Predictive maintenance method for transformer

Abstract

The invention provides a predictive maintenance method for a transformer, and belongs to the technical field of fault prediction based on machine learning. The design life of the transformer is divided into three stages of a running-in stage, a green and strong stage and a decline stage, an oil chromatography online monitoring system and a pickup are deployed in the running-in stage, gas concentration and audio data are collected, characteristic gas data are preprocessed and then input into a three-ratio model to monitor faults, and the gas data in the stage is used for training an oil chromatography model. Dividing positive and negative samples of audio data according to whether a fault occurs, and training a voiceprint recognition model; oil chromatography monitoring is carried out in the green and strong period, and fault monitoring is carried out only through a voiceprint recognition model. And recalibrating and starting an oil chromatography monitoring system in a decline period, and carrying out fault early warning by fusing an oil chromatography model and a voiceprint recognition model, so that missed judgment and misjudged judgment are reduced. According to the invention, the comprehensive cost performance of fault monitoring is improved through the staged and dynamically adaptive fault monitoring model, and the exclusive adaptation of the fault identification model and the target transformer is realized.

Description

Technical Field

[0001] This invention belongs to the field of fault prediction technology based on machine learning, and particularly relates to a predictive maintenance method for transformers. Background Technology

[0002] Predictive maintenance of transformers is a key aspect of power equipment operation and maintenance management. It refers to the process of collecting equipment operation data through multi-dimensional monitoring methods, combining data analysis models to predict potential equipment failures and assess health status, and formulating targeted maintenance strategies. The goal is to transform "fault repair" into "pre-emptive prediction and precise maintenance," ensuring the safe and stable operation of transformers throughout their entire life cycle and reducing the risk of unplanned downtime.

[0003] Full life cycle management is the core management concept that coordinates the operation and maintenance decisions of transformers throughout the entire process from commissioning to decommissioning. Its core is to dynamically adapt monitoring methods, analysis models and maintenance strategies according to the performance characteristics, fault patterns and operation and maintenance needs of equipment at different operating stages, so as to achieve the rational allocation of operation and maintenance resources, maximize the utilization of equipment life, improve the operation and maintenance efficiency of power equipment, control operation and maintenance costs, and ensure the reliable power supply of the power system, which directly affects the operation stability and operation and maintenance economic benefits of the power network.

[0004] In the current transformer operation and maintenance scenarios of the power industry, predictive maintenance based on full life cycle management still relies on traditional models, which mainly have the following problems: (1) The operation and maintenance strategy is fixed and lacks stage adaptability, resulting in low accuracy of fault prediction. Traditional transformer maintenance mostly adopts fixed strategies and does not design monitoring and maintenance schemes according to the different performance degradation patterns, fault types and occurrence probabilities of equipment during the break-in period, prime period and decline period. This leads to redundant monitoring methods and wasted operation and maintenance costs in some stages, and insufficient monitoring dimensions in some stages, making it impossible to capture potential fault signals in time. The failure rate and misjudgment rate of fault prediction are relatively high. (2) The monitoring methods are too limited and the data integration is insufficient, resulting in incomplete state perception. Existing predictive maintenance relies on a single monitoring technology to judge the equipment status. The data sources and characteristic parameters of different monitoring methods have not been systematically integrated, which makes it impossible to perceive the transformer's operating status from multiple dimensions and makes it difficult to accurately locate the root cause and development trend of the fault. (3) Inefficient resource allocation and poor maintenance economy result in low overall lifecycle operation and maintenance benefits. Under the traditional model, either excessive maintenance leads to ineffective consumption of human and material resources, increasing operation and maintenance costs; or insufficient maintenance leads to frequent equipment failures, which not only increases emergency repair costs but may also cause greater power supply losses due to unplanned shutdowns. Without dynamically allocating monitoring and maintenance resources according to the fault risk level of the equipment at different stages, it is impossible to achieve the optimal balance between safe and stable equipment operation and operation and maintenance cost control, resulting in low overall operation and maintenance economic benefits throughout the entire lifecycle.

[0005] In summary, existing predictive maintenance for transformers suffers from several key pain points, including rigid maintenance strategies and a lack of adaptability to different stages, limited monitoring methods and insufficient data integration, inefficient resource allocation, and poor maintenance economics. These shortcomings fail to meet the power industry's development needs for safe and reliable transformer operation, refined control of maintenance costs, and maximization of benefits throughout the entire transformer lifecycle. Summary of the Invention

[0006] To address the above problems, this invention proposes a predictive maintenance method for transformers, comprising the following steps: The service life of a transformer is divided into three stages: the break-in period, the prime period, and the decline period. During the break-in period, an online oil chromatography monitoring system was deployed to calculate the concentration data of H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 gases, and a microphone was used to collect transformer operating status data. The H2, CH4, C2H6, C2H4, and C2H2 gas concentration data were preprocessed and used as input to the transformer three-ratio model, which was used to monitor transformer faults. At the end of the break-in period, the concentration data of H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 gases during the break-in period were used as a training set to train a transformer oil chromatography model. The audio data collected by the microphone was divided into two categories: the audio data corresponding to the fault was used as a negative sample, and the others were used as positive samples to train a transformer fault soundprint recognition model. During its prime, the microphone collects audio data as input to the transformer fault acoustic fingerprint recognition model, and the transformer's operating status is monitored through the trained transformer fault acoustic fingerprint recognition model. During the decline period, the online oil chromatography monitoring system is activated, and transformer fault early warning is performed using the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model.

[0007] Preferably, based on the designed service life, 10% is used as the threshold for dividing the break-in period and the prime period, and 80% is used as the threshold for dividing the prime period and the decline period; the stage determination rules are clearly defined: if the actual running time after commissioning is less than 10% of the designed service life, it is determined to be the break-in period; if the actual running time is less than 80% of the designed service life, it is determined to be the prime period; if the actual running time is greater than or equal to 80% of the designed service life, it is determined to be the decline period.

[0008] Preferably, the online oil chromatography monitoring system is as follows: A special sampling probe is installed at the sampling point of the transformer oil circulation pipeline. The sampling point is selected in the oil circulation branch at the bottom of the transformer oil tank, and a small amount of transformer oil sample is continuously extracted. Using polymer membrane degassing technology, when the oil sample flows through the permeable membrane, the dissolved gas in the oil can pass through the membrane and enter the detection gas path, while the oil phase is returned to the transformer; The degassed gas enters a miniature online gas chromatograph, where it is separated by a miniature chromatographic column and detected by a miniature detector. The system has built-in calibration curves and concentration conversion programs, which directly calculate and display the real-time concentrations of dissolved gases H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 in the oil.

[0009] Preferably, the method of monitoring transformer faults using the transformer three-ratio model specifically includes: Define the calculation formula for the three-ratio model of the transformer: the first ratio is C2H2 / C2H4, the second ratio is CH4 / H2, and the third ratio is C2H4 / C2H6; Define the relationship between the first, second, and third ratios and the code; if the first, second, and third ratios are less than 0.1, the code is 0; if 0.1 is less than the first, second, and third ratios and less than 1, the code is 1; if the first, second, and third ratios are greater than or equal to 1, the code is 2; and combine the first, second, and third ratios into one code. The gas concