Transformer fault detection method and system based on multi-parameter feature fusion

Abstract

The application discloses a transformer fault detection method and system based on multi-parameter feature fusion, which comprises the following steps: a gas detection unit and a micro-flow particle detection unit are arranged in the same transformer insulation oil circulation bypass, and the dissolved gas feature parameters and the oil particle feature parameters of the insulation oil are synchronously and timely collected; the dissolved gas feature parameters and the oil particle feature parameters are time-stamped and aligned and pretreated, and time-synchronized multi-dimensional time sequence feature data are generated; the multi-dimensional time sequence feature data are input into a trained fault detection model, the gas feature sub-vector and the particle feature sub-vector are fused and analyzed by the fault detection model, and a transformer fault detection result is output. Through synchronous detection and fusion analysis of the gas feature parameters and the particle feature parameters in the insulation oil, a multi-parameter joint feature model is constructed, the problems of traditional DGA technology detection lag and single-dimensional feature limitation are broken through, and the accuracy of fault identification and the early warning capability are improved.

Description

Technical Field

[0001] This invention relates to the field of transformer fault diagnosis technology, and in particular to a transformer fault detection method and system based on multi-parameter feature fusion. Background Technology

[0002] Power transformers are core equipment in power systems, and their operating status directly affects the safety and stability of the entire power grid. A transformer failure often triggers a chain reaction of power outages and equipment damage, causing severe economic losses and systemic safety risks. Insulating oil is not only the main insulating medium in transformers but also plays a crucial role in heat dissipation and cooling.

[0003] Currently, the most widely used transformer condition assessment method in the engineering field is Dissolved Gas Analysis (DGA). This method identifies insulation fault types such as internal discharge and overheating by monitoring the concentration changes and ratios of characteristic gases such as hydrogen (H2), acetylene (C2H2), and methane (CH4). After years of development, DGA technology has become an important technological foundation for power equipment condition monitoring. However, in practical applications, it still suffers from problems such as detection lag, system complexity, and limited diagnostic accuracy. Because the generation, diffusion, and dissolution processes of gases in insulating oil have significant time dependence, the detection results often cannot reflect the transient state changes of the equipment in real time, especially in the early stages of a fault, where weak discharge or overheating signals may be delayed or even masked. Furthermore, traditional DGA systems typically require multiple steps, including oil-gas separation, gas extraction, and quantitative analysis, resulting in complex system structures, long detection cycles, and high maintenance costs, making it difficult to meet the real-time, accurate, and continuous monitoring requirements of smart grids for transformer operating status.

[0004] During the manufacturing, transportation, installation, maintenance, and operation of transformers, impurity particles inevitably become mixed into the insulating oil. These particles easily migrate and accumulate in high-field regions under strong electric fields, leading to localized electric field distortion and inducing partial discharge, insulation aging, and even insulation breakdown. The quantity, size, and morphological distribution of particles in the oil are important indicators reflecting the health status of the transformer insulation system. Although current standards specify the concentration and size range of particles in the oil, existing particle detection methods mostly rely on offline sampling and laboratory analysis. These methods require manual intervention, are highly periodic, and lack real-time accuracy. This traditional approach struggles to reflect the dynamic changes in equipment during operation, easily leading to the neglect of latent degradation signals and allowing potential faults to accumulate and worsen undetected.

[0005] With the continuous expansion of power grid scale and the constant improvement of its intelligence level, the demand for real-time online monitoring of power equipment is becoming increasingly prominent. Relying solely on dissolved gas analysis is no longer sufficient to meet the comprehensive diagnostic needs of complex, multi-source fault mechanisms. Changes in oil particle size, as an important parameter reflecting insulation degradation and oil quality deterioration, are significantly complementary to gas characteristics. Therefore, there is an urgent need to develop an online condition monitoring method for transformers that can integrate dissolved gas and particle characteristic information in oil. This method would enable the simultaneous detection and fusion analysis of gas and particle characteristic parameters in insulating oil, construct a multi-parameter fault characteristic detection model, overcome the limitations of traditional DGA technology in detection lag and single-parameter characteristics, and improve the sensitivity and accuracy of transformer fault diagnosis and prediction.

