Transformer health assessment and technical improvement decision optimization method

Abstract

The invention relates to the field of transformers, in particular to a transformer health assessment and technical improvement decision optimization method. Comprising the following steps: S1, acquiring a multi-modal original data set by using an infrared dual-spectrum unmanned aerial vehicle, a vibration sensor array and a gas-in-oil sensor; s2, respectively preprocessing the image, the vibration signal and the oil chromatography data through a generative adversarial network, wavelet transform and moving average processing; s3, inputting the preprocessed data into a multi-modal deep learning model, and generating gas trend prediction, hot spot area positioning and mechanical fault judgment results; s4, calculating health scores of all dimensions and fusing the health scores into a comprehensive health score; and S5, generating a hierarchical maintenance strategy, and constructing a multi-objective optimization model to solve an optimal technical improvement scheme sequence when technical improvement conditions are met. The problems that in the prior art, data collection is not comprehensive, fault prediction depends on artificial experience, long-term health trend prediction is lacked, and a scientific basis cannot be provided for technical improvement decision making are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of transformers, in particular to a transformer health assessment and technical improvement decision optimization method. BACKGROUND

[0002] With the continuous development of power systems, as one of the core equipment, the health management and fault prevention of transformers become particularly important. Traditional transformer monitoring methods usually rely on periodic inspection and manual diagnosis, mainly through manual reading of sensor data or manual analysis of test results to judge the running state of the equipment. These methods have obvious limitations. First, the data collection is not comprehensive enough to reflect the real-time health status of the equipment. Second, fault prediction relies on human experience, which is difficult to be accurate and efficient, and may lead to failure to discover equipment failure in time, or even over-maintenance. In addition, traditional methods usually lack long-term health trend prediction for equipment, and cannot provide scientific basis for technical improvement decision of equipment. Although existing technologies have gradually introduced sensors and monitoring systems, combined with some automatic equipment for real-time data monitoring, these technologies still face problems such as single data and low processing efficiency. Common monitoring technologies such as temperature, vibration and oil chromatographic data can provide certain equipment running state information, but these data often exist in isolation, lacking effective fusion and comprehensive analysis. Therefore, although there are some automatic monitoring systems, it is still difficult to achieve comprehensive and accurate assessment of the health status of transformer equipment, and most technologies cannot accurately predict the future operation risk of equipment, lacking predictability of potential faults. For example, the power distribution transformer health index evaluation method mentioned in patent CN105404936A can evaluate the health degree by obtaining the state index data of the transformer, but there are still deficiencies in the fusion processing of multi-modal data and the application of deep learning model.

[0003] Therefore, the present application proposes a transformer health assessment and decision optimization method based on multi-modal data and deep learning model. SUMMARY

[0004] The purpose of the present application is to provide a transformer health assessment and technical improvement decision optimization method to solve the problems of incomplete data collection, fault prediction relying on human experience, lack of long-term health trend prediction for equipment and inability to provide scientific basis for technical improvement decision in the prior art.

[0005] To achieve the above purpose, the following technical solutions are adopted.

[0006] A transformer health assessment and technical improvement decision optimization method, comprising the following steps,

[0007] Step S1, collect transformer surface image and infrared temperature field data by infrared dual-spectrum unmanned aerial vehicle, collect vibration signal by vibration sensor array, and collect oil chromatogram data of transformer oil by online gas sensor to form a multi-modal original data set;

[0008] Step S2, adopt a generative adversarial network to process the image data collected in step S1 to remove fog, adopt wavelet transform to denoise the vibration signal, and adopt sliding average smoothing processing to the oil chromatogram data to form a preprocessed data set;

[0009] Step S3, input the preprocessed data set into a multi-modal deep learning model, process the oil chromatogram data by a time series analysis module to generate a gas trend prediction result, process the defogging image by an image segmentation module to generate a hot spot area positioning result, and process the denoised vibration signal by a vibration frequency spectrum analysis module to generate a mechanical fault judgment result in combination with a partial discharge detection result;

[0010] Step S4, according to the fault recognition result of step S3, calculate the health scores of the oil chromatogram, temperature, vibration and partial discharge dimensions respectively, and fuse them into a comprehensive health score through a dynamic weight distribution strategy;

[0011] Step S5, based on the comprehensive health score and historical fault records, generate a graded maintenance strategy, when the technical improvement trigger condition is met, construct a multi-objective optimization model containing health improvement, cost control and power outage constraint, and solve the optimal technical improvement scheme sequence.

[0012] Optionally, the step S2 includes adopting a generative adversarial network to process the infrared dual-spectrum unmanned aerial vehicle image data collected in step S1 to remove fog, generating a defogging image data set, and optimizing the image clarity through the adversarial training of the generator and the discriminator of the generative adversarial network;

[0013] Perform wavelet transform decomposition on the vibration sensor array signal collected in step S1, and reconstruct the denoised vibration signal set after removing high-frequency noise;

[0014] Perform sliding average smoothing processing on the oil chromatogram data collected in step S1 to generate a stable oil chromatogram data set;

[0015] Combine the defogging image data set, the denoised vibration signal set and the stable oil chromatogram data set to form a preprocessed data set for inputting into the multi-modal deep learning model of step S3.

[0016] Optionally, in step S3, the time series analysis module adopts a long short-term memory network LSTM to analyze the time series characteristics of the oil chromatogram data, and when the target gas ratio exceeds a dynamically adjusted first threshold value and the monthly change amount of hydrogen concentration exceeds a dynamically adjusted second threshold value, a high temperature overheating fault judgment signal is triggered;

[0017] The image segmentation module adopts an improved U-Net model to locate the hotspot area in the infrared thermal image, and the U-Net model eliminates the environmental temperature interference through a temperature compensation module in the encoder;

[0018] The vibration spectrum analysis module adopts a one-dimensional convolutional neural network 1D-CNN to extract the frequency spectrum features of the vibration signal, and generates a core loose fault determination result when the energy of a specific frequency band exceeds a dynamically adjusted third threshold value and the Mel frequency distribution of the voiceprint signal is abnormal.

