Method and system for evaluating insulation aging state of converter transformer

Abstract

The invention provides a method and a system for evaluating the insulation aging state of a converter transformer. The method comprises the following steps: loading multiple physical stresses and constructing an experimental platform; collecting and preprocessing multi-source time sequence data; a DF-BiLSTM model is constructed and trained; evaluating and verifying the aging state; and integrating and deploying the system. The problem that the generalization ability of the model is insufficient is solved; the prediction precision encounters a bottleneck and is not sensitive to the subtle change of the aging state, so that early warning is difficult to realize; the reliability is low, and the judgment of the model lacks robustness and is easily interfered by noise in data; the feature expression capability is limited; the insulation aging time sequence evolution rule cannot be effectively captured, and the state prediction accuracy is influenced.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, specifically to a method and system for assessing the insulation aging condition of a converter transformer. Background Technology

[0002] As a core component of the UHVDC transmission system, the operational reliability of the UHVDC converter transformer directly impacts the safety and stability of the entire power grid. The valve-side outgoing line device, serving as a crucial link connecting the converter valve and the transformer, endures the combined effects of AC / DC voltage, high electric field strength, large current, and complex mechanical stress over extended periods, making it the most vulnerable point for insulation failure. The oil-paper insulation material inside this device is continuously subjected to the combined effects of electrical, thermal, and mechanical fields during operation, leading to irreversible degradation of its insulation performance over time. This inevitable material aging has become a major factor restricting the long-term operational reliability of the equipment. Statistics show that insulation aging is one of the main causes of converter transformer failures. Therefore, accurate assessment and prediction of the insulation status of the valve-side outgoing line device are of paramount importance for ensuring the safe and stable operation of UHVDC transmission projects.

[0003] Although machine learning and deep learning algorithms have been initially applied in the field of power equipment condition assessment, such as using support vector machines (SVM) or ordinary fully connected neural networks (FNN) to perform regression prediction on single feature parameters of insulation materials (such as peak recovery voltage), or using one-way long short-term memory networks (LSTM) to fit trends in time series data, these existing models still have significant limitations when facing the complex problem of insulation aging assessment of valve-side outgoing line devices of UHV converter transformers.

[0004] Conventional neural network models are essentially static mapping models, making it difficult to effectively capture the inherent temporal dependencies in the dynamic evolution of insulation aging. While simple LSTM models can process sequential data, their unidirectional structure means they can only rely on historical information in one direction, failing to fully utilize the value of "future" contextual information in monitoring data for current state judgment. This limits their understanding of the complete evolutionary patterns of the state. More importantly, both conventional neural networks and ordinary LSTMs can be considered "shallow" models with limited feature extraction capabilities. They typically rely on tedious and highly expert-dependent manual feature engineering, making it difficult for them to automatically learn and extract deep, abstract essential features strongly correlated with aging states from raw, high-dimensional, nonlinear, and multi-source heterogeneous data.

[0005] This leads to the following problems with existing models: First, the models lack generalization ability; their performance significantly degrades when equipment operating conditions and environmental factors change. Second, prediction accuracy encounters a bottleneck; they are insensitive to subtle changes in aging conditions, making early warning difficult. Third, reliability is questionable; the models lack robustness and are easily affected by noise in the data. Therefore, existing technologies are insufficient to meet the practical engineering needs for accurate, reliable, and forward-looking condition assessment of ultra-high voltage converter transformers, a critical piece of equipment.

[0006] Therefore, there is an urgent need for an advanced dedicated deep learning architecture to address the dual challenges of automatic feature extraction and deep temporal dependency modeling.

[0007] The invention application document CN119538071A, entitled "A Method for Constructing a Diagnostic Mode for Thermal Aging of Oil-Paper Insulation under High-Frequency Pulsed Voltage," describes an existing method that constructs a high-frequency pulsed partial discharge experimental platform, acquires signals based on a sliding phase window model, and extracts pass-frequency aging features such as phase jumps and phase concentration. Finally, it uses a support vector machine (SVM) optimized by particle swarm optimization to identify aging stages. While this method optimizes the SVM parameters through an optimization algorithm, it still fundamentally relies on the traditional support vector machine model. As a shallow machine learning algorithm, SVM has inherent limitations when processing high-dimensional, nonlinear insulation aging data: First, SVM struggles to effectively capture the complex nonlinear relationships between multi-source heterogeneous data (such as frequency domain dielectric spectra, temperature time series, and electric field distributions); second, SVM is highly dependent on feature engineering, requiring manual feature extraction based on expert experience, and cannot automatically uncover deep temporal patterns and contextual dependencies in the data; third, SVM has high computational complexity and limited generalization ability when processing large-scale high-dimensional data, making it difficult to adapt to the complex and variable operating conditions of converter transformers. These shortcomings result in insufficient accuracy and reliability in identifying insulation aging conditions, which may ultimately affect the accurate assessment of the remaining life of the equipment.