[0006] Chinese patent application number 202311356012.3 proposes a transformer fault diagnosis method based on variable-weighted VAE and dual-channel deep feature fusion. It uses single-parameter data of dissolved gases in oil and achieves fault classification through variable-weighting, VAE data augmentation, and a CNN-BiLSTM-Attention model. This invention represents an algorithmic optimization of traditional DGA data, but it does not solve the problems of detection lag, poor real-time performance, and insufficient feature dimensions in existing DGA methods.

[0007] Chinese patent application number 202510289548.0 provides a method for classifying and identifying impurity particles in oil. It employs fluid flow dynamic imaging technology combined with a YOLOv9 network to identify and statistically analyze impurities such as fiber particles, carbon particles, and copper particles. This invention focuses on particle image detection, and the results are used to assist in condition assessment. However, it does not address multi-parameter fusion mechanisms or online fault prediction, thus failing to overcome the limitations of single-dimensional features. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a transformer fault detection method and system based on multi-parameter feature fusion. To solve the problems of single parameters, detection lag, and inability to achieve real-time monitoring in existing power transformer fault diagnosis technologies, this invention constructs a multi-parameter fault detection model by synchronously detecting and fusing the gas and particle characteristic parameters in insulating oil. This overcomes the detection lag and single-dimensional feature limitations of traditional DGA technology, improving the accuracy of fault identification and early warning capabilities.

[0009] The present invention adopts the following technical solution. The first aspect of the present invention provides a transformer fault detection method based on multi-feature parameter fusion, comprising the following steps: Step 1: By using a gas detection unit and a microfluidic particle detection unit located in the same transformer insulating oil circulation bypass, the characteristic parameters of dissolved gas and the characteristic parameters of particles in the oil are collected synchronously and in real time. Step 2: The dissolved gas characteristic parameters and oil particle characteristic parameters are time-stamp aligned and preprocessed to generate time-synchronized multi-dimensional time-series characteristic data. The time-series characteristic data of each time step consists of gas characteristic sub-vectors and particle characteristic sub-vectors. Step 3: Input the multidimensional time-series feature data into the trained fault detection model. The fault detection model then performs a fusion analysis on the gas feature vector and the particle feature vector, and outputs the transformer fault detection result.

[0010] Further, the preprocessing in step 2 includes: Based on the operating parameters of the gas detection unit and the image parameters obtained by the microfluidic particle detection unit, data quality weights for the corresponding time steps are generated. Missing or outlier values ​​in the dissolved gas characteristic parameters and oil particle characteristic parameters are marked with missing tags.

[0011] Furthermore, the fault detection model described in step 3 includes: The encoding layer constructed by the Long Short-Term Memory network includes a gas feature sub-encoder and a particle feature sub-encoder. The gas feature sub-encoder is used to encode the gas temporal feature data and output the gas hidden state sequence; the particle feature sub-encoder is used to encode the particle temporal feature data and output the particle hidden state sequence. The attention layer performs attention weighting on the encoded temporal features along the time dimension. The decoding layer, constructed from a Long Short-Term Memory network, generates a comprehensive state representation based on the weighted features from the attention layer; and, The output layer outputs the transformer fault detection results based on the comprehensive state representation.

[0012] Furthermore, the fault detection model also includes a gated fusion layer connected between the coding layer and the attention layer, which is used to receive the gas hidden state sequence, the particle hidden state sequence and the data quality weight. At each time step, the fusion weight of the gas features and the fusion weight of the particle features are calculated through a gating mechanism to achieve adaptive weighted fusion and generate a fused feature sequence.

[0013] Furthermore, The gated fusion layer performs the following operations at each time step: The gas hidden state, particle hidden state, and data quality weights are combined to generate a combined feature vector; The combined feature vector is input into a trainable gating unit to generate normalized gas feature fusion weights and particle feature fusion weights. By using the fusion weights of gas features and particle features, the hidden states of gas and particles at the current time step are weighted and summed to obtain the fusion features for that time step.

[0014] Furthermore, the output layer includes three independent output branches, specifically including: Output transformer fault type classification results; The output reflects the health status index of the transformer's operating condition; In addition, the predicted value of the future operating trend of the output transformer.

[0015] Furthermore, it also includes step 4, which involves post-processing the transformer fault detection results, specifically including: The transformer fault type classification results are subjected to probability calibration; Hysteresis judgment is performed on the pressure vessel failure probability value based on preset entry threshold, exit threshold and duration; The transformer fault type is then checked for consistency with the diagnostic results obtained based on the dissolved gas ratio method.

