damages the insulating paper fibers, while the high temperature accelerates the breakage of cellulose chains.

[0082] Equation implementation: The system calculates the rate of decrease in insulation polymerization in real time using the formula dDP / dt = -A•exp(-Ea / kT)•B^α•DP^β. A higher magnetic flux BB results in a faster aging rate, thus quantifying the impact of harmonics and overvoltages on lifetime.

[0083] (3) The quantification principle of the Comprehensive Health Index (CHI) is that the system transforms the complex physical state into a scalar value of 0 to 100: Multidimensional integration: The health indices of four dimensions—mechanical (MHI), electromagnetic (EHI), thermal (THI), and insulation (IHI)—are integrated according to their weights.

[0084] Trend quantification: For example, EHI calculates the accumulated overexcitation damage by integrating Σ[(B_rms / B_rated)^n-1]×Δt; IHI calculates the remaining lifetime percentage by monitoring the distance between the current aggregation degree DP and the initial DP.

[0085] Tiered output: The final CHI value intuitively corresponds to four levels: "normal, attention, abnormal, and dangerous", transforming obscure physical parameters into decision-making basis that operations and maintenance personnel can understand.

[0086] (4) Weibull Remaining Life (RUL) Prediction and Uncertainty Quantification.

[0087] Probability distribution: The failure probability distribution function R(t) of the equipment is described by the two-parameter Weibull distribution.

[0088] Dynamic parameters: The distributed parameter εη (characteristic lifetime) is not fixed, but is dynamically updated with the current degree of aggregation and the historical temperature integral.

[0089] Confidence interval: By using Monte Carlo stochastic simulation (simulating 10,000 possible operating scenarios), considering the random fluctuations of parameters, the lifetime prediction interval with a 95% confidence level is output, making the prediction more robust.

[0090] IV. Closed-Loop Feedback Principle: Differentiated Operation and Maintenance Decisions. The system's final output is not just numerical data, but action instructions: Policy matching: Based on the CHI level, the system automatically matches the pre-set operation and maintenance policy library. For example, if the CHI drops to the "abnormal" level (<60), the system automatically generates work orders for "limiting the load to <80%" and "inspection within one week".

[0091] Closed-loop optimization: After maintenance personnel perform maintenance, they feed back the measured results (such as the actual state after the core inspection) to the cloud to calibrate the parameters of the PINN model and realize the model's self-evolution.

[0092] In summary, this online detection and health assessment system for distribution transformers transforms multi-physical field signals, such as mechanical vibration, electromagnetic field changes, thermal effects, and chemical reactions during transformer operation, into six-dimensional digital features. It achieves real-time "filtering" through edge computing and utilizes a cloud-based "physical + AI" dual-drive model for deep decoupling and fusion. Ultimately, it quantifies these signals into a health index to guide differentiated operation and maintenance decisions, forming a closed-loop ecosystem of "monitoring-assessment-decision-feedback".

[0093] The above description is only a specific embodiment of the present invention, but the structural features of the present invention are not limited thereto. Any changes or modifications made by those skilled in the art within the scope of the present invention are covered by the patent scope of the present invention.

Claims

1. An online detection and health assessment system for distribution transformers, characterized in that, include: Device sensing layer, edge computing layer, and cloud intelligent platform layer; The device's sensing layer deploys multi-parameter fusion sensors to achieve real-time five-dimensional sensing of vibration, magnetic flux, temperature, noise, and ultrasound. It is also equipped with an IoT smart meter, a partial discharge current sensor, and an infrared temperature sensor to construct a six-dimensional feature space. The edge computing layer includes an AI chip and an edge gateway built into the multi-parameter fusion sensor. The AI ​​chip is used to realize real-time anomaly detection on the edge side, and the edge gateway is used to realize data aggregation, protocol conversion and encrypted cloud upload. The cloud-based intelligent platform layer includes a magnetic flux-temperature coupled aging model, a temperature-oil chromatography chemical aging model, a PINN physical information neural network fusion model, a comprehensive health index CHI calculation engine, a Weibull remaining life prediction model, and a differentiated operation and maintenance decision engine.

2. The online detection and health assessment system for distribution transformers according to claim 1, characterized in that, The Comprehensive Health Index (CHI) is calculated using a four-index fusion formula, as follows: CHI=0.25×MHI+0.25×EHI+0.25×THI+0.25×IHI; Among them, MHI is the Mechanical Health Index, calculated based on vibration and acoustic signature characteristics; EHI is the Electromagnetic Health Index, calculated based on magnetic flux and electrical characteristics; THI is the Thermal Health Index, calculated based on temperature and load characteristics; and IHI is the Insulation Health Index, calculated based on temperature history and oil chromatographic characteristics. The CHI value ranges from 0 to 100, corresponding to four health levels: normal (CHI≥80), caution (60≤CHI<80), abnormal (40≤CHI<60), and dangerous (CHI<40).

