An edge-intelligent data acquisition method, device and medium for transformer windings

Abstract

The application discloses a kind of edge intelligent data acquisition methods, equipment and medium for transformer winding, belong to transformer winding state monitoring technical field, including constructing multi-channel high-precision synchronous sampling unit, the real-time acquisition of the multiple physical quantities of transformer winding is carried out;Based on the electromagnetic thermal coupling characteristics of transformer winding, construct digital twin model to assist state evaluation;For the multi-source heterogeneous data of transformer winding, singular value decomposition and attention mechanism are combined to extract features;Introduce reinforcement learning algorithm, dynamically optimize the extracted features and dynamically adjust the sampling strategy according to the winding state, and detect the abnormality of key state indicators, and evaluate the performance through comprehensive performance indicators.The application realizes the fundamental improvement of transformer winding state monitoring performance through the systematic innovation of data acquisition and intelligent analysis integration.

Description

Technical Field

[0001] This invention relates to the field of transformer winding condition monitoring technology, specifically to an edge intelligent data acquisition method, device, and medium for transformer windings. Background Technology

[0002] As a key piece of equipment in the power system, the safe and stable operation of transformers is crucial to the reliability of the power grid. In recent years, with the advancement of smart grid construction, higher requirements have been placed on the real-time performance, accuracy, and intelligence level of transformer winding condition monitoring. Currently, transformer winding condition monitoring technology has made some progress. Existing technologies disclose a big data-based intelligent operation and maintenance method and system for transformers, realizing the collection and analysis of data throughout the transformer's entire life cycle. Existing technologies also propose an online monitoring system for transformer windings, which uses a multi-channel synchronous sampling unit to collect winding voltage and current data in real time and calculates short-circuit reactance to monitor deformation.

[0003] However, existing technologies still have some limitations. First, traditional centralized data processing architectures result in high data transmission latency and insufficient real-time performance, making it difficult to meet the millisecond-level response requirements of smart grids. Second, existing monitoring systems lack the ability to dynamically model the multi-physics coupling characteristics of transformer windings, leading to a high error rate in state assessment and affecting the accuracy of fault early warning. Furthermore, the design pattern of separating data acquisition and analysis modules results in overall system inefficiency, insufficient resource utilization, and difficulty in adapting to the resource constraints of edge computing environments.

[0004] While some recent technologies and existing methods and systems for assessing the condition of rural power grid transformers have begun to utilize multi-source data, they have not yet fully leveraged the advantages of edge computing and artificial intelligence. Currently, there is no transformer winding condition monitoring system that can integrate data acquisition and intelligent analysis on edge devices, accurately model the multi-physics coupling characteristics of windings, and possess adaptive optimization capabilities. Therefore, developing an edge intelligent data acquisition module and system for transformer windings to improve the real-time performance, accuracy, and resource utilization efficiency of condition monitoring has significant theoretical and practical value.

[0005] Current transformer winding condition monitoring technology in power systems faces three main problems. First, traditional data acquisition systems employ a centralized processing architecture, resulting in high data transmission latency and insufficient real-time performance, making it difficult to meet the millisecond-level response requirements of smart grids. Second, existing monitoring systems lack the ability to dynamically model the multi-physics coupling characteristics of transformer windings, leading to a condition assessment error rate as high as 15%-20%, affecting the accuracy of fault early warning. Third, the separation of data acquisition and analysis modules results in low overall system efficiency, with resource utilization below 60%, making it difficult to adapt to the resource constraints of edge computing environments.

[0006] These issues severely restrict the performance improvement of transformer winding condition monitoring systems. Insufficient real-time performance prevents the system from timely detecting and handling potential faults, increasing the risk of unplanned equipment outages. Inadequate dynamic modeling capabilities make it difficult for the system to accurately capture minute changes in winding conditions, reducing the effectiveness of predictive maintenance. Furthermore, the separation of acquisition and analysis modules leads to a waste of computing resources and limits the system's deployment capabilities on edge devices. Summary of the Invention

[0007] In view of the above-mentioned problems, the present invention is proposed.

[0008] Therefore, the technical problem solved by this invention is: how to design an edge intelligent data acquisition method, device, and medium for transformer windings to achieve the integration of data acquisition and intelligent analysis, thereby improving the real-time performance, accuracy, and resource utilization efficiency of condition monitoring. This system should be able to accurately model the multi-physics coupling characteristics of the windings under limited edge computing resources, and continuously optimize the condition assessment strategy through adaptive algorithms, thereby significantly improving the overall performance of transformer winding condition monitoring.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an edge intelligent data acquisition method for transformer windings, comprising,

[0010] A multi-channel high-precision synchronous sampling unit is constructed to collect multiple physical quantities of the transformer winding in real time. Based on the electromagnetic and thermodynamic coupling characteristics of the transformer winding, a digital twin model is constructed to assist in state assessment. For multi-source heterogeneous data of the transformer winding, a combination of singular value decomposition and attention mechanism is used for feature extraction. A reinforcement learning algorithm is introduced to dynamically optimize the extracted features and dynamically adjust the sampling strategy according to the winding state. Anomaly detection is performed on key state indicators, and performance is evaluated through comprehensive performance indicators.

[0011] As a preferred embodiment of the edge intelligent data acquisition method for transformer windings described in this invention, the construction of a multi-channel high-precision synchronous sampling unit includes using a high-resolution ADC with a resolution of not less than 24 bits and an adjustable sampling rate range of 10-100kHz to construct a multi-channel high-precision synchronous sampling unit for real-time acquisition of multiple physical quantities of the transformer windings. The sampling process is represented as follows:

[0012] ,

[0013] in, for Channel sampled data vector, including voltage Current ,temperature ,magnetic field physical quantity, Let R be the sampling noise vector, R be the set of real numbers, and T be the transpose matrix;

[0014] Sampling time synchronization error is controlled within the microsecond level:

[0015] ,

[0016] in, Indicates the first Sampling time of the channel Let M be the reference time, M be the number of sampling channels, and i be the variable index.

[0017] As a preferred embodiment of the edge intelligent data acquisition method for transformer windings described in this invention, the digital twin model includes: constructing a digital twin model based on the electromagnetic-thermal coupling characteristics of the transformer windings; accurately reproducing the dynamic behavior of the windings through the virtual model; and assisting in state assessment and predictive maintenance.

[0018] ,

[0019] in, It is a nonlinear state transition matrix. It is a nonlinear control matrix. To control the input vector, Let n be the process noise vector, and n represent the dimension of the system state variables. For at t+1 The channel sampled data vector, where m represents the dimension of the control input vector.

[0020] As a preferred embodiment of the edge intelligent data acquisition method for transformer windings described in this invention, the feature extraction using a combination of singular value decomposition and attention mechanism includes, considering the characteristics of multi-source heterogeneous data from transformer windings, using a combination of singular value decomposition and attention mechanism for feature extraction.

[0021] ,

[0022] in, For parameterized encoder functions, For encoder parameters, The result of SVD decomposition. For attention mechanism functions, Represents the Hadamard product. This represents the result of feature extraction;

[0023] Feature extraction results The state estimate is input as the system state encoding vector and compared with the prediction results of the digital twin model to update the state estimate.

[0024] As a preferred embodiment of the edge intelligent data acquisition method for transformer windings described in this invention, the dynamic optimization of the extracted features includes introducing a reinforcement learning algorithm to define a comprehensive value function by dynamically optimizing the data acquisition and processing strategy. for:

[0025] ,

[0026] in, and Representing the virtual model and the physical system state respectively. The reward function for winding state evaluation. For the discount factor of the future τ steps, , , These are the weighting coefficients. This is a penalty term for model uncertainty or environmental fluctuations. This is the expected combined value function under the current state and action. This represents the expected value of the sampling trajectory in the next τ steps under the empirical data distribution D, where t is the time step. The total time step, This indicates that at time step t... Actions performed during feature extraction;

[0027] Based on the current comprehensive value function, update the Q function:

[0028] ,

[0029] in, and These are the Q-value functions based on predictions from digital twin models and actual data, respectively. As a balance factor, This is the final integrated policy function. The optimal action for the next time step. express The result of feature extraction at that time, For the future Discount factor for each step;

[0030] Simultaneously, an edge computing efficiency factor is introduced. Correct Q function update:

[0031] ,

[0032] in, This is a Q-value estimation function based solely on local data.

[0033] As a preferred embodiment of the edge intelligent data acquisition method for transformer windings described in this invention, the step of dynamically adjusting the sampling strategy according to the winding state includes setting an adaptive sampling frequency adjustment function and dynamically adjusting the sampling strategy according to the winding state.

[0034] ,

[0035] in, for sampling frequency at any time For learning rate, This represents the gradient operator with respect to frequency.

[0036] As a preferred embodiment of the edge intelligent data acquisition method for transformer windings described in this invention, the anomaly detection includes defining an anomaly evaluation function for highly sensitive anomaly detection of key status indicators on the edge device side. The anomaly evaluation function is as follows:

[0037] ,

[0038] in, The outlier score is the score obtained from the outlier evaluation function. To represent the actual sampled state value at time t, This represents the predicted value of the system state based on the constructed digital twin model. Represents the L2 norm. Let KL divergence be the KL divergence. These are the weighting coefficients. and These are the probability distributions of the actual observed values ​​and the predicted values, respectively.

[0039] The outlier score obtained from the outlier evaluation function serves as the input variable for performance evaluation and sampling strategy adjustment.

[0040] As a preferred embodiment of the edge intelligent data acquisition method for transformer windings described in this invention, the performance evaluation through comprehensive performance indicators includes:

[0041] ,

[0042] in, To evaluate the performance results through comprehensive performance indicators, This is the accuracy function for winding condition diagnosis. For the system delay function, For energy efficiency function, , , These are the weighting coefficients for the three types of comprehensive performance indicators;

[0043] Introducing meta-learning update rules to improve adaptive capabilities:

[0044] ,

[0045] in, To update the model parameters after introducing meta-learning rules, These are the parameters of the original model. The meta-learning rate, The meta-objective function is... and These are the training set and the validation set, respectively. Indicates the parameter The gradient operator.

[0046] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the edge intelligent data acquisition method for transformer windings.

[0047] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the edge intelligent data acquisition method for transformer windings.

[0048] The beneficial effects of this invention: The edge intelligent data acquisition module and system for transformer windings proposed in this invention have significant advantages over existing technologies. Firstly, the embodiments of this invention integrate data acquisition and intelligent analysis by combining digital twin and reinforcement learning technologies. This innovative architecture breaks through the limitations of traditional centralized processing models, offloading complex data processing and analysis tasks to edge devices, significantly reducing data transmission latency and improving system response speed. Specifically, the embodiments of this invention improve the real-time performance of winding state estimation to within 10ms, nearly 70% faster than existing technologies, effectively meeting the millisecond-level response requirements of smart grids.

[0049] Secondly, the multi-channel high-precision synchronous sampling unit designed in this embodiment of the invention, combined with a digital twin system based on a nonlinear state-space model, significantly enhances the dynamic modeling capability of the multi-physics coupling characteristics of transformer windings. By employing a 24-bit high-resolution ADC and microsecond-level synchronous sampling technology, along with a feature extraction method coupled with SVD and attention mechanisms, this embodiment of the invention can accurately capture minute changes in the winding state. This high-precision, multi-dimensional state monitoring method enables the system's state assessment accuracy to exceed 95%, an improvement of 15%-20% compared to existing technologies, significantly enhancing the reliability of fault early warning.

[0050] Finally, by introducing an edge computing efficiency factor and an adaptive sampling strategy, this embodiment of the invention effectively solves the problem of insufficient resource utilization. Compared with existing technologies, the edge computing efficiency of this embodiment is improved by 200%, and the overall system energy consumption is reduced by 40%. Simultaneously, the meta-learning-based model optimization method endows the system with strong adaptive capabilities, enabling it to quickly adapt to changing operating environments. These innovations not only significantly improve the system's resource utilization efficiency but also enable this embodiment to achieve high-performance transformer winding status monitoring with limited edge computing resources, providing reliable technical support for the intelligent upgrading of power systems. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a general flowchart of an edge intelligent data acquisition method for transformer windings provided in one embodiment of the present invention.

[0053] Figure 2 This paper compares the data processing latency of different edge intelligent data acquisition methods for transformer windings, as provided in one embodiment of the present invention.

[0054] Figure 3 This paper presents a comparison of the accuracy of equipment under typical operating conditions for an edge intelligent data acquisition method for transformer windings, as provided in an embodiment of the present invention.

[0055] Figure 4 This invention provides a comparison of equipment resource utilization under typical operating conditions for an edge intelligent data acquisition method for transformer windings, as provided in one embodiment of the present invention.

[0056] Figure 5A comparison of device response times under typical operating conditions for an edge intelligent data acquisition method for transformer windings provided in an embodiment of the present invention.

[0057] Figure 6 A comparison of the robustness of an edge-intelligent data acquisition method for transformer windings under typical operating conditions, as provided in an embodiment of the present invention.

[0058] Figure 7 Different algorithms for edge intelligent data acquisition methods for transformer windings, as provided in one embodiment of the present invention, are discussed in terms of CPU utilization on edge devices.

[0059] Figure 8 Different algorithms for an edge intelligent data acquisition method for transformer windings, as provided in one embodiment of the present invention, are used in the memory of the edge device.

[0060] Figure 9 This invention provides an embodiment of an edge intelligent data acquisition method for transformer windings, which includes different algorithms for edge device energy consumption. Detailed Implementation

[0061] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0062] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides an edge intelligent data acquisition method for transformer windings, including:

[0063] To address the problems of insufficient real-time performance, inadequate dynamic modeling capabilities, and low overall system efficiency in current transformer winding condition monitoring technologies in power systems, this invention aims to provide an edge intelligent data acquisition module and system for transformer windings, integrating data acquisition and intelligent analysis to improve the real-time performance, accuracy, and resource utilization efficiency of condition monitoring.

[0064] Existing technologies, when addressing the aforementioned pain points, often suffer from drawbacks such as high data transmission latency, high condition assessment error rates (15%-20%), and insufficient resource utilization (below 60%). These problems severely restrict the performance improvement of transformer winding condition monitoring systems and increase the risk of unplanned equipment downtime.

[0065] This invention can be widely applied to scenarios such as transformer condition monitoring and predictive maintenance in smart grids.

[0066] Its technical value lies in its ability to significantly improve the real-time performance (less than 10ms), accuracy (over 95%), and resource utilization efficiency (200% improvement) of status monitoring. Given its outstanding advantages in improving power grid reliability and reducing maintenance costs, this invention has good application prospects in the field of smart grid upgrading and transformation.

[0067] S1. Construct a multi-channel high-precision synchronous sampling unit to collect multiple physical quantities of the transformer winding in real time.