[0016] A second aspect of the present invention provides a transformer fault detection system based on multi-feature parameter fusion, specifically including: Edge acquisition module: Deployed at the transformer site, used to synchronously and in real time acquire characteristic parameters of dissolved gas in insulating oil and characteristic parameters of particles in the oil; Data communication module: used to realize the real-time transmission of data collected by the insulation acquisition module; Cloud analytics and storage module: Used to receive, store, and analyze real-time data from the edge acquisition module; Prediction and Diagnosis Module: Based on the fault detection model, it outputs the detection results of the transformer's real-time status; User client display module: Used to receive the transformer status detection results output by the prediction and diagnosis module, and to perform visualization display and alarm management.

[0017] A third aspect of the present invention provides a terminal, including a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of a transformer fault detection method based on multi-feature parameter fusion.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when executed by a processor, the program implements the steps of a transformer fault detection method based on multi-feature parameter fusion.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates a gas detection unit and a microfluidic particle detection unit on the transformer oil circulation bypass. Through a unified fluid control and data acquisition mechanism, it achieves real-time synchronous acquisition of characteristic parameters of dissolved gases and particles in the insulating oil. By synchronously detecting and fusing gas and particle characteristics, the fault evolution of the transformer can be reflected from two complementary physical dimensions. Compared with the traditional DGA or particle detection process of "manual sampling + laboratory analysis", it shortens the sampling, separation and analysis link, allowing the insulating oil to continuously flow through the detection unit in the operating state, realizing real-time status capture and reducing errors caused by sample retention and environmental changes.

[0020] This invention constructs a fault detection model for multi-dimensional time-series feature data of "gas + particles". It uses a front-end LSTM encoding layer to extract deep time-series features to alleviate the gradient vanishing problem in long sequences. In the middle, gating and attention mechanisms are introduced to weight the hidden states at different times to obtain context vectors reflecting key periods of fault evolution. Then, the back-end LSTM decoding and output layers output transformer fault type classification results, a health status index reflecting the transformer's operating condition, and a predicted value for the transformer's future operating trend. This elevates transformer fault detection from "static classification" to a diagnostic model integrating trend prediction, early warning, and health management. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the transformer fault detection method of the present invention; Figure 2 This is a schematic diagram of the transformer fault prediction model structure of the present invention; Figure 3 This is a schematic diagram of the transformer fault detection system of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0023] Reference Figure 1 and Figure 2 Embodiment 1 of the present invention provides a transformer fault detection method based on multi-feature parameter fusion, specifically including the following steps: Step 1: By setting up a gas detection unit and a microfluidic particle detection unit in the same transformer insulating oil circulation bypass, the characteristic parameters of dissolved gas and particulate matter in the insulating oil are collected synchronously and in real time.

[0024] Specifically, a gas detection unit and a microfluidic particle detection unit are installed in the same oil circulation bypass pipeline of the insulating oil. The insulating oil passes through the gas detection unit and the microfluidic particle detection unit in sequence under constant flow. The two detection units share the same oil circuit, flow rate, temperature and pressure conditions.

[0025] The gas detection unit employs sensing methods suitable for online monitoring, such as non-dispersive infrared absorption (NDIR), tunable laser absorption spectroscopy (TDLAS), or solid-state electrochemical sensing, to detect characteristic gases such as hydrogen (H2), acetylene (C2H2), and methane (CH4) in oil samples in real time. The instantaneous concentration values ​​of the corresponding gases are obtained by measuring changes in absorption intensity at characteristic wavelengths, cavity-enhanced absorption signals, or the electrochemical response of the sensitive membrane. Noise filtering, temperature and pressure compensation, and standardization are performed on the original concentration sequence. Simultaneously, the rate of gas change is calculated based on continuous sampling to enhance sensitivity to sudden discharges or thermal faults. Based on the generation mechanism characteristics of gases in the oil, preliminary judgments of internal transformer faults are made through gas ratios and trend changes, including CH4 / H2, C2H6 / CH4, C2H4 / C2H6, and C2H2 / C2H4.