[0008] The invention patent application document CN115372766A, entitled "A Method for Evaluating the Aging Status of Converter Transformer Oil-Paper Insulation Based on Degree of Polymerization," describes an existing method that prepares nano-TiO2 modified insulating paperboard, measures its viscosity characteristics, establishes a degradation model based on the degree of polymerization, and performs reliability analysis. This method is an offline, destructive testing method that requires sampling for laboratory analysis, making real-time online monitoring of equipment status impossible. Furthermore, this method focuses solely on the single chemical indicator of the degree of polymerization, completely ignoring the combined effects of multiple stresses such as electric fields, moisture, and mechanical vibrations on aging during operation. The evaluation results are lagging and one-sided, failing to reflect the true aging state of the insulation material and easily leading to errors in maintenance strategy formulation, resulting in over-maintenance or failure risks.

[0009] Existing assessment methods are mostly based on a single stress aging model, which fails to fully consider the multi-stress coupling effects such as electrical, thermal, mechanical, and moisture stresses, resulting in deviations between assessment results and actual operating conditions. Traditional feature extraction methods rely on human experience and are difficult to adaptively process high-dimensional time-series data such as frequency domain dielectric spectra and temperature fields, thus having limited feature representation capabilities. Existing models lack time-series dynamic modeling capabilities and cannot effectively capture the time-series evolution of insulation aging, affecting the accuracy of state prediction.

[0010] In summary, existing technologies suffer from several technical problems: insufficient model generalization ability; bottleneck in prediction accuracy, insensitivity to subtle changes in aging state, making early warning difficult; low reliability, lack of robustness in model judgment, and susceptibility to noise interference in data; limited feature representation ability; and inability to effectively capture the temporal evolution of insulation aging, thus affecting the accuracy of state prediction. Summary of the Invention

[0011] The technical problems to be solved by this invention are: how to solve the following problems in the existing technology: insufficient generalization ability of the model; bottleneck in prediction accuracy; insensitivity to subtle changes in aging state, making it difficult to achieve early warning; low reliability; lack of robustness in model judgment, and susceptibility to noise interference in data; limited feature expression ability; and inability to effectively capture the temporal evolution law of insulation aging, which affects the accuracy of state prediction.

[0012] This invention solves the above-mentioned technical problems by employing the following technical solution: A method for assessing the insulation aging condition of a converter transformer, comprising:

[0013] S1. Applying multiple physical stresses to construct an experimental platform;

[0014] S2. Using the experimental platform, collect and preprocess multi-source time-series data to obtain real-time monitoring data;

[0015] S3. Construct and train the deep feature extraction-bidirectional long-term memory network model DF-BiLSTM to model the mapping relationship between the aging state of the converter transformer and the temperature and electric field distribution.

[0016] S4. Input the real-time monitoring data into the trained deep feature extraction-bidirectional long-term memory network model DF-BiLSTM to perform aging state assessment and verification, and output the insulation aging state assessment results.

[0017] S5. Based on the insulation aging status assessment results, perform system integration and deployment.

[0018] This invention is based on a fusion model of DF-BiLSTM (Deep Feature Extraction-Bidirectional Long Short-Term Memory Network). Through its unique structure, it achieves end-to-end deep feature learning and bidirectional temporal context modeling of multi-source heterogeneous data, thereby significantly improving the accuracy and reliability of insulation aging status assessment.

[0019] In a more specific technical solution, in S1, the multiple physical stresses include: temperature stress, electric field stress, mechanical stress, and humidity stress.

[0020] This invention constructs an experimental platform capable of simultaneously applying stresses from multiple physical fields, including electrical, thermal, mechanical, and moisture fields. This realistically replicates the actual operating conditions of the valve-side outgoing line device of a converter transformer, overcoming the limitations of traditional methods that only consider a single stress factor. This makes the aging condition assessment results more accurate and reliable, and significantly improves the consistency with the actual insulation degradation process.