3. The online detection and health assessment system for distribution transformers according to claim 2, characterized in that, The electromagnetic health index EHI takes into account the magnetic flux density ratio and harmonic effects, and is calculated using the formula: EHI = 100 × (1 - D_flux / D_limit); where D_flux is the overexcitation cumulative damage index, D_flux = Σ[(B_rms / B_rated)^n-1] × Δt, B_rms is the effective value of magnetic flux density, B_rated is the rated magnetic flux density, n is the exponential factor with a value range of 4 to 6, Δt is the time interval, and D_limit is the damage limit threshold.

4. The online detection and health assessment system for distribution transformers according to claim 2, characterized in that, The Insulation Health Index (IHI) is calculated based on the quantitative mapping relationship between oil chromatography data and the degree of polymerization (DP). The formula is: IHI = 100 × (DP_current - DP_failure) / (DP_initial - DP_failure); where DP_failure is the alarm threshold, with a value of 250; DP_initial is the initial degree of polymerization; and DP_current is the current degree of polymerization, determined by the oil chromatography data mapping function DP = f(CO, furfural The value is estimated by tanδ (temperature history integral).

5. The online detection and health assessment system for distribution transformers according to claim 1, characterized in that, The flux-temperature coupled aging model comprehensively considers the synergistic effects of core flux density B and winding hot spot temperature θ_hs on insulation aging. The core equation is: dDP / dt =-A•exp(-Ea / kT)•B^α•DP^β; where DP is the degree of insulation polymerization, A is the pre-exponential factor, Ea is the activation energy, k is the Boltzmann constant, T is the absolute temperature, α is the magnetic susceptibility index, ranging from 0.8 to 1.2, and β is the aging level, ranging from 0.9 to 1.

1.

6. The online detection and health assessment system for distribution transformers according to claim 1, characterized in that, The PINN physical information neural network fusion model embeds electromagnetic constraints, thermal constraints, aging constraints, and oil chromatography constraints into the loss function. The composite loss function is: L_total = •L_data+ •L_physics+ •L_boundary+ • L_reg; where L_data is the data loss, L_physics is the physical constraint loss, L_boundary is the boundary condition loss, and L_reg is the regularization loss; These are the corresponding weighting coefficients.

7. The online detection and health assessment system for distribution transformers according to claim 6, characterized in that, The backbone network of the PINN physical information neural network fusion model adopts a 5-layer fully connected structure, with the node configuration as FC128→FC256→FC512→FC256→FC128. The activation function is Swish, combined with residual connections and LayerNorm regularization. The physical constraint layer encodes Maxwell's equations, the heat conduction equation, the Arrhenius-Coupled aging equation, and the gas generation rate equation as regularization terms.

8. The online detection and health assessment system for distribution transformers according to claim 1, characterized in that, The system employs a two-level health status assessment engine. The first level is real-time assessment at the edge, where an isolated forest anomaly detection algorithm and a lightweight CNN fault classification model are run on the AI ​​chip built into the multi-parameter fusion sensor, outputting the operating status and severity level with a response time of less than 5 seconds. The second level is cloud-based fusion assessment, where a PINN physical information neural network fusion model is run on the cloud intelligent platform layer for in-depth assessment, outputting the comprehensive health index (CHI) and remaining life prediction results. When an anomaly is detected at the edge, a data uploading mechanism is triggered, uploading high-frequency sampling data from the anomaly period to the cloud intelligent platform layer.

9. The online detection and health assessment system for distribution transformers according to claim 1, characterized in that, The Weibull remaining lifetime prediction model uses a two-parameter Weibull distribution function: R(t)=exp[-(t / η)^β]; where η is the characteristic lifetime, which is a function of the current degree of aggregation, temperature history integral, and load rate; β is the shape parameter, which depends on the aging acceleration factor and maintenance history; the model combines Monte Carlo stochastic simulation to consider parameter uncertainty, with N≥10000 simulations, and outputs the confidence interval of the remaining lifetime RUL.

10. The online detection and health assessment system for distribution transformers according to claim 1, characterized in that, The six-dimensional feature space includes: V-vibration domain features, F-magnetic flux domain features, T-temperature domain features, A-acoustic signature domain features, E-electrical domain features, and H-history domain features; V-Vibrational Domain Characteristics: including vibration RMS, peak value, kurtosis, 100Hz frequency band energy, and 50Hz frequency band energy; F-Fluid domain characteristics: including B_rms, flux harmonicity THD_B, saturation factor K_sat, and overexcitation accumulation D_flux; T-temperature domain characteristics: including hot spot temperature θ_hs, temperature rise rate dT / dt, and cumulative damage index CDI; A-Acoustic domain features: including PRPD spectrum, TEO energy operator, and wavelet entropy; E-Electrical Domain Characteristics: including load factor, current harmonic distortion rate (THD-I), active power (P), reactive power (Q), and power factor; H-History domain characteristics: including parameter trends and cumulative damage.

Images