[0026] The microfluidic particle detection unit captures particle images through pulsed light source illumination and a high-speed camera. Based on image recognition algorithms such as adaptive threshold segmentation, edge detection, and morphological feature analysis, it extracts morphological feature parameters such as particle size distribution, number concentration, roundness, and aspect ratio. The acquired raw images undergo illumination equalization and noise suppression processing to eliminate local brightness differences caused by the pulsed light source and background noise caused by oil flow disturbances. Subsequently, a segmentation method based on grayscale gradient or Otsu adaptive thresholding is used to separate the particles from the background, and the particle edge contours are extracted using Canny or Sobel algorithms. Based on contour recognition, the system uses minimum circumcircle and minimum circumcircle ellipse fitting methods to calculate parameters such as particle size, major and minor axis dimensions, roundness, aspect ratio, and shape skewness. Simultaneously, by statistically analyzing the number density of particles within the detection window, it obtains counting features reflecting the degree of contamination, characterizing the changing trend of the transformer's internal insulation state.

[0027] This approach utilizes a combined labeling method based on operational records, maintenance results, and experimental fault tests. For in-service transformers, the corresponding state categories at each moment are determined: based on operational logs recorded by the maintenance unit, such as discharge accidents, gas relay operation events, abnormal temperature rise records, and electrical test conclusions, the corresponding time periods are labeled as normal, partial discharge, mild overheating, and severe overheating. Based on the actual defects found in tap changers, windings, and insulation paperboard after power outage maintenance, the corresponding historical monitoring data are traced and labeled as the corresponding defect types. For experimental platforms, the response process of gas / particles in the oil can be recorded by applying controlled discharge power, particulate contamination sources, or thermal fault sources, constructing an experimental dataset with clearly defined fault type labels. State labels include: Normal, Partial Discharge (PD), Low Thermal Fault, High Thermal Fault, Insulation Aging, and Carbonization. Alternatively, binary labels (normal / abnormal) or multi-level health level labels can be used depending on the specific application scenario.

[0028] Step 2: The dissolved gas characteristic parameters and oil particle characteristic parameters are time-stamp aligned and preprocessed to generate time-synchronized multidimensional time-series characteristic data. The time-series characteristic data of each time step consists of gas characteristic sub-vectors and particle characteristic sub-vectors.

[0029] Specifically, the collected dissolved gas characteristic parameters and oil particle characteristic parameters undergo preprocessing including timestamp alignment, data cleaning, missing value imputation, and feature scale normalization. The data is then sliced ​​according to a preset time window length to construct a three-dimensional training sample set with dimensions [number of samples, time step, number of features], as shown in the formula: ; in, For the sample size, For time step, For characteristic number, In time step The fused feature vector, ; ; For gas feature vectors; For particle feature vectors; ; in, For the first Gas characteristics at each time step For the first Particle features at each time step For the dimensions of gas characteristics, Dimension of the particle features.

[0030] Furthermore, based on the operating status parameters of the gas detection unit and quality indicators such as image clarity, background noise level, and insulating oil flow stability acquired by the microfluidic imaging particle detection unit, data quality weights are generated for corresponding time steps. Simultaneously, for dissolved gas characteristic parameters and oil particle characteristic parameters, missing or abnormal data that remain after data cleaning and imputation are marked with missing tags.

[0031] Step 3: Input the multidimensional time-series feature data into the trained fault detection model. The fault detection model then performs a fusion analysis on the gas feature vector and the particle feature vector, and outputs the transformer fault detection result.

[0032] Specifically, the fault detection model includes: The encoding layer constructed by the Long Short-Term Memory network includes a gas feature sub-encoder and a particle feature sub-encoder. The gas feature sub-encoder is used to encode the gas temporal feature data and output the gas hidden state sequence; the particle feature sub-encoder is used to encode the particle temporal feature data and output the particle hidden state sequence.

[0033] Specifically, the gas feature sub-encoder uses gas feature sequences. Using the forget gate, input gate, and output gate as input, a recursive calculation outputs a gas hidden state sequence that encodes the long-term dependence of the thermal decomposition and discharge chemical processes of oil-paper insulation reflected by changes in gas concentration and ratio relationships. .

[0034] Particle feature sub-encoder with particle feature sequence As input, through recursive calculations involving forget gates, input gates, and output gates, the output is a sequence of hidden particle states that encodes the dynamic evolution patterns of physical processes such as mechanical wear, insulation degradation, and contaminant accumulation reflected by changes in particle number and morphology. .