[0021] In a more specific technical solution, S2 includes multi-source time-series data such as dielectric response data, temperature field data, electric field distribution data, and aging characteristic parameters.

[0022] In a more specific technical solution, S3 includes the Deep Feature Extraction-Bidirectional Long Memory Network Model DF-BiLSTM, which includes: a Deep Feature Extraction (DF) module: using a three-layer one-dimensional convolutional neural network 1D-CNN to extract deep abstract features from the original high-dimensional data and output feature vectors of a preset dimension.

[0023] The Bi-LSTM timing modeling module includes two bidirectional LSTM layers to capture the bidirectional timing dependencies of the aging process.

[0024] This invention employs a deep feature extraction (DF) module (based on 1D-CNN), which can automatically learn and extract deep abstract features related to aging state directly from raw, high-dimensional, multi-source monitoring data (such as frequency domain dielectric spectrum, spatial temperature field, and electric field distribution). This avoids reliance on complex professional knowledge and manual feature engineering, and improves the efficiency and effectiveness of feature expression.

[0025] In a more specific technical solution, in the Bi-LSTM timing modeling module, for each time step t, the following logic is used to perform a forward LSTM unit:

[0026] Input gate: i t =σ(W xi x t +W hi h t-1 +b i )

[0027] Forgotten Gate: f t =σ(W xf xt +W hf h t-1 +b f )

[0028] Output gate: o t =σ(W xo x t +W ho h t-1 +b o )

[0029] Candidate cell status:

[0030] Cell renewal status:

[0031] Hidden state output: h t =o t ⊙tanh(C t ).

[0032] In a more specific technical solution, during the backward LSTM unit computation process of the Bi-LSTM time series modeling module, the backward LSTM unit executes the computation process in the forward LSTM unit in the reverse time direction to obtain the backward hidden state sequence.

[0033] In a more specific technical solution, bidirectional output fusion is performed in the Bi-LSTM time series modeling module.

[0034] This invention utilizes a bidirectional long short-term memory (BiLSTM) network module to capture the bidirectional long-term dependencies in the insulation aging process, enabling a more comprehensive understanding and modeling of the dynamic evolution of aging states. This achieves a leap from static assessment to dynamic prediction, significantly improving the accuracy of aging trend and remaining lifetime prediction.

[0035] In a more specific technical solution, during the bidirectional output fusion process, the forward and backward hidden states are evaluated to obtain the output at each time step t:

[0036]

[0037] In the formula, For forward LSTM output, For backward LSTM output, H t This is the final output.

[0038] In a more specific technical solution, S3 trains and configures the deep feature extraction-bidirectional long-term memory network model DF-BiLSTM, using the Adam optimizer with mean squared error (MSE) as the loss function.

[0039] In a more specific technical solution, a converter transformer insulation aging condition assessment system includes:

[0040] The stress loading module is used to load multiple physical stresses and construct an experimental platform;

[0041] The monitoring and acquisition module is used to collect and preprocess multi-source time-series data using the experimental platform to obtain real-time monitoring data. The monitoring and acquisition module is connected to the stress loading module.

[0042] The model building and training module is used to build and train the deep feature extraction-bidirectional long-term memory network model DF-BiLSTM to model the mapping relationship between the aging state of the converter transformer and the temperature and electric field distribution.

[0043] The state assessment module is used to input real-time monitoring data into the trained deep feature extraction-bidirectional long-term memory network model DF-BiLSTM to perform aging state assessment and verification, and output insulation aging state assessment results. The state assessment module is connected to the model building and training module and the monitoring and acquisition module.

[0044] The system deployment module is used to integrate and deploy the system based on the insulation aging status assessment results. The system deployment module is connected to the status assessment module.

[0045] The present invention has the following advantages over the prior art:

[0046] This invention is based on a fusion model of DF-BiLSTM (Deep Feature Extraction-Bidirectional Long Short-Term Memory Network). Through its unique structure, it achieves end-to-end deep feature learning and bidirectional temporal context modeling of multi-source heterogeneous data, thereby significantly improving the accuracy and reliability of insulation aging status assessment.

[0047] This invention constructs an experimental platform capable of simultaneously applying stresses from multiple physical fields, including electrical, thermal, mechanical, and moisture fields. This realistically replicates the actual operating conditions of the valve-side outgoing line device of a converter transformer, overcoming the limitations of traditional methods that only consider a single stress factor. This makes the aging condition assessment results more accurate and reliable, and significantly improves the consistency with the actual insulation degradation process.