[0068] It should be noted that, in the embodiments of the present invention, a high-resolution ADC with a resolution of not less than 24 bits and an adjustable sampling rate range of 10-100kHz is used to construct a multi-channel high-precision synchronous sampling unit for real-time acquisition of multiple physical quantities of the transformer winding. The sampling process is represented as follows:

[0069] ,

[0070] in, for Channel sampled data vector, including voltage Current ,temperature ,magnetic field physical quantity, Let R be the sampling noise vector, R be the set of real numbers, and T be the transpose matrix;

[0071] In an optional embodiment, a single 24-bit high-resolution ADC can be replaced by multiple 16-bit ADC chips forming an interleaved sampling array, extending the effective resolution to an equivalent 24 bits through time interleaving technology. This alternative meets the requirements for dynamic sampling rate adjustment and can compensate for inter-channel mismatch errors through calibration algorithms, maintaining the technical specifications of time synchronization accuracy. Furthermore, the time synchronization mechanism can use the IEEE 1588 precise time protocol instead of hardware-triggered synchronization, achieving sub-microsecond-level synchronization error compensation through a clock servo control system.

[0072] Sampling time synchronization error is controlled within the microsecond level:

[0073] ,

[0074] in, Indicates the first Sampling time of the channel Let M be the reference time, M be the number of sampling channels, and i be the variable index.

[0075] S2. Based on the electromagnetic-thermal coupling characteristics of transformer windings, a digital twin model is constructed for auxiliary state assessment.

[0076] It should be noted that, based on the electromagnetic-thermal coupling characteristics of transformer windings, this embodiment of the invention constructs a digital twin model. Its purpose is to accurately reproduce the dynamic behavior of the windings through a virtual model without directly interfering with the physical system, thereby assisting in state assessment and predictive maintenance. The digital twin model provides a highly abstract state estimation mechanism, enabling the system to perform fast, low-latency state simulation and analysis on edge devices. The digital twin model is expressed as follows:

[0077] ,

[0078] in, This is a nonlinear state transition matrix that reflects the dynamic changes in the internal state of the winding. It is a nonlinear control matrix. To control the input vector (such as load regulation). Let n be the process noise vector, and n represent the dimension of the system state variables. for Channel sampling data vector, for The channel sampled data vector, where m represents the dimension of the control input vector.

[0079] For the construction of digital twin models, the nonlinear state equation of an optional embodiment of the present invention It can be transformed into a hybrid modeling architecture. Specifically, the nonlinear state transition matrix can be... Decomposed into linear principal terms and nonlinear correction terms Construct a piecewise linearized model Alternatively, Long Short-Term Memory (LSTM) networks can be used instead of traditional differential equation modeling, through... Establish hidden state evolution relationships to maintain the ability to characterize the dynamic characteristics of the winding, wherein, Let t be the hidden state vector of the LSTM at time t.

[0080] This invention proposes an edge-intelligent data acquisition module and system for transformer windings. By integrating digital twin and reinforcement learning technologies, it achieves the integration of data acquisition and intelligent analysis. Specifically, the embodiments of this invention design a multi-channel high-precision synchronous sampling unit, employing a high-resolution ADC to achieve real-time acquisition of multiple physical quantities of the transformer windings. The sampling accuracy reaches 24 bits, the sampling rate is adjustable from 10 to 100 kHz, and the sampling time synchronization error is controlled at the microsecond level. Simultaneously, based on the electromagnetic-thermal coupling characteristics of the transformer windings, a digital twin system of a nonlinear state-space model is constructed to accurately describe the dynamic changes in the internal state of the windings.

[0081] S3. For multi-source heterogeneous data of transformer windings, a combination of singular value decomposition and attention mechanism is used for feature extraction.

[0082] It should be noted that, considering the characteristics of multi-source heterogeneous data of transformer windings, this embodiment of the invention uses a method combining singular value decomposition (SVD) and attention mechanism for feature extraction. The purpose is to compress redundant information and highlight key feature dimensions in order to improve the accuracy of state coding and decision performance. This invention overcomes the dynamic modeling problem of multi-physics coupling characteristics of transformer windings. Specifically, by constructing a digital twin system of nonlinear state-space model and combining the feature extraction methods of SVD and attention mechanism, an accurate description of complex physical processes is achieved.

[0083] The results of feature extraction are as follows:

[0084] ,

[0085] in, For parameterized encoder functions, For encoder parameters, The result of SVD decomposition. For attention mechanism functions, Represents the Hadamard product. This represents the result of feature extraction;

[0086] Feature extraction results The state estimate is input as the system state encoding vector and compared with the prediction results of the digital twin model to update the state estimate.

[0087] The SVD-attention fusion architecture of this invention can be replaced with other matrix factorization and feature weighting mechanisms. Specifically, nonnegative matrix factorization (NMF) can be used instead of singular value factorization to construct... The low-rank approximation is combined with a gated attention mechanism to achieve feature selection: .in, θ is the encoder function with parameter θ, used to encode the fused features; W is the trainable weight matrix; For Hadamard product (i.e., element-wise multiplication); GateAttn(⋅) is the gated attention mechanism function used to extract important information from the input; This is the transpose of the transformation matrix; For the original input features at time step t, alternatively, wavelet packet transform can be used instead of SVD for time-frequency feature extraction.

[0088] ,

[0089] in, To represent the relationship between input x and basis functions in dimension k. The inner product or projection coefficient; Let k be the k-th frequency domain basis function; Summation is performed across all basis function dimensions; This represents the feature encoding result obtained by frequency domain projection over time t; Let be the basis functions of the j-th decomposition layer.

[0090] The multi-scale decomposition coefficients are obtained, and then the feature encoding is completed by combining the convolutional attention mechanism.

[0091] In terms of feature extraction, this embodiment of the invention employs a method combining SVD and an attention mechanism to effectively process multi-source heterogeneous data from transformer windings. By parameterizing the encoder function, high-dimensional sampled data is mapped to low-dimensional feature vectors, improving data processing efficiency. To optimize system performance, a dynamic optimization strategy based on reinforcement learning is introduced. Specifically, a comprehensive value function is defined, considering the differences between the virtual model and the physical system state, the winding state evaluation reward, and model uncertainty. Adaptive optimization of the system is achieved through iterative updates of the Q-function.

[0092] S4. Introduce reinforcement learning algorithms to dynamically optimize the extracted features, dynamically adjust the sampling strategy according to the winding state, and perform anomaly detection on key state indicators. Evaluate performance through comprehensive performance indicators.

[0093] It should be noted that by introducing reinforcement learning algorithms and dynamically optimizing data acquisition and processing strategies, the system's adaptive capabilities are improved; a comprehensive value function is defined. for:

[0094] ,

[0095] in, and Representing the virtual model and the physical system state respectively. The reward function for winding state evaluation. For the future Discount factor of step , , These are the weighting coefficients. This is a penalty term for model uncertainty or environmental fluctuations. This is the expected combined value function under the current state and action. This represents the prediction of the future under the empirical data distribution D. The expected value of the sampled trajectory, where t is the time step. The total time step, This indicates that at time step t... Actions performed during feature extraction;

[0096] Based on the current comprehensive value function, update the Q function:

[0097] ,

[0098] in, and These are the Q-value functions based on predictions from digital twin models and actual data, respectively. As a balance factor, This is the final integrated policy function. The optimal action for the next time step is used to drive sampling and control decisions; express The result of feature extraction at that time, For the future Discount factor for each step;

[0099] In one optional embodiment of the present invention, the discount factor can be... Extended to time-varying function It balances short-term and long-term rewards through an adaptive decay mechanism, in which... This is the discount factor at time step t, used to balance the importance of current rewards and future rewards; is the initial discount factor, typically ranging from (0,1); s is a coefficient controlling the rate at which the discount decays over time; the larger the value, the faster the discount factor decays. This is the decay rate coefficient.

[0100] To adapt to the resource constraints of the edge computing environment, an edge computing efficiency factor is introduced. Correct Q function update:

[0101] ,

[0102] in, This is a Q-value estimation function based solely on local data.

[0103] Regarding performance optimization under edge computing resource constraints, this invention adopts a dynamic optimization strategy based on reinforcement learning, and introduces an edge computing efficiency factor to correct the Q function update, thereby achieving the goal of effectively reducing resource consumption while ensuring performance.

[0104] It should also be noted that an adaptive sampling frequency adjustment function is set to dynamically adjust the sampling strategy based on the winding state:

[0105] ,

[0106] in, for sampling frequency at any time For learning rate, This represents the gradient operator with respect to frequency.

[0107] Furthermore, an outlier evaluation function is defined to perform highly sensitive anomaly detection on the edge device side for key status indicators (such as winding temperature, stress, and deformation); the outlier evaluation function is as follows:

[0108] ,

[0109] in, The outlier score is the score obtained from the outlier evaluation function. To represent the actual sampled state value at time t, This indicates that the predicted value of the system state based on the constructed digital twin model has been expressed through the state transition equation. It can be concluded that; Represents the L2 norm. Let KL divergence be the KL divergence. These are the weighting coefficients. and These are the probability distributions of the actual observed values ​​and the predicted values, respectively.

[0110] This detection function combines numerical error and distribution bias to improve its response to events such as non-Gaussian anomalies and sudden deformations. The resulting anomaly scores can be used as input variables for subsequent performance evaluation and sampling strategy adjustments.

[0111] To adapt to the resource constraints of edge computing environments, this invention introduces an edge computing efficiency factor to correct the Q-function update process and balance the contributions of the global model and local data. Furthermore, an adaptive sampling frequency adjustment function is designed to dynamically adjust the sampling strategy based on the winding state, improving system resource utilization efficiency. In terms of anomaly detection, an anomaly metric function based on the L2 norm and KL divergence is defined, paying particular attention to anomalies in key indicators such as winding temperature and deformation, thus improving the accuracy of fault warnings.

[0112] Furthermore, comprehensive performance metrics are defined to evaluate the overall performance of the system:

[0113] ,

[0114] in, To evaluate the performance results through comprehensive performance indicators, This is the accuracy function for winding condition diagnosis. For the system delay function, For energy efficiency function, , , These are the weighting coefficients for the three types of comprehensive performance indicators;

[0115] Regarding the improvement of system adaptability, this invention introduces meta-learning update rules to achieve dynamic optimization of model parameters, enabling the system to quickly adapt to changing operating environments.

[0116] Introducing meta-learning update rules to improve adaptive capabilities:

[0117] ,

[0118] in, To update the model parameters after introducing meta-learning rules, These are the parameters of the original model. The meta-learning rate, The meta-objective function is... and These are the training set and the validation set, respectively. Indicates the parameter The gradient operator.

[0119] This invention proposes a comprehensive performance evaluation index, combining factors such as winding condition diagnosis accuracy, system latency, and energy efficiency to comprehensively evaluate system performance. To further enhance the system's adaptive capability, a meta-learning-based model optimization method is introduced. By dynamically updating model parameters, the system can quickly adapt to changing operating environments. Through these technological innovations, this invention significantly improves the real-time performance, accuracy, and resource utilization efficiency of transformer winding condition monitoring, providing strong support for the safe and stable operation of power systems.

[0120] Through the aforementioned technologies, this invention solves the problems of high data transmission delay, insufficient real-time performance, high error rate in state assessment due to insufficient dynamic modeling capabilities, and low overall system efficiency caused by the separation of data acquisition and analysis modules in existing transformer winding condition monitoring systems. This invention integrates data acquisition and intelligent analysis, significantly improving the real-time performance, accuracy, and resource utilization efficiency of condition monitoring. Specifically, embodiments of this invention improve sampling accuracy to 24 bits, control time synchronization error to within 1 μs, achieve real-time winding condition estimation of less than 10 ms, achieve an accuracy exceeding 95%, improve anomaly detection sensitivity by 30%, reduce false alarm rate by 50%, reduce overall system energy consumption by 40%, improve edge computing efficiency by 200%, and improve predictive maintenance accuracy to 90%. These technological innovations provide a comprehensive solution for upgrading transformer winding condition monitoring systems, effectively ensuring the safe and stable operation of power systems. Simulation tests were conducted on an intelligent transformer winding condition monitoring system to verify the effectiveness of the proposed method. The simulation environment uses the MATLAB / Simulink platform, with hardware configuration including an Intel Xeon Gold 6248 processor (2.5GHz), 256GB DDR4 RAM, and 1TB NVMe SSD storage. The simulation model includes key components such as a multi-channel high-precision synchronous sampling unit, a digital twin algorithm module, and a reinforcement learning algorithm module. Parameter settings reference the technical specifications of the actual transformer winding monitoring system.

[0121] Example 2, refer to Figures 2-9 This invention provides an edge-intelligent data acquisition method for transformer windings, as one embodiment of the present invention. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0122] In an embodiment of the present invention, Figure 2 The box plot comparison of different methods in this invention in terms of data processing latency mainly demonstrates the significant advantages of the fusion method in this paper in terms of latency and stability compared with traditional methods, digital twin methods and reinforcement learning methods, with a performance improvement of up to 63.55%.

[0123] Figure 2 This diagram compares the cumulative distribution function (CDF) results of different methods in terms of data processing latency. The graph was generated using MATLAB code, and the main steps included data preparation, random sample generation, and CDF curve plotting. Figure 2As can be seen, the fusion method proposed in this embodiment has an average latency of 45.82 ms and a standard deviation of 3.91, which is significantly better than the traditional method (152.37 ms, 12.45), the digital twin method (98.64 ms, 8.73), and the reinforcement learning method (87.21 ms, 7.59). This result shows that the embodiment of this invention has a significant advantage in terms of data processing latency, reducing the average latency by 63.55% compared to the traditional method, 53.48% compared to the digital twin method, and 47.32% compared to the reinforcement learning method.

[0124] Furthermore, Figure 3 The paper presents a comparative analysis of the system accuracy over time under typical working conditions, mainly showing the learning and optimization process of the proposed method, which continuously improves from an initial 95.31% to 99.57%. Compared with digital twin modeling, reinforcement learning optimization and traditional methods, it shows better convergence performance and final accuracy.

[0125] Figure 4 The paper presents a time-series comparison of different algorithms in terms of resource utilization, mainly demonstrating the resource utilization efficiency of the proposed method from 15.32% to 14.12%, highlighting its significant advantages and continuous optimization capabilities in resource consumption control compared to other methods.

[0126] Figure 5 The paper demonstrates the dynamic trends of each method in terms of system response time, mainly showcasing the performance improvement of the proposed method from 78.32ms to 75.47ms, highlighting its significant real-time advantage compared to digital twin modeling and reinforcement learning optimization methods.

[0127] Figure 6 The time series comparison of different methods in terms of system robustness mainly demonstrates that the method in this paper maintains a robustness level of over 97%, showing its advantages in stability and reliability in complex environments compared to other methods.

[0128] Figures 3-6 This further illustrates a comparison of equipment status identification accuracy under typical operating conditions. The figure, generated using MATLAB code, comprises four subplots comparing the changes in accuracy, resource utilization, response time, and robustness over time. Figure 3 As can be seen, the accuracy of the method in this embodiment of the invention has been continuously improved from the initial 95.31% to 99.57%, which is significantly higher than other methods. Figure 4 The results show that the resource utilization rate of this method gradually decreased from 15.32% to 14.12%, which is far lower than that of the comparison method. Figure 5 This indicates that the response time of this method has been optimized from 78.32 milliseconds to 75.47 milliseconds, demonstrating a significant advantage. Figure 6This confirms that the proposed method has high robustness, stabilizing at around 97%. These results fully demonstrate the superiority of the embodiments of the present invention in data feature extraction, decision optimization, resource scheduling, and anomaly handling.

[0129] It should be noted that, Figure 7 The paper presents a comparative analysis of different algorithms in terms of CPU utilization on edge devices, mainly demonstrating the significant advantages of the proposed method in terms of processor resource consumption compared with traditional methods, analysis method 1, and analysis method 2, with an efficiency improvement of 60.17%.

[0130] Figure 8 The study demonstrates the changing trends in memory consumption of various methods, highlighting the outstanding performance of the proposed method in terms of memory usage optimization compared to traditional methods, with a memory efficiency improvement of 45.08%, meeting the requirements of edge computing deployment.

[0131] Figure 9 The graph shows a comparative analysis of different algorithms in terms of energy consumption, mainly demonstrating the significant improvement in energy efficiency of the method of this invention compared with traditional methods, with energy consumption reduced by 43.83%, reflecting the effectiveness of the meta-learning strategy for energy efficiency optimization.

[0132] Figures 7-9 This paper presents a comparative analysis of the resource consumption of different algorithms on edge devices. The graph, also generated using MATLAB code, includes three sub-graphs: CPU utilization, memory usage, and energy consumption. Figure 7 It can be observed that the initial CPU utilization rate of the method in this embodiment of the invention is the lowest (23.47%), and it only increases to 34.82% after 30 seconds, which is much lower than other methods. Figure 8 The memory usage of this method increased from 156.23MB to 214.62MB, a relatively slow increase. Figure 9 The results show that the energy consumption of this method increases gradually from 67.32 mWh to 100.50 mWh, with a limited increase. These results confirm the significant advantages of the embodiments of the present invention in improving computing efficiency, memory management, and energy saving.

[0133] Based on the simulation results above, it can be inferred that the embodiments of the present invention solve the problems of insufficient real-time performance, high error rate in state assessment, and low system efficiency in traditional transformer winding condition monitoring systems. This method, by integrating digital twin and reinforcement learning technologies, achieves the integration of data acquisition and intelligent analysis, significantly improving the real-time performance, accuracy, and resource utilization efficiency of condition monitoring. Specifically, the embodiments of the present invention reduce data processing latency by more than 47%, increase the state recognition accuracy to 99.57%, and significantly reduce resource consumption. These improvements directly solve key technical challenges in edge intelligent data acquisition, providing strong support for the upgrade of intelligent transformer winding condition monitoring systems and effectively ensuring the safe and stable operation of the power system.

[0134] Compared to solutions using traditional centralized data processing architectures, this invention has significant advantages in real-time data processing and system response speed, mainly due to the design of the edge intelligent data acquisition module.

[0135] Compared to solutions that rely on passive response control that separates simple data acquisition and analysis, this invention, by introducing a collaborative mechanism of digital twins and reinforcement learning, demonstrates superior performance in terms of state assessment accuracy and resource utilization efficiency.

[0136] This invention is not a simple improvement or addition to existing monitoring systems, but a systematic innovation that integrates data acquisition and intelligent analysis, which fundamentally improves the performance of transformer winding condition monitoring.

[0137] Example 3 is an embodiment of the present invention. This embodiment also provides an electronic device applicable to an edge intelligent data acquisition method for transformer windings, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the edge intelligent data acquisition method for transformer windings as proposed in the above embodiments.

[0138] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements an edge intelligent data acquisition method for transformer windings as proposed in the above embodiment.

[0139] The storage medium proposed in this embodiment and the edge intelligent data acquisition method for transformer windings proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0140] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0141] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for edge intelligent data acquisition for transformer windings, characterized by: The method comprises the steps of: A multi-channel high-precision synchronous sampling unit is constructed to collect real-time multi-physical quantities of the transformer winding; Based on the electromagnetic thermal coupling characteristics of the transformer winding, a digital twin model is constructed to assist in state evaluation; For multi-source heterogeneous data of the transformer winding, singular value decomposition and attention mechanism are combined for feature extraction; A reinforcement learning algorithm is introduced to dynamically optimize the extracted features, dynamically adjust the sampling strategy according to the winding state, and detect abnormalities in key state indicators, and evaluate performance through comprehensive performance indicators; The digital twin model comprises constructing a digital twin model based on the electromagnetic thermal coupling characteristics of the transformer winding, accurately reproducing the dynamic behavior of the winding through a virtual model, and assisting in state evaluation and predictive maintenance: x t+1 = A(x t )u t +B(x t )u t +w t where A(x t )∈R n×n is a nonlinear state transition matrix, B(x t )∈R n×m is a nonlinear control matrix, u t ∈R m is a control input vector, w t ∈R n is a process noise vector, n represents the dimension of the system state variable, x t+1 is an M-channel sampling data vector at t+1, and m represents the dimension of the control input vector. The dynamic optimization of the extracted features includes introducing a reinforcement learning algorithm, defining a comprehensive value function J(s t ,a t ) as follows: wherein D virtual physical respectively represent the virtual model and the physical system state, R winding (t+τ) is the winding state evaluation reward function, γ τ is a discount factor for future τ steps, β1, β2, β3 are weight coefficients, ε t+τ is a model uncertainty term or an environmental fluctuation penalty term, is the expected comprehensive value function under the current state and action; represents the expected value of the sampling trajectory for future τ steps under the experience data distribution D, t is a time step, Time is the total time step, a t represents the action performed at time step t in s t characteristic extraction;​ Based on the current comprehensive value function, the Q function is updated: wherein Q model and Q data are the Q-value functions based on the digital twin model prediction and actual data, respectively, λ∈[0, 1] is a balancing factor, Q integrated is the final integrated policy function, a’ is the optimal action at the next time step, s t+1 represents the result of feature extraction at t+1. At the same time, introduce the edge computing efficiency factor η edge , the modified Q function update: Q edge (s t ,a t )=η edge Q integrated (s t ,a t )+(1-η edge )Q local (s t ,a t ) where Q local is the Q-value estimate function based only on local data.

2. A method of edge smart data collection for a transformer winding as claimed in claim 1, wherein: The multi-channel high-precision synchronous sampling unit comprises a high-resolution ADC with a resolution of not less than 24 bits and a sampling rate adjustable range of 10-100 kHz, and is used to collect real-time multi-physical quantities of the transformer winding, and the sampling process is represented as: wherein x t ∈R M is a vector of M-channel sampled data, including physical quantities of voltage current temperature magnetic field n t ∈R M is a vector of sampling noise, R is a real number set, and T is a transpose matrix; The sampling time synchronization error is controlled within microseconds: wherein t i denotes the sampling time of the i-th channel, t0is the reference time, M is the number of sampling channels, and i is the index of the variable.

3. A method of edge smart data collection for a transformer winding as claimed in claim 2, wherein: The singular value decomposition and attention mechanism are combined for feature extraction, which is suitable for the characteristics of multi-source heterogeneous data of the transformer winding: where Enc θ (·) is a parameterized encoder function, θ is the encoder parameters, UΣV T is the SVD decomposition result, Attn(·) is an attention mechanism function, ⊙ denotes the Hadamard product, s t denotes the result of feature extraction; Results of feature extraction s t The state evaluation is input as a system state encoding vector, compared with the prediction result of the digital twin model, and the state estimate is updated.

4. A method of edge smart data collection for a transformer winding as claimed in claim 3, wherein: The adaptive sampling frequency adjustment function is set to dynamically adjust the sampling strategy according to the winding state: f s (t+1) = f s (t) + κ · ∇ f Q edge (s t , a t ) where f s (t) is the sampling frequency at time t, k is the learning rate, abla f denotes the gradient operator with respect to the frequency.

5. A method of edge smart data collection for a transformer winding as claimed in claim 4, wherein: The abnormal value evaluation function is defined for high-sensitivity abnormal detection of key state indicators on the edge device side, and the abnormal value evaluation function is as follows: where Ano(t) is the anomaly score obtained by the anomaly evaluation function, x t is the actual sampling state value at time t, represents the predicted value of the system state based on the constructed digital twin model, represents the L2 norm, KL(·||·) is the KL divergence, ω is the weight coefficient, p(x t ) and are the probability distributions of the actual observation value and the predicted value, respectively; The abnormal score obtained by the abnormal value evaluation function is used as an input variable for performance evaluation and sampling strategy adjustment.

6. A method of edge smart data collection for a transformer winding as claimed in claim 5, wherein: The performance is evaluated through comprehensive performance indicators, which comprises: Wherein, Perfor is the evaluation result of evaluating performance by comprehensive performance index, Accuracy winding (t) is the winding state diagnosis accuracy rate function, Latency -1 (t) is the system delay function, EnergyE(t) is the energy efficiency function, and α1, α2, α3 are weight coefficients of the three types of comprehensive performance indexes. The meta-learning update rule is introduced to improve the adaptive ability: z t+1 = z t - δabla ζ L(D trin , D val ; z t ) where ζ t+1 is the model parameter after introducing the meta-learning update rule, ζ t is the original model parameter, and δ is the meta-learning rate, is the meta-objective function, D trin and D val are the training set and the validation set, respectively, and abla ζ denotes the gradient operator with respect to the parameter ζ. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the edge intelligent data acquisition method for the transformer winding in any one of claims 1-6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to realize the steps of the edge intelligent data acquisition method for the transformer winding in any one of claims 1-6.

Images