[0035] The gated fusion layer receives the gas hidden state sequence output by the gas feature sub-encoder, the particle hidden state sequence output by the particle feature sub-encoder, and the data quality weights at each time step. Based on the gas hidden state output by the gas feature sub-encoder, the particle hidden state output by the particle feature sub-encoder, and the data quality weight, the fusion weight of gas features and particle features in each time step is calculated through a gating mechanism to achieve adaptive weighted fusion of gas time series features and particle time series features, generating a fused feature sequence.

[0036] Specifically, the hidden states of gas, particles, and data quality weights are concatenated. In this embodiment, the trainable gating unit is a learnable multilayer perceptron (MLP) and a softmax function. Normalized fusion weights are generated through the trainable gating unit, using the following formula: ; ; in, The gating coefficient for gas features in the current time step fusion. Let be the gating coefficient of the particle features in the fusion at the current time step, and , Data quality weights for particle image parameters. Data quality weights for gas operating parameters.

[0037] If the gas operating parameter data for the corresponding time step is missing, then let If particle image parameter data is missing, then let Then, normalize it again.

[0038] Based on the gas hidden state, particle hidden state, and data quality weights at the current time step, a weighted fusion calculation is performed to obtain the fused features, as shown in the formula: ; The fused feature sequence is obtained as follows: .

[0039] The attention layer applies attention weights to the fused feature sequence passed through the gated fusion layer in the time dimension.

[0040] Specifically, for each time step in the sequence Fusion hidden state Calculate its hidden state compared to the decoder's previous time step. correlation score The formula is: ; in, For trainable attention weight vectors, and For a trainable weight matrix, Indicates the decoding layer at the 1st level The hidden state of the step, Indicates the coding layer at the 1st The hidden state at each time step Trainable bias vectors.

[0041] The relevance scores are normalized using the Softmax function, and the attention weights are calculated using the following formula: ; in, For the first At the decoding moment, the encoder's first decoding moment... The contribution weight of the predicted target at each time step; The context vector is obtained by weighted summation of the hidden states at all time steps of the coding layer based on the attention weights, and the calculation formula is as follows: ; The attention layer enables the fault detection model to automatically identify and focus on key event points in the fault evolution process, overcoming the shortcomings of traditional methods that treat all time points equally, and significantly enhancing the detection sensitivity and diagnostic accuracy of early faults and transient anomalies.

[0042] The decoding layer, constructed from a long short-term memory network, generates a comprehensive state representation based on the weighted features from the attention layer.

[0043] Specifically, taking the context vector and its previous state as input, the state is updated and output to generate the comprehensive state. The calculation process is as follows: ; in, This is the output of the decoder itself from its previous state; The hidden state of the last time step of the decoder is taken as the comprehensive state representation of the entire input sequence, as follows: .

[0044] The output layer, based on the comprehensive state representation, outputs the transformer fault detection result. .

[0045] Specifically, it outputs the probability of transformer fault types, enabling accurate classification of transformer fault types such as transformer normal, partial discharge, low temperature overheating, and insulation aging. The output is a continuous scalar between 0 and 1, reflecting the health index of the transformer's operating status, and realizing a quantitative and continuous assessment of the insulation health of the transformer. by Given the initial state, the predicted value of the transformer's future operating trend is output through a fully connected sequence.

[0046] Step 4: Post-process the transformer fault detection results.

[0047] Specifically, the Platt scaling method is adopted, and a logistic regression mapping is learned using the validation set to correct the systematic bias of the original output probability of the model and improve the reliability of probability threshold-based decision-making.

[0048] Set entry threshold Exit threshold and minimum duration , for The fault status at any given time, 0 for normal and 1 for fault. for The fault category probability after time calibration, and the fault state determination follow the following: like , And the duration is greater than or equal to ,but ; like , And the duration is greater than or equal to ,but ; If the above conditions are not met, The transformer's state remains unchanged.

[0049] The transformer fault type is compared with the diagnostic results obtained based on the dissolved gas ratio method. If the two diagnostic results are consistent, the transformer fault type classification result of the fault detection model is output; if the two diagnostic results are inconsistent, a review prompt message containing conflict markers is output.

[0050] The LSTM encoding layer can memorize evolutionary information over a long time span, enabling the model to comprehensively consider the trajectory of changes over a period of time when judging the current state, thereby improving the ability to identify chronic faults and gradual degradation patterns. The attention mechanism adaptively weights the latent states at different times. When abnormal combinations such as "gas mutation + granularity surge" occur within certain time windows, the corresponding weights are significantly amplified, causing the model to automatically focus on time segments more sensitive to fault evolution, enhancing its response to latent faults and sudden anomalies. The decoding layer outputs state parameters and health status indices for several future time points based on the weighted temporal representation, enabling quantitative assessment of the transformer insulation status and prediction of future degradation trends, thus providing early warning before insulation breakdown or partial discharge deterioration. Therefore, this LSTM-Attention-LSTM model elevates this invention from "static classification" to a diagnostic model integrating trend prediction, early warning, and health management, significantly different from existing methods that only classify and identify DGA or voiceprint signals.

[0051] Reference Figure 3 Embodiment 2 of the present invention provides a transformer fault detection system based on multi-feature parameter fusion, specifically including: Edge acquisition module: Deployed at the transformer site, used to synchronously and in real time acquire characteristic parameters of dissolved gas in insulating oil and characteristic parameters of particles in the oil; Data communication module: Used to realize the real-time transmission of data collected by the insulation acquisition module; communication methods can be adopted through Ethernet, 5G, LoRa, or fiber optic communication, ensuring the stability and low latency of data transmission. The module has built-in encryption algorithms and data verification mechanisms to ensure the security and integrity of data transmission. The system can dynamically adjust the upload frequency according to changes in the transformer's operating status, achieving adaptive communication and bandwidth optimization.

[0052] The cloud-based analysis and storage module is responsible for receiving, storing, and analyzing real-time data from the edge acquisition module. The system uses a time-series database for unified management of multi-source data and employs feature fusion algorithms for multi-dimensional analysis of gas and particle parameters. The cloud module can perform trend fitting, anomaly detection, and feature correlation analysis, enabling long-term tracking and health assessment of transformer condition changes. All detection data and results can be automatically generated into historical curves and operational reports in the cloud, providing a basis for maintenance decisions.

[0053] Prediction and Diagnosis Module: This module constructs a transformer fault prediction model based on neural network or deep learning algorithms. By inputting fused gas and particle feature vectors, the fault detection model can automatically identify different types of faults and output a state health index and future trend prediction results. This module supports model self-learning and parameter updates, and can be personalized for training and optimization according to the different operating characteristics of transformers, thereby improving the accuracy and robustness of fault diagnosis. The user client display module is used to visualize and remotely interact with monitoring results. Users can view the real-time operating status of the transformer, gas and particle characteristic curves, alarm records, and prediction results through a computer, mobile terminal, or control center interface. The system supports multi-level access control, allowing different data access and control operations based on user roles. When monitoring results exceed set thresholds, the client automatically generates an alarm prompt and can trigger on-site audible and visual alarm devices or SMS notification systems.

[0054] By dividing the functions of edge acquisition and cloud analysis, preprocessing and basic judgment, which have high real-time requirements, can be completed locally on-site, while the computationally intensive deep learning models can be deployed to the cloud. This ensures the stability of the acquisition side and improves the overall computing power and model update capabilities. With a unified time-series database and feature fusion algorithm, the system can record and trace back operational data over long periods, supporting trend analysis, fault review, and long-term health assessment. Through a visual client and intelligent alarm mechanism, maintenance personnel can promptly grasp the equipment status before and after anomalies occur, supporting condition-based maintenance (CBM) and maintenance strategy optimization, avoiding reliance solely on scheduled maintenance or reactive repairs. Simultaneously, this closed-loop architecture facilitates future integration of other monitoring parameters such as temperature, load, and current, possessing excellent scalability and cross-device migration capabilities. Thus, this invention is no longer an isolated detection or algorithm solution, but rather forms a complete transformer health management platform spanning "acquisition—analysis—decision-making," significantly enhancing the engineering application value of the technical solution.

[0055] Embodiment 3 of the present invention provides a terminal, including a processor and a storage medium; The storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the transformer fault detection method based on multi-feature parameter fusion.

[0056] Embodiment 4 of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a transformer fault detection method based on multi-feature parameter fusion.

[0057] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0058] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0059] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0060] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A transformer fault detection method based on multi-feature parameter fusion, characterized in that, Specifically, the following steps are included: Step 1: By using a gas detection unit and a microfluidic particle detection unit located in the same transformer insulating oil circulation bypass, the characteristic parameters of dissolved gas and the characteristic parameters of particles in the oil are collected synchronously and in real time. Step 2: The dissolved gas characteristic parameters and oil particle characteristic parameters are time-stamp aligned and preprocessed to generate time-synchronized multi-dimensional time-series characteristic data. The time-series characteristic data of each time step consists of gas characteristic sub-vectors and particle characteristic sub-vectors. Step 3: Input the multidimensional time-series feature data into the trained fault detection model. The fault detection model then performs a fusion analysis on the gas feature vector and the particle feature vector, and outputs the transformer fault detection result.

2. The transformer fault detection method based on multi-feature parameter fusion according to claim 1, characterized in that, Step 2 of the preprocessing includes: Based on the operating parameters of the gas detection unit and the image parameters obtained by the microfluidic particle detection unit, data quality weights for the corresponding time steps are generated. Missing or outlier values ​​in the dissolved gas characteristic parameters and oil particle characteristic parameters are marked with missing tags.

3. The transformer fault detection method based on multi-feature parameter fusion according to claim 1, characterized in that, The fault detection model described in step 3 includes: The encoding layer constructed by the Long Short-Term Memory network includes a gas feature sub-encoder and a particle feature sub-encoder. The gas feature sub-encoder is used to encode the gas temporal feature data and output the gas hidden state sequence; the particle feature sub-encoder is used to encode the particle temporal feature data and output the particle hidden state sequence. The attention layer performs attention weighting on the encoded temporal features along the time dimension. The decoding layer, constructed from a Long Short-Term Memory network, generates a comprehensive state representation based on the weighted features from the attention layer; and, The output layer outputs the transformer fault detection results based on the comprehensive state representation.

4. The transformer fault detection method based on multi-feature parameter fusion according to claim 3, characterized in that, The fault detection model also includes a gated fusion layer connected between the coding layer and the attention layer, which is used to receive the gas hidden state sequence, the particle hidden state sequence and the data quality weight. At each time step, the fusion weight of the gas features and the fusion weight of the particle features are calculated through a gating mechanism to achieve adaptive weighted fusion and generate a fused feature sequence.

5. The transformer fault detection method based on multi-feature parameter fusion according to claim 4, characterized in that, The gated fusion layer performs the following operations at each time step: The gas hidden state, particle hidden state, and data quality weights are combined to generate a combined feature vector; The combined feature vector is input into a trainable gating unit to generate normalized gas feature fusion weights and particle feature fusion weights. By using the fusion weights of gas features and particle features, the hidden states of gas and particles at the current time step are weighted and summed to obtain the fusion features for that time step.

6. The transformer fault detection method based on multi-feature parameter fusion according to claim 3, characterized in that, The output layer includes three independent output branches, specifically: Output transformer fault type classification results; The output reflects the health status index of the transformer's operating condition; In addition, the predicted value of the future operating trend of the output transformer.

7. The transformer fault detection method based on multi-feature parameter fusion according to claim 1 or 6, characterized in that, It also includes step 4, which involves post-processing the transformer fault detection results, specifically including: The transformer fault type classification results are subjected to probability calibration; Hysteresis judgment is performed on the pressure vessel failure probability value based on preset entry threshold, exit threshold and duration; The transformer fault type is then checked for consistency with the diagnostic results obtained based on the dissolved gas ratio method.

8. A transformer fault detection system based on multi-feature parameter fusion, used to implement the transformer fault detection method based on multi-feature parameter fusion as described in any one of claims 1-7, characterized in that, Specifically, it includes: Edge acquisition module: Deployed at the transformer site, used to synchronously and in real time acquire characteristic parameters of dissolved gas in insulating oil and characteristic parameters of particles in the oil; Data communication module: used to realize the real-time transmission of data collected by the insulation acquisition module; Cloud analytics and storage module: Used to receive, store, and analyze real-time data from the edge acquisition module; Prediction and Diagnosis Module: Based on the fault detection model, it outputs the detection results of the transformer's real-time status; User client display module: Used to receive the transformer status detection results output by the prediction and diagnosis module, and to perform visualization display and alarm management.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-7.

Images