[0048] This invention employs a deep feature extraction (DF) module (based on 1D-CNN), which can automatically learn and extract deep abstract features related to aging state directly from raw, high-dimensional, multi-source monitoring data (such as frequency domain dielectric spectrum, spatial temperature field, and electric field distribution). This avoids reliance on complex professional knowledge and manual feature engineering, and improves the efficiency and effectiveness of feature expression.

[0049] This invention utilizes a bidirectional long short-term memory (BiLSTM) network module to capture the bidirectional long-term dependencies in the insulation aging process, enabling a more comprehensive understanding and modeling of the dynamic evolution of aging states. This achieves a leap from static assessment to dynamic prediction, significantly improving the accuracy of aging trend and remaining lifetime prediction.

[0050] This invention addresses the technical problems existing in the prior art, such as insufficient model generalization ability; bottleneck in prediction accuracy; insensitivity to subtle changes in aging state, making early warning difficult; low reliability; lack of robustness in model judgment, making it susceptible to noise interference in the data; limited feature representation ability; and inability to effectively capture the temporal evolution law of insulation aging, thus affecting the accuracy of state prediction. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the basic steps of a converter transformer insulation aging condition assessment method according to Embodiment 1 of the present invention;

[0052] Figure 2 This is a schematic diagram of data stream processing for a converter transformer insulation aging status assessment method according to Embodiment 1 of the present invention;

[0053] Figure 3 This is a network unit structure diagram of the DF-BiLSTM model in Embodiment 1 of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0055] Example 1

[0056] like Figure 1 As shown, the present invention provides a method for assessing the insulation aging condition of a converter transformer, comprising the following basic steps:

[0057] S1. Multi-physics stress loading and experimental platform construction;

[0058] In this embodiment, a multi-stress accelerated aging experiment is conducted; specifically, by coupling multiple physical field stresses such as electrical, thermal, mechanical and humidity, the oil-paper insulation sample is synchronously and synergistically loaded to accurately simulate the material degradation process under actual operating conditions.

[0059] In this embodiment, a multi-stress accelerated aging test platform capable of accurately simulating the actual operating conditions of the valve side of a converter transformer is constructed. This platform must be able to achieve precise loading and coordinated control of multi-physical field stresses on the oil-paper insulation sample.

[0060] In this embodiment, multiphysics stress includes, but is not limited to:

[0061] Temperature stress: A high-precision temperature control system, specifically a PID-controlled circulating oil bath, is used to subject the sample to a constant temperature aging environment of 130±2℃, in accordance with the IEC 60076-14 thermal aging test standard, to simulate the thermal aging process of transformers during long-term operation.

[0062] Electric field stress: A 20kV / mm power frequency AC electric field is applied through a high-voltage generator, specifically a Haefely Hipotronics high-voltage generator, and a 15% content of high-order harmonics (mainly 3rd and 5th harmonics) is superimposed on it to simulate the actual electric field environment of the converter valve.

[0063] Mechanical stress: An electromagnetic vibration table, specifically the LDS V900 series electromagnetic vibration table, was used to apply continuous mechanical vibration at a frequency of 50 Hz and an acceleration of 0.5g to simulate the electromagnetic force vibration effect during transformer operation.

[0064] Humidity stress: By using a humidity control system, specifically a saturated salt solution method or a steam generator, the moisture content of the environment around the sample is maintained at 3.0 ± 0.5% to simulate the moisture absorption and aging process of insulating paper.

[0065] Each stress parameter needs to be monitored and recorded in real time using high-precision sensors. For example, temperature is measured using a type K thermocouple (accuracy ±0.5℃), electric field using a field strength meter (accuracy ±1%), and humidity using a dew point sensor (accuracy ±0.1%). The core of the platform lies in achieving precise coupling and independent control of multiple stresses to reproduce the most severe operating conditions in actual operation.

[0066] S2. Multi-source time-series data acquisition and preprocessing;

[0067] In this embodiment, the sampling frequency f is set. s Specifically, the frequency can be set to 1Hz to ensure complete capture of transient processes and steady-state responses, satisfying the requirements of Shannon's sampling theorem.

[0068] In this embodiment, multi-parameter synchronous acquisition is performed: a multi-parameter synchronous acquisition system is used to continuously acquire the following data at a sampling frequency of 1Hz:

[0069] Dielectric response data acquisition: A frequency domain dielectric spectrometer, specifically the IDAX-300, was used. -3 ~10 3The complex capacitance and loss factor are measured within the Hz frequency range to obtain dielectric spectrum data;

[0070] Temperature field data acquisition: The temperature distribution of 32 measuring points on the sample surface was acquired using a distributed fiber optic temperature measurement system, specifically the OZ Optics distributed fiber optic temperature measurement system, with a spatial resolution of 0.1m.

[0071] Electric field distribution data: A field strength sensor array, specifically the Probes EP100 field strength sensor array, was used to measure the electric field distribution in the space surrounding the sample, including 16 measurement points.

[0072] Aging characteristic parameters: Periodically sample and measure the degree of polymerization, i.e., DP value, using the viscosity method according to IEC 60450 standard; simultaneously measure the breakdown voltage according to IEC 60243 standard.

[0073] In this embodiment, data preprocessing is performed; specifically, the collected multi-source heterogeneous data is preprocessed as follows:

[0074] Specifically, data alignment is performed; among which, GPS timestamp-based synchronization technology is used to ensure the consistency of the time series of each parameter.

[0075] Normalization is performed; specifically, each parameter is normalized using min-max normalization to eliminate the influence of dimensions.

[0076]

[0077] Outlier handling is performed; specifically, for various types of time-series data collected, such as dielectric response, temperature field, and electric field distribution, outlier data points are removed using the 3σ criterion.

[0078] Complete missing data; specifically, estimate and fill in missing values ​​reasonably based on the continuity assumption of the time series using linear interpolation.

[0079] S3 and DF-BiLSTM model construction and training;

[0080] In this embodiment, time series data X is used. t , representing the multi-source heterogeneous data vector collected at time point t. Where n is the dimension of the feature space.

[0081] Construct a deep feature extraction-bidirectional long-term memory network model, the structure of which is as follows: Figure 2 As shown:

[0082] In this embodiment, the deep feature extraction (DF) module employs a three-layer one-dimensional convolutional neural network (1D-CNN), with specific parameters including:

[0083] First layer: 64 convolutional kernels, kernel size 5, ReLU activation function.

[0084] Second layer: 128 convolutional kernels, kernel size 3, ReLU activation function.

[0085] The third layer: 256 convolutional kernels, kernel size 3, ReLU activation function.

[0086] This module is used to automatically extract deep abstract features from raw high-dimensional data such as dielectric spectrum, temperature field, and electric field distribution, and output a 256-dimensional feature vector.

[0087] In this embodiment, the Bi-LSTM timing modeling module includes two bidirectional LSTM layers with the following structural parameters:

[0088] First layer: 128 hidden units, return_sequences = True

[0089] Second layer: 64 hidden units

[0090] This module is used to capture the bidirectional temporal dependencies of the aging process, and its forward propagation process can be described in detail as follows:

[0091] Forward LSTM cell computation process:

[0092] For each time step t, the forward LSTM unit calculates as follows:

[0093] Input gate: i t =σ(W xi x t +W hi h t-1 +b i )

[0094] Forgotten Gate: f t =σ(W xf x t +W hf h t-1 +b f )

[0095] Output gate: o t =σ(W xo x t +W ho h t-1 +b o )

[0096] Candidate cell status:

[0097] Cell renewal status:

[0098] Hidden state output: ht =o t ⊙tanh(C t )

[0099] Backward LSTM unit computation process: At the same time, the backward LSTM unit performs the same computation in the reverse time direction to obtain the backward hidden state sequence.

[0100] Bidirectional output fusion: Ultimately, the output at each time step t is composed of the forward and backward hidden states.

[0101]

[0102] in, For forward LSTM output, For backward LSTM output, H t This is the final output.

[0103] In this embodiment, the training configuration is performed; specifically, the Adam optimizer is used (learning rate 0.001, β1 = 0.9, β2 = 0.999), with mean squared error (MSE) as the loss function:

[0104]

[0105] The training cycle is 500 rounds, the batch size is 32, an early stopping strategy (patience=20) is used to prevent overfitting, and Dropout (rate=0.2) is used to increase generalization ability.

[0106] S4. Aging status assessment and verification;

[0107] In this embodiment, real-time condition assessment is performed; specifically, real-time monitoring data is input into the trained DF-BiLSTM model, and the insulation aging condition assessment result is output:

[0108] Cohesion degree prediction: The output DP value prediction result has a mean absolute error (MAE) of less than 5% compared with the actual measured value.

[0109] Aging level classification: The insulation condition is divided into 4 levels (Good: DP>800; Caution: 600<DP≤800; Warning: 400<DP≤600; Danger: DP≤400), with a classification accuracy rate of over 95%.

[0110] Remaining lifetime prediction: Outputs the probability distribution of remaining lifetime, providing lifetime predictions at a 90% confidence interval;

[0111] In this embodiment, model validation is performed; specifically, k-fold cross-validation (k=5) is used, and the dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Compared with traditional detection methods (such as high-performance liquid chromatography for furfural content determination and viscometry for DP value determination), the model evaluation accuracy should reach above 96.2%. Simultaneously, uncertainty is quantified through Monte Carlo simulation to assess the reliability of the prediction results.

[0112] Degree of polymerization (DP value): A key indicator characterizing the average degree of polymerization of cellulose polymer chains in insulating paper. It is accurately measured using the capillary viscosity method according to IEC 60450 standard. The initial value DP0 ≥ 1000, and the failure threshold is DP. crit =250.

[0113] Frequency domain dielectric spectrum (FDS): in a wide bandwidth f∈

[10] -3 10 3 Within 1 Hz, the complex capacitance C is measured by frequency sweeping. * Spectral characteristics of tanδ(ω) and dielectric loss factor tanδ(ω).

[0114] 3σ outlier removal criterion: Based on the normal distribution assumption, for data point x... i If |x i If -μ|>3σ (where μ is the sample mean and σ is the sample standard deviation), it is considered a statistical outlier and should be removed.

[0115] S5, System Integration and Deployment.

[0116] In this embodiment, hardware system integration is performed, and the above method is integrated into an embedded online monitoring system, which includes:

[0117] The data acquisition unit includes a multi-parameter synchronous acquisition card (such as NI PXle-51771R), with a sampling rate of 100kS / s and 16-bit resolution.

[0118] The preprocessing module includes: data preprocessing and feature extraction based on FPGA (such as Xilinx Kintex-7).

[0119] The real-time evaluation unit includes: deploying a trained DF-BiLSTM model using an embedded GPU (such as NVIDIA Jetson AGX Orin) to perform inference evaluation with a cycle of less than 1 second.

[0120] In this embodiment, software system development is carried out; specifically, the human-machine interface displays insulation status assessment results, aging trend curves, and early warning information (such as audible and visual alarms). The system integrates with existing monitoring systems (such as SCADA) via the OPC UA protocol, supporting real-time data upload and remote monitoring.

[0121] In this embodiment, a standardized system deployment is implemented; specifically, all hardware units must meet the IEC 61850 standard to ensure reliable operation in the harsh electromagnetic environment of the substation.

[0122] In summary, this invention is based on a fusion model of DF-BiLSTM (Deep Feature Extraction-Bidirectional Long Short-Term Memory Network). Through its unique structure, it achieves end-to-end deep feature learning and bidirectional temporal context modeling of multi-source heterogeneous data, thereby significantly improving the accuracy and reliability of insulation aging status assessment.

[0123] This invention constructs an experimental platform capable of simultaneously applying stresses from multiple physical fields, including electrical, thermal, mechanical, and moisture fields. This realistically replicates the actual operating conditions of the valve-side outgoing line device of a converter transformer, overcoming the limitations of traditional methods that only consider a single stress factor. This makes the aging condition assessment results more accurate and reliable, and significantly improves the consistency with the actual insulation degradation process.

[0124] This invention employs a deep feature extraction (DF) module (based on 1D-CNN), which can automatically learn and extract deep abstract features related to aging state directly from raw, high-dimensional, multi-source monitoring data (such as frequency domain dielectric spectrum, spatial temperature field, and electric field distribution). This avoids reliance on complex professional knowledge and manual feature engineering, and improves the efficiency and effectiveness of feature expression.

[0125] This invention utilizes a bidirectional long short-term memory (BiLSTM) network module to capture the bidirectional long-term dependencies in the insulation aging process, enabling a more comprehensive understanding and modeling of the dynamic evolution of aging states. This achieves a leap from static assessment to dynamic prediction, significantly improving the accuracy of aging trend and remaining lifetime prediction.

[0126] This invention addresses the technical problems existing in the prior art, such as insufficient model generalization ability; bottleneck in prediction accuracy; insensitivity to subtle changes in aging state, making early warning difficult; low reliability; lack of robustness in model judgment, making it susceptible to noise interference in the data; limited feature representation ability; and inability to effectively capture the temporal evolution law of insulation aging, thus affecting the accuracy of state prediction.

[0127] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing the aging condition of converter transformer insulation, characterized in that, The method includes: S1. Applying multiple physical stresses to construct an experimental platform; S2. Using the experimental platform, collect and preprocess multi-source time-series data to obtain real-time monitoring data; S3. Construct and train the deep feature extraction-bidirectional long-term memory network model DF-BiLSTM to model the mapping relationship between the aging state of the converter transformer and the temperature and electric field distribution. S4. Input the real-time monitoring data into the trained deep feature extraction-bidirectional long-term memory network model DF-BiLSTM to perform aging state assessment and verification, and output the insulation aging state assessment result. S5. Based on the insulation aging status assessment results, perform system integration and deployment.

2. The method for assessing the insulation aging condition of a converter transformer according to claim 1, characterized in that, In S1, the multiple physical stresses include: temperature stress, electric field stress, mechanical stress, and humidity stress.

3. The method for assessing the insulation aging condition of a converter transformer according to claim 1, characterized in that, In S2, the multi-source time-series data includes: dielectric response data, temperature field data, electric field distribution data, and aging characteristic parameters.

4. The method for assessing the insulation aging condition of a converter transformer according to claim 1, characterized in that, In S3, the deep feature extraction-bidirectional long-term memory network model DF-BiLSTM includes: a deep feature extraction (DF) module: using a three-layer one-dimensional convolutional neural network 1D-CNN to extract deep abstract features from the original high-dimensional data and output a feature vector of a preset dimension; The Bi-LSTM timing modeling module includes two bidirectional LSTM layers to capture the bidirectional timing dependencies of the aging process.

5. The method for assessing the insulation aging condition of a converter transformer according to claim 4, characterized in that, In the Bi-LSTM timing modeling module, for each time step t, the forward LSTM unit is performed using the following logic: Input gate: i t =σ(W xi x t +W hi h t-1 +b i ) Forgotten Gate: f t =σ(W xf x t +W hf h t-1 +b f ) Output gate: o t =σ(W xo x t +W ho h t-1 +b o ) Candidate cell status: Cell renewal status: Hidden state output: h t =o t ☉tanh(C t ).

6. The method for assessing the insulation aging condition of a converter transformer according to claim 4, characterized in that, During the backward LSTM unit computation process of the Bi-LSTM time series modeling module, the backward LSTM unit executes the computation process in the forward LSTM unit in the reverse time direction to obtain the backward hidden state sequence.

7. The method for assessing the insulation aging condition of a converter transformer according to claim 4, characterized in that, In the Bi-LSTM time series modeling module, bidirectional output fusion is performed.

8. The method for assessing the insulation aging condition of a converter transformer according to claim 7, characterized in that, During the bidirectional output fusion process, the forward and backward hidden states are evaluated to obtain the output at each time step t: In the formula, For forward LSTM output, For backward LSTM output, H t This is the final output.

9. The method for assessing the insulation aging condition of a converter transformer according to claim 1, characterized in that, In step S3, the deep feature extraction-bidirectional long-term memory network model DF-BiLSTM is trained and configured, wherein the Adam optimizer is used and the mean squared error (MSE) is used as the loss function.

10. A converter transformer insulation aging condition assessment system, characterized in that, The system includes: The stress loading module is used to load multiple physical stresses and construct an experimental platform; The monitoring and acquisition module is used to collect and preprocess multi-source time-series data using the experimental platform to obtain real-time monitoring data. The monitoring and acquisition module is connected to the stress loading module. The model building and training module is used to build and train the deep feature extraction-bidirectional long-term memory network model DF-BiLSTM to model the mapping relationship between the aging state of the converter transformer and the temperature and electric field distribution. The state assessment module is used to input the real-time monitoring data into the trained deep feature extraction-bidirectional long-term memory network model DF-BiLSTM to perform aging state assessment and verification, and output insulation aging state assessment results. The state assessment module is connected to the model construction and training module and the monitoring and acquisition module. The system deployment module is used to perform system integration and deployment based on the insulation aging status assessment results. The system deployment module is connected to the status assessment module.

Images