[0019] Optionally, in the step S4, the oil chromatographic health score is based on a weighted calculation of the target gas ratio and the hydrogen concentration change amount, and the weight is dynamically adjusted according to historical fault data;

[0020] The temperature health score is determined according to the difference between the maximum temperature difference of the hotspot area and a preset threshold value, and the preset threshold value is set in combination with the transformer model;

[0021] The vibration health score is calculated based on the deviation of the vibration energy from a normal threshold value, and the normal threshold value is dynamically updated through baseline data;

[0022] The partial discharge health score is determined according to the over-standard situation of the discharge intensity and the pulse frequency, and the over-standard threshold value is dynamically optimized based on an insulation aging model;

[0023] The comprehensive health score is dynamically assigned a weight of each dimension by an entropy weight method and is calculated by weighted averaging, and a high-risk equipment priority processing instruction is triggered when the comprehensive score is lower than a preset risk threshold value.

[0024] Optionally, in the step S5, the hierarchical maintenance strategy includes emergency repair, planned maintenance and continuous monitoring, and when the annual recurrence number of similar defects exceeds a dynamically adjusted fourth threshold value or the cumulative maintenance cost exceeds a device residual value proportion threshold value, a technical improvement project evaluation is triggered;

[0025] The multi-objective optimization model aims to maximize the health score improvement, minimize the transformation cost and power outage time, generates a non-dominated solution set through an improved NSGA-III algorithm, and the algorithm adopts a dynamic hierarchical screening mechanism to divide the scheme levels.

[0026] Optionally, the improved NSGA-III algorithm includes:

[0027] A candidate scheme set that meets the health score improvement and cost constraints is screened out through non-dominated sorting;

[0028] The high-dimensional target space is uniformly partitioned based on an adaptive reference point strategy to ensure the diversity and coverage of the solution set;

[0029] A hybrid selection strategy is adopted to combine tournament selection and regional preference guidance to drive the population to evolve towards a region with high health score improvement rate and optimal cost-effectiveness.

[0030] The dynamic crossover and mutation operators are designed based on decision variables, and the global exploration and local development capabilities are balanced.

[0031] Optionally, in the step S1, the infrared dual-spectrum unmanned aerial vehicle synchronously collects visible light and infrared images, and marks the coordinates of the key area through GPS positioning, and the coordinates are matched with the three-dimensional model of the transformer to realize the visualization of the fault area.

[0032] The vibration sensor array is arranged on the transformer base and the shell, and the multi-axis sensor is used to collect vibration signals in real time, and the multi-axis signals are filtered by Kalman filtering to eliminate installation error interference.

[0033] The oil gas sensor monitors the concentration of dissolved gas in the oil, and generates oil chromatogram time sequence data in combination with the time stamp.

[0034] Optionally, in the step S3,

[0035] The partial discharge detection collects the discharge pulse signal through the high-frequency sensor, and when the discharge intensity exceeds the dynamically adjusted fifth threshold value or the pulse frequency exceeds the dynamically adjusted sixth threshold value, the electrical insulation fault is determined.

[0036] The cross-modal matching is realized by comparing the vibration frequency spectrum characteristics with the Mel frequency distribution of the voiceprint signal, and the Mel frequency distribution is extracted after the voiceprint signal is preprocessed by the filter bank.

[0037] Optionally, in the step S4,

[0038] The dynamic weight distribution strategy is based on the influence degree of each dimension on the reliability of the equipment, and is dynamically optimized by the entropy weight method combined with historical fault data.

[0039] When the health score of any dimension is lower than the preset alarm threshold, the weight proportion of the dimension is forcibly increased, and the real-time health state warning is triggered.

[0040] Optionally, in the step S5, the technical improvement scheme optimization includes budget constraint, high-risk equipment coverage constraint and power outage time constraint, and the budget constraint is allocated by a dynamic programming algorithm.

[0041] The non-dominated solution set is screened by the crowding distance to select the scheme with the highest diversity, and the crowding distance is determined based on the comprehensive calculation of the health score improvement rate and the cost benefit rate.

[0042] Compared with the prior art, the present application has the following beneficial effects:

[0043] The method of the present application comprehensively obtains the operation state information of the transformer through multi-modal data acquisition, including the transformer surface image and infrared temperature field data collected by the infrared dual-spectrum unmanned aerial vehicle, the vibration signal collected by the vibration sensor array, the transformer oil chromatogram data monitored by the gas sensor in the oil, etc., forming a rich multi-modal original data set, providing a data basis for subsequent accurate evaluation. Then, the collected data is respectively processed by the generative adversarial network to remove fog from the image data, the wavelet transform to denoise the vibration signal, and the moving average smoothing to process the oil chromatogram data, effectively improving the quality and reliability of the data, making the subsequent analysis more accurate. Then, the preprocessed data is input into the multi-modal deep learning model, and different types of data are processed by the time series analysis module, the image segmentation module, the vibration spectrum analysis module, etc., to generate gas trend prediction results, hot spot area positioning results, mechanical fault determination results, etc., realizing intelligent identification of multiple potential faults of the transformer. Then, the health scores of each dimension are calculated according to the fault identification results, and the dynamic weight distribution strategy is used to fuse them into a comprehensive health score, so as to comprehensively and accurately reflect the overall health status of the transformer. Finally, based on the comprehensive health score and historical fault records, a graded maintenance strategy is generated, and a multi-objective optimization model is constructed to solve the optimal technical improvement scheme sequence when the technical improvement trigger condition is met, realizing scientific decision and optimization of transformer maintenance and technical improvement, improving the intelligent level of equipment management, reducing unplanned shutdown, and reducing the production cost of the power grid.

[0044] The present application further refines the processing method of different types of data in step S2, making the fog removal, denoising and smoothing process more specific and concrete, which helps to further improve the effect and quality of data preprocessing, provides more reliable input data for subsequent deep learning model analysis, and thus improves the accuracy of fault identification and the precision of health assessment.

[0045] The present application limits the specific implementation methods and determination conditions of each module of the multi-modal deep learning model in step S3, such as using long short-term memory network LSTM to analyze the time series characteristics of oil chromatogram data in the time series analysis module, using an improved U-Net model to locate the hot spot area in the image segmentation module, and using one-dimensional convolutional neural network 1D-CNN to extract the frequency spectrum characteristics of the vibration signal in the vibration spectrum analysis module. These specific implementation methods can more effectively extract features and information from the data, improve the accuracy of fault determination, and make the fault identification more scientific and reliable.

[0046] The application specifies the calculation method of each dimension health score and the fusion strategy of comprehensive health score in step S4. Through calculation based on specific indicators and weights, the health status of the transformer in various aspects can be more accurately quantified. Through dynamic weight distribution and entropy weight method weighted average calculation of comprehensive health score, the scoring result is more scientific and reasonable, and better reflects the real health status of the transformer, providing a strong basis for subsequent maintenance and technical improvement decision.

[0047] The application specifies the hierarchical maintenance strategy and multi-objective optimization model in step S5. According to the fault severity, different maintenance strategies are formulated, which can more reasonably arrange maintenance resources and time, and improve maintenance efficiency. At the same time, the multi-objective optimization model is constructed to maximize the health score improvement, minimize the transformation cost and power outage time, and the optimal transformation scheme sequence is solved by improved NSGA-III algorithm, which realizes the optimization selection of transformation scheme under the condition of meeting various constraints, balances the relationship between equipment health improvement and cost control, and improves the scientificity and economy of transformation decision.

[0048] The application describes the specific steps and strategies of the improved NSGA-III algorithm. Through non-dominated sorting, adaptive reference point strategy, hybrid selection strategy, and dynamic crossover and mutation operators, high-quality transformation schemes can be effectively searched and selected to ensure the diversity and coverage of the solution set, improve the global exploration ability and local development ability of the algorithm, and provide strong support for obtaining better transformation decision.

[0049] The application supplements the specific data collection methods and related contents of infrared dual-spectrum unmanned aerial vehicle, vibration sensor array and oil gas sensor in step S1, such as synchronously collecting visible light and infrared images by unmanned aerial vehicle and marking key area coordinates by GPS positioning, deploying position of vibration sensor array and eliminating installation error interference by Kalman filter, generating oil chromatogram time series data by oil gas sensor combined with time stamp, etc. These details further improve the data collection process, which helps to improve the accuracy and integrity of the data, and provides better data support for subsequent analysis and evaluation.

[0050] The application limits the specific method of partial discharge detection and cross-modal matching related content in step S3, and specifies that the discharge pulse signal is collected by high-frequency sensor and the electrical insulation fault is determined according to specific conditions, and the Mel frequency distribution is extracted after the voiceprint signal is preprocessed by filter bank for cross-modal matching. The way can more accurately identify the partial discharge phenomenon and related faults, and improve the monitoring and evaluation ability of the electrical insulation state of the transformer.

[0051] The application further refines the content of the dynamic weight allocation strategy in step S4, emphasizes the influence degree of each dimension on the reliability of the device, and dynamically optimizes through the entropy weight method combined with historical failure data, and in specific cases, the mechanism of forcibly increasing the weight proportion of a certain dimension and triggering real-time health state warning, so that the calculation of the health score is more in line with the actual situation, the potential risks of the transformer can be found and focused in time, and the reliability and warning ability of the evaluation result are enhanced.

[0052] The application supplements the specific content of the optimization of the technical improvement scheme in step S5, including the specific implementation mode of the budget constraint, the high-risk device coverage constraint and the power outage time constraint, and the screening method of the non-dominated solution set, and through the dynamic programming algorithm, the funds are allocated and the scheme with the highest diversity is screened based on the congestion distance, etc., further perfecting the optimization process of the technical improvement decision, ensuring the feasibility and effectiveness of the technical improvement scheme under the satisfaction of various constraint conditions, and improving the scientificity and rationality of the technical improvement decision. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a step flow schematic diagram of an embodiment of the transformer health evaluation and technical improvement decision optimization method of the application.

[0054] Figure 2 is a step 2 sub-step schematic diagram of an embodiment of the transformer health evaluation and technical improvement decision optimization method of the application.

[0055] Figure 3 is a step 3 sub-step schematic diagram of an embodiment of the transformer health evaluation and technical improvement decision optimization method of the application.

[0056] Figure 4 is a step 4 sub-step schematic diagram of an embodiment of the transformer health evaluation and technical improvement decision optimization method of the application.

[0057] Figure 5 is a step 5 sub-step schematic diagram of an embodiment of the transformer health evaluation and technical improvement decision optimization method of the application. DETAILED DESCRIPTION

[0058] The application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments. It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0059] The following detailed description is exemplary description, which is intended to provide further detailed description of the application. Unless otherwise specified, all technical terms used in the application are the same as the meanings understood by the general technical personnel in the field to which the application belongs. The terms used in the application are only for describing the specific embodiments, and are not intended to limit the exemplary embodiments according to the application.

[0060] The application provides a transformer health evaluation and technical improvement decision optimization method based on multi-modal data and a deep learning model, which provides decision support for the technical improvement and maintenance of power grid equipment and improves the scientificity and rationality of projects.

[0061] The above technical purpose of the application is achieved by the following technical scheme:

[0062] As shown in Figures 1-5 A transformer health evaluation and technical improvement decision optimization method based on multi-modal data and a deep learning model, which specifically comprises the following steps:

[0063] Step S1, data acquisition is performed, and a variety of sensors and devices are deployed to comprehensively acquire data of the transformer. High-resolution images and infrared temperature field data of the surface of the transformer are acquired by using an infrared dual-spectrum unmanned aerial vehicle, which are used to detect temperature abnormalities and surface damage; vibration signals of the transformer are acquired by using a vibration sensor array, which are used to analyze the running state of the equipment and identify potential mechanical faults; meanwhile, oil chromatographic data of the transformer are monitored by using an oil gas sensor, which are used to evaluate the influence of the external environment on the equipment.

[0064] The infrared dual-spectrum unmanned aerial vehicle is used to fly and acquire images of the surface of the transformer. Meanwhile, temperature field data are acquired by using an infrared sensor, and special attention is paid to the key areas of the transformer. The time stamp, area position and image content of the images and the temperature data are recorded.

[0065] Vibration sensors are deployed at the key positions of the transformer, including the base and the shell. The vibration signals of the equipment under different running conditions are recorded. The vibration data are acquired in real time by using the sensors, and the time stamp and frequency information of each data point are recorded.

[0066] The oil gas sensor is deployed, and the oil chromatographic data of the transformer are acquired regularly. The environmental parameters including the salt density and the humidity and the time stamp of each data point are recorded. The sensors are used to monitor in real time according to the change rate.

[0067] The electrical parameters such as the current and the voltage of the transformer are monitored in real time by using the SCADA system.

[0068] The load, the current and the voltage data of the equipment are recorded regularly, which provide background information for the health evaluation.

[0069] Step S2, after the data acquisition, the fog and haze interference in the images is removed by using a generative adversarial network, the wavelet transform is used to denoise the vibration signals, and the signal-to-noise ratio of the signals is enhanced. The stability of the oil chromatographic data is improved by using data smoothing processing. Finally, the cleaned and enhanced data provide a reliable basis for the subsequent defect identification and health evaluation.

[0070] Image data fog removal and enhancement

[0071] The generative adversarial network performs image defogging processing through two opposing network structures, a generator and a discriminator. The generator is responsible for generating defogged images, while the discriminator evaluates the authenticity of the generated images. Through adversarial training, the defogging effect of the generator is gradually optimized, thereby removing the haze interference in the image, enhancing the clarity of the image, and ensuring that the quality of the image is not affected by the haze during subsequent defect identification.

[0072] Adversarial loss function:

[0073]

[0074] Cycle consistency loss:

[0075]

[0076] Total loss function:

[0077]

[0078] Generator structure: deconvolutional convolution

[0079] G(z) = Dec 反卷积 (Enc 卷积 (z))

[0080] Discriminator structure:

[0081] D(x) = PatchGAN(x)

[0082] Where G represents the generator, the input haze image z, and the output clear image G(z). D represents the discriminator, which distinguishes between real images x and generated images G(z). data p represents the probability distribution of clear images.

[0083] p z represents the probability distribution of haze images. F represents the inverse generator (clear → haze) λ cyc = 10 represents the cycle loss weight. Convolution Enc 卷积 represents a 4-layer convolutional encoder (channel 64 → 256). Deconvolution Dec 反卷积 represents a 4-layer deconvolutional decoder (channel 256 → 64). PatchGAN represents a local image block discrimination network (output 30 × 30 matrix)

[0084] Wavelet denoising of vibration signals

[0085] Wavelet transform effectively distinguishes useful parts from noise in the signal by decomposing the signal into components of different frequencies. It can separate high-frequency noise from low-frequency effective signals in vibration signals, retain important low-frequency features, and remove high-frequency environmental noise or equipment background noise, thereby improving the signal-to-noise ratio of vibration signals and making subsequent fault diagnosis more accurate.

[0086] Wavelet decomposition:

[0087]

[0088] Adaptive threshold:

[0089]

[0090] Improved soft threshold:

[0091]

[0092] Signal reconstruction:

[0093]

[0094] where, represents the original vibration signal, sampling rate 10 kHz, ψ j，k represents the db5 wavelet basis.

[0095] d j represents the jth layer detail coefficient. σ j represents the jth layer noise standard deviation estimate. N represents the signal length, default 10,000 points. a5 represents the 5th layer approximation coefficient. represents the denoised detail coefficient

[0096] Oil chromatogram data smoothing

[0097] Smooth the oil chromatogram data to reduce short-term fluctuations in the data and improve data stability.

[0098] Sliding window filtering:

[0099]

[0100] Boundary extension:

[0101] r -1 = r1, r0 = r2, r n+1 = r n-1 , r n+2 = r n-2

[0102] Outlier correction:

[0103]

[0104] where r k represents the kth hourly change rate (unit: μg / cm 2 ·d). represents the smoothed ith hourly data. W=5 represents the size of the sliding window. σ represents the standard deviation of the data within the window.

[0105] n is the total length of the data.

[0106] Step S3, intelligent identification and maintenance decision of transformer defects are realized by constructing a multi-modal deep learning model. Combining visual and vibration signal data sources, deep learning technology is used for defect identification and technical improvement maintenance decision. First, the LSTM network is used to analyze the oil chromatogram data to predict the gas growth trend. When the ratio exceeds 0.1 and the hydrogen concentration change exceeds 50 ppm / month, it is automatically determined as a high temperature overheating fault. Then, the improved U-Net model is used to process the infrared thermal imaging data to locate the hot spot area of the transformer and determine whether there is an overheating fault through temperature difference analysis. Next, the 1D-CNN model is used for frequency spectrum analysis of the vibration signal, and the soundprint signal is combined for cross-modal matching. When the vibration energy exceeds 3σ and the soundprint pulse count is abnormal, it is determined as a core loose fault. In addition, the partial discharge detection method is used to identify electrical faults to ensure timely detection of insulation problems.

[0107] (1) LSTM network analyzes oil chromatogram data

[0108] Long Short-Term Memory (LSTM) is a special type of Recurrent Neural Network (RNN) that is well suited for processing time series data. For oil chromatogram data, LSTM can capture the trend of gas concentration changes over time and predict future gas growth trends, thereby determining whether the equipment has a high temperature overheating fault. By inputting oil chromatogram data, the LSTM model can identify the pattern of gas concentration changes and automatically determine a high temperature overheating fault when the ratio exceeds 0.1 and the hydrogen concentration change exceeds 50 ppm / month.

[0109] The core of the LSTM network is to capture long-term dependencies in time series through hidden states. The network calculates the following formula:

[0110] Input gate (controls the flow of new information):

[0111] i t = σ(W xi x t +W hi h t-1 +b i )

[0112] Based on the current input and the hidden state at the previous time, the activation value of the input gate is calculated to determine how much information to retain.

[0113] Forget gate (filtering historical memory):

[0114] f t = σ(W xf x t +W hf h t-1 +b f )

[0115] Based on the current input x t and the hidden state h t-1 of the last moment, decide how much memory information to forget.

[0116] Memory cell update (fuse new and old information):

[0117] C t = f t ⊙C t-1 +i t ⊙tanh(W xc x t +W hc h t-1 +b c )

[0118] The cell state C t is controlled by the forget gate f t The cell state C t-1 of the previous moment and the new information controlled by the input gate i t , is updated through the weighted combination of the current input x t and the hidden state h t-1 of the last moment.

[0119] Output gate (generate prediction results):

[0120] h t = o t ⊙tanh(C t ), o t = σ(W xo x t +W ho h t-1 +b o )

[0121] Based on the current input x t and the hidden state h t-1 of the last moment, decide the activation value of the output gate.

[0122] Where W x , W h represent the weight matrix, used for linear transformation of input x t and the hidden state ht-1 , control the strength of information transmission. x acts on the current input x t , W h acts on the hidden state h t-1 of the previous moment. i , b f , b c ,

[0123] b o : represents the bias term, used to adjust the activation value of each gate. σ represents the Sigmoid activation function, with an output range of [0, 1], used to control the proportion of information flow. tanh represents the hyperbolic tangent activation function, with an output range of [-1, 1], used to generate the output of the cell state and hidden state.

[0124] Fault determination rule

[0125] Trigger high-temperature overheating alarm when the following conditions are met simultaneously: and

[0126] and ΔH2> 50 ppm / month

[0127] (2) Improved U-Net model for processing infrared thermal imaging data

[0128] U-Net is a convolutional neural network (CNN) architecture for image segmentation, especially suitable for regional identification of medical images. In this patent, the improved U-Net model is used to process the infrared thermal imaging data of the transformer, locate the hot spot area of the device, and determine whether there is an overheating fault through temperature difference analysis.

[0129] To address the problem of environmental temperature interference on outdoor equipment, a temperature compensation module is embedded in the improved U-Net structure to eliminate the influence of environmental temperature difference on hot spot detection.

[0130] Temperature compensation formula: Corrected measured environmental

[0131] T 修正 = T 实测 - 0.8 (T 环境 - 25℃)

[0132] Where, the corrected T 修正 is the corrected temperature value, representing the temperature obtained after environmental compensation. The measured T 实测 is the measured temperature value, the original collected temperature data. The environmental T 环境 represents the ambient temperature, referring to the temperature of the environment where the sensor is located.

[0133] Encoder-decoder structure:

[0134] Encoder (4 convolutional layers, 64→256 channels):

[0135] F enc =Conv 3×3 (ReLU(BN(x)))

[0136] Where F enc This represents the encoder's output feature map. `Conv3x` represents a 3x3 convolution operation, indicating that a convolution operation is applied to the input feature map to extract spatial features. `ReLU` is an activation function, a non-linear transformation. `BN` represents batch normalization, used to accelerate training and stabilize the learning process. `x` is the input data or feature map.

[0137] Decoder (4 layers of deconvolution, 256 channels → 64):

[0138] M hotspot =DeConv 3×3 (ReLU(BN(F enc )))

[0139] Among them, M hotspot This represents the thermal feature map output by the decoder, used to identify regions with abnormal temperatures. `DeConv3x3` represents a 3x3 deconvolution operation used to restore image resolution. `ReLU` represents the activation function, adding non-linearity. `BN` represents batch normalization, used to normalize the data.

[0140] Fault Judgment Rules

[0141] An alarm is triggered when the temperature difference in critical areas exceeds the threshold: Correct adjacent...

[0142] ΔT=T 修正 -T 相邻 >15K

[0143] (3) 1D-CNN model analysis of vibration signals

[0144] 1D Convolutional Neural Networks (1D-CNNs) are specifically designed for processing one-dimensional sequential data. In vibration signal analysis, 1D-CNNs extract features from vibration signals through convolution operations to identify early signs of faults. Combined with acoustic signature signals, cross-modal matching is performed to further determine whether mechanical faults, such as loose iron cores, exist.

[0145] Vibration signal analysis:

[0146]

[0147] Where E vib Represents vibrational energy, with units of m / s. 2 |X(f)| 2The square of the spectrum amplitude of the vibration signal represents the energy of the vibration signal at a specific frequency f. f represents the frequency, and the integral range is 100 Hz to 400 Hz.

[0148] Convolution kernel response:

[0149]

[0150] where y i represents the output of the convolution layer. x i+k-1 represents the i+k-1th value in the input signal x. w k represents the kth weight value of the convolution kernel. b represents the bias term. ReLU is a rectified linear unit, an activation function, used for nonlinear mapping.

[0151] This formula represents the processing of the input signal x by the convolution operation using 5 convolution kernels w k to extract signal features. Then the output of the convolution is nonlinearly transformed by the ReLU activation function.

[0152] Voiceprint signal processing:

[0153]

[0154] where Mel(f) is the Mel frequency scale, representing the frequency of the audio signal. f represents the frequency.

[0155] When the following conditions are met simultaneously, it is determined that the core is loose: and the next minute

[0156] E vib > 3σ vib and N audio > 20 times / minute

[0157] Partial discharge detection

[0158] The detection of partial discharge is usually through high-frequency sensors or current sensors to collect signals. Its signal is usually a pulse signal, which can be analyzed by analyzing its amplitude, frequency and distribution characteristics to evaluate the health.

[0159] When the partial discharge intensity PD intensity and the pulse count N pd exceed the set threshold, it indicates that the equipment has an electrical insulation problem and needs further inspection. The judgment conditions are as follows:

[0160] or

[0161] PD intensity > threshold pd or N pd > threshold count

[0162] wherein:

[0163] threshold pd : set threshold of partial discharge intensity.

[0164] threshold count : set threshold of partial discharge pulse count.

[0165] When one of the two conditions is met, the device is determined to have an electrical fault.

[0166] Step S4, through the quantitative analysis of the diagnosis results such as vibration signal, infrared thermal image data, partial discharge signal, etc., a single health score will be obtained for each dimension. Then, combined with the weight of each dimension, these individual scores will be weighted and averaged to obtain a comprehensive health score. The lower the comprehensive score, the worse the health of the device, and the higher the risk of failure.

[0167] (1) Oil chromatographic data analysis

[0168] The change of gas concentration in transformer oil is analyzed, and the health condition of the oil is determined by the set threshold. By calculating the ratio of characteristic gases and the change of gas concentration, the oil chromatographic health score is obtained.

[0169] Oil chromatographic health score calculation formula:

[0170]

[0171] wherein: C C2H2 and C C2H4 are the concentrations of characteristic gases, ΔH2 is the change of hydrogen concentration, and α and β are weight coefficients, which can be adjusted according to actual conditions.

[0172] (2) Temperature health score

[0173] Based on the infrared thermal imaging data of the transformer, the temperature abnormality of the transformer bushing joint area is evaluated by analyzing the temperature difference. According to the temperature difference threshold, the temperature health condition is determined, and the temperature health score is calculated.

[0174] Temperature health score calculation formula:

[0175]

[0176] wherein: ΔT max is the maximum temperature difference, T threshold is the set temperature difference threshold, and if the temperature difference exceeds the threshold, the score will be reduced.

[0177] (3) Vibration health score

[0178] By analyzing the vibration spectrum data of the transformer, the vibration characteristics of a specific frequency band are extracted to evaluate the mechanical health status. When the vibration energy exceeds the normal range, it is determined as a potential mechanical failure, and the vibration health score is calculated accordingly.

[0179] Vibration health score calculation formula:

[0180]

[0181] Where:

[0182] E vibration is the actual measured vibration energy,

[0183] E threshold is the preset normal energy threshold, and if the vibration energy exceeds the standard, the score will be reduced.

[0184] (4) Partial discharge health score

[0185] Partial discharge detection evaluates the electrical insulation health status of the transformer by analyzing the frequency and intensity of partial discharge signals. When partial discharge is frequent and strong, it indicates that the electrical insulation is at risk of deterioration, thereby affecting the health score.

[0186] Partial discharge health score calculation formula:

[0187]

[0188] Where:

[0189] PD intensity is the intensity of partial discharge,

[0190] PD max is the maximum discharge intensity set, and if the discharge intensity is large, the health score is lower.

[0191] (5) Calculation of comprehensive health score

[0192] The comprehensive health score of the transformer is calculated by weighted average of each dimension health score. The health score of each dimension is assigned a corresponding weight according to its impact on the overall health of the device.

[0193] Comprehensive health score calculation formula:

[0194] HI = w1·HI oil + w2·HI temp + w3·HI vibration + w4·HI pd

[0195] Where:

[0196] w1, w2, w3, w4 are the weight coefficients of oil chromatography, temperature, vibration and partial discharge, respectively, satisfying:

[0197] W1+W2+W3+W4=1

[0198] The comprehensive health score can reflect the overall health condition of the transformer, providing a basis for subsequent maintenance or technical improvement decisions.

[0199] Step S5, according to the severity of different faults, generate hierarchical maintenance strategy, including emergency repair, planned maintenance and continuous monitoring. When the same device has the same defect for more than 3 years or the cumulative maintenance cost exceeds 50% of the device residual value, it belongs to the scope of technical improvement project. Further, based on the health status of the equipment, the cost of improvement and the constraints of power grid operation, a multi-objective optimization model is constructed, and the improved NSGA-III algorithm is used to solve the optimal technical improvement scheme, ensuring that high-risk and high-benefit projects are disposed of within the budget.

[0200] Maintenance and technical improvement decisions

[0201] According to the fault level and specific judgment conditions of the transformer, a hierarchical maintenance strategy is developed. The fault level is divided into three categories: emergency repair, planned maintenance and continuous monitoring. Emergency repair is suitable for equipment with poor health status or sudden serious faults, which needs to be shut down immediately and completed within 24 hours. Planned maintenance is suitable for equipment with less serious faults, such as CHI value between 0.4 and 0.7, and the same defect recurs 1-2 times, and the maintenance can be completed within 30 days. Continuous monitoring is suitable for equipment with light faults, CHI value greater than or equal to 0.7, and the equipment shows slight abnormalities. In this case, enhanced data collection and real-time tracking are adopted.

[0202]

[0203] The trigger condition of technical improvement project is based on the health status of the equipment and historical fault data. When the same fault of the equipment recurs more than 3 times a year, or the residual value of the equipment is lower than the set standard, the technical improvement evaluation will be started. The residual value of the equipment is calculated by the formula, considering the original value, design life and running time of the equipment. If the cumulative maintenance cost exceeds 50% of the residual value of the equipment, the technical improvement evaluation will also be triggered. Through these judgment conditions, timely technical improvement of the equipment is ensured, avoiding further fault risk and high maintenance cost.

[0204] Objective function

[0205] Objective 1: Maximize equipment health state improvement

[0206] Maximize S total =w1·S oil +w2·S temp +w3·Svibration +w4·S pd

[0207] Meaning Explanation:

[0208] S total : Comprehensive health score.

[0209] S oil , S temp , S vibration , S pd : Oil spectrum, temperature, vibration and partial discharge health score respectively.

[0210] w1, w2, w3, w4: Corresponding weight coefficients, satisfying w1+w2+w3+w4=1.

[0211] Objective 2: Minimize the cost of transformation

[0212]

[0213] C total : Total transformation cost.

[0214] C i : The transformation cost of the i-th project.

[0215] x i ∈{0,1}: Whether to choose this technical improvement project.

[0216] Objective 3: Minimize downtime

[0217]

[0218] T total : Total downtime.

[0219] T i : The estimated downtime of the first technical improvement project.

[0220] x i ∈{0,1}: Whether to choose this project.

[0221] (3) Constraints

[0222] Constraint 1: Budget constraint

[0223]

[0224] B max : Annual budget limit.

[0225] Constraint 2: Power outage window constraint

[0226]

[0227] Tmax : Maximum outage duration that can be tolerated.

[0228] Constraint 3: Health score threshold triggers tech improvement

[0229] S oil ≤ S oil，threshold , S vibration ≤ S vibration，threshold

[0230] If any dimension health score is below a set threshold, enter tech improvement evaluation process.

[0231] Constraint 4: Variable limits

[0232] x i ∈ {0, 1}, for i = 1, 2,..., m

[0233] (4) Solution approach

[0234] Non-dominated sorting

[0235] For an individual i, its dominance relation is defined as:

[0236] Individual i dominates individual j if:

[0237] f i，k ≤ f j,k , and

[0238] f i，k < f j，k ,

[0239] Crowding distance

[0240]

[0241] Selection operation

[0242] Use tournament selection to compare the dominance relation of two individuals i and j, select the individual with stronger dominance relation.

[0243] Crossover operation

[0244] Crossover operation generates new offspring individuals:

[0245] x child = C(x1, x2)

[0246] Mutation operation

[0247] Mutation operation increases population diversity by adjusting decision variable x i

[0248] ​x mutated = M(x i )

[0249] Pareto frontier selection

[0250] Select those individuals with greater diversity by crowding distance:

[0251] F1, F2, F3,...

[0252] f i，k : objective function value of individual i on target k.

[0253] x i : decision variable vector of individual i.

[0254] C(x1, x2): crossover operation, generating offspring individuals.

[0255] M(x i ): mutation operation, generating mutated individuals.

[0256] d i : crowding distance of individual i, used to measure the diversity of individuals.

[0257] F1, F2, F3,...: different levels of Pareto frontier, F1 is the optimal solution set, F2 is the suboptimal solution set, and so on.

[0258] The result of this solution scheme provides multiple Pareto optimal solutions for transformer technical improvement projects, reflecting the trade-off between different objectives. Through optimization algorithm, the best compromise scheme in terms of budget limitation, downtime and equipment health status is obtained. According to the solution result, the most suitable technical improvement scheme can be selected according to actual demand, so as to ensure the efficient operation and economy of the transformer. It provides strong support for technical improvement decision, which helps to improve the long-term stability and safety of the equipment.

[0259] From the technical knowledge, the application can be realized by other embodiments without departing from the spirit or essential characteristics thereof. Therefore, the above disclosed embodiments are only illustrative in all aspects, and are not the only ones. All changes within the scope of the application or within the scope equivalent to the application are included in the application.

Claims

1. A transformer health assessment and technical improvement decision optimization method, characterized in that, The method comprises the following steps, Step S1, collecting transformer surface image and infrared temperature field data by infrared dual-spectrum unmanned aerial vehicle, collecting vibration signal by vibration sensor array, and monitoring transformer oil chromatogram data online by gas sensor in oil to form a multi-modal original data set; Step S2, the image data collected in step S1 is subjected to de-fogging processing by a generative adversarial network, the vibration signal is subjected to wavelet transform denoising, and the oil chromatogram data is subjected to sliding average smoothing processing to form a pre-processed data set; Step S3, the pre-processed data set is input into a multi-modal deep learning model, the oil chromatogram data is processed by a time series analysis module to generate a gas trend prediction result, the de-fogging image is processed by an image segmentation module to generate a hot spot area positioning result, and the de-noised vibration signal is processed by a vibration frequency spectrum analysis module in combination with a partial discharge detection result to generate a mechanical fault judgment result; Step S4, according to the fault recognition result of step S3, the health scores of the oil chromatogram, temperature, vibration and partial discharge dimensions are calculated respectively, and a comprehensive health score is fused by a dynamic weight distribution strategy; Step S5, based on the comprehensive health score and historical fault records, a graded maintenance strategy is generated, when the technical improvement trigger condition is met, a multi-objective optimization model containing health improvement, cost control and power outage constraint is constructed, and an optimal technical improvement scheme sequence is solved.

2. The transformer health assessment and technical improvement decision optimization method of claim 1, wherein, The step S2 comprises, The infrared dual-spectrum unmanned aerial vehicle image data collected in step S1 is subjected to de-fogging processing by a generative adversarial network to generate a de-fogging image data set, and the generative adversarial network optimizes the image clarity through the adversarial training of the generator and the discriminator; The vibration sensor array signal collected in step S1 is subjected to wavelet transform decomposition, and after removing the high-frequency noise, it is reconstructed into a de-noised vibration signal set; The oil chromatogram data collected in step S1 is subjected to sliding average smoothing processing to generate a stable oil chromatogram data set; The de-fogging image data set, the de-noised vibration signal set and the stable oil chromatogram data set are combined into a pre-processed data set for input into the multi-modal deep learning model of step S3.

3. The transformer health assessment and technical improvement decision optimization method of claim 2, wherein, In the step S3, The time series analysis module uses a long short-term memory network LSTM to analyze the time series characteristics of the oil chromatogram data, when the target gas ratio exceeds the dynamically adjusted first threshold value and the monthly change amount of hydrogen concentration exceeds the dynamically adjusted second threshold value, a high temperature overheating fault judgment signal is triggered; The image segmentation module uses an improved U-Net model to locate the hot spot area in the infrared thermal image, and the U-Net model eliminates the environmental temperature interference through the temperature compensation module in the encoder; The vibration frequency spectrum analysis module uses a one-dimensional convolutional neural network 1D-CNN to extract the frequency spectrum characteristics of the vibration signal, when the energy of a specific frequency band exceeds the dynamically adjusted third threshold value and the Mel frequency distribution of the voiceprint signal is abnormal, an iron core loosening fault judgment result is generated.

4. The transformer health assessment and technical improvement decision optimization method of claim 3, wherein, In the step S4, The oil chromatogram health score is calculated based on the weighted calculation of the target gas ratio and the change amount of hydrogen concentration, and the weight is dynamically adjusted according to the historical fault data; The temperature health score is determined according to the difference between the maximum temperature difference of the hot spot area and the preset threshold value, and the preset threshold value is set in combination with the transformer model. The vibration health score is calculated based on the deviation of vibration energy from a normal threshold, which is dynamically updated by baseline data; The partial discharge health score is determined according to the exceeding of discharge intensity and pulse frequency, and the exceeding threshold is dynamically optimized based on an insulation aging model; The comprehensive health score is dynamically allocated to each dimension weight by entropy weight method and weighted average calculation, and when the comprehensive score is lower than the preset risk threshold, the high-risk equipment priority processing instruction is triggered.

5. The transformer health assessment and technical improvement decision optimization method of claim 1, wherein, In the step S5, The hierarchical maintenance strategy includes emergency repair, planned maintenance and continuous monitoring, and when the annual recurrence number of similar defects exceeds the dynamically adjusted fourth threshold or the cumulative maintenance cost exceeds the proportion threshold of the equipment residual value, technical improvement project evaluation is triggered; The multi-objective optimization model aims to maximize the health score improvement, minimize the transformation cost and power outage time, and generates a non-dominated solution set through an improved NSGA-III algorithm, which uses a dynamic hierarchical screening mechanism to divide the scheme levels.

6. The transformer health assessment and technical improvement decision optimization method of claim 5, wherein, The improved NSGA-III algorithm includes: Through non-dominated sorting, a candidate scheme set that meets the health score improvement and cost constraints is screened out; Based on the adaptive reference point strategy, the high-dimensional target space is uniformly partitioned to ensure the diversity and coverage of the solution set; A hybrid selection strategy is adopted to combine tournament selection and regional preference guidance to drive the population to evolve towards the region with high health score improvement rate and optimal cost benefit; Based on the dynamic crossover and mutation operators designed for decision variables, the global exploration and local development capabilities are balanced.

7. The transformer health assessment and technical improvement decision optimization method of claim 1, wherein, In the step S1, The infrared dual-spectrum unmanned aerial vehicle synchronously collects visible light and infrared images, and marks the key area coordinates through GPS positioning, which are matched with the transformer three-dimensional model to realize the visualization of the fault area; The vibration sensor array is deployed on the transformer base and shell, and real-time vibration signals are collected through multi-axis sensors, and the multi-axis signals are filtered through Kalman filtering to eliminate installation error interference; The oil gas sensor monitors the concentration of dissolved gas in oil and generates oil chromatogram time series data combined with time stamps.

8. The transformer health assessment and technical improvement decision optimization method of claim 1, wherein, In the step S3, The partial discharge detection collects discharge pulse signals through high-frequency sensors, and when the discharge intensity exceeds the dynamically adjusted fifth threshold or the pulse frequency exceeds the dynamically adjusted sixth threshold, it is determined as an electrical insulation fault; Cross-modal matching is realized by comparing the vibration frequency spectrum characteristics and the Mel frequency distribution of the voiceprint signal, which is extracted after pre-processing the voiceprint signal through a filter bank.

9. The transformer health assessment and technical improvement decision optimization method of claim 1, wherein, In the step S4, The dynamic weight allocation strategy is based on the influence of each dimension on the reliability of the equipment, and is dynamically optimized by entropy weight method combined with historical fault data; When the health score of any dimension is lower than the preset alarm threshold, the weight proportion of that dimension is forcibly increased, and real-time health state warning is triggered.

10. The transformer health assessment and technical improvement decision optimization method of claim 1, wherein, In the step S5, The technical improvement scheme optimization includes budget constraint, high-risk equipment coverage constraint and power outage time constraint, and the budget constraint is allocated through a dynamic programming algorithm; The non-dominated solution set is screened for the most diverse scheme through the crowding distance, and the crowding distance is determined based on the comprehensive calculation of the health score improvement rate and the cost benefit rate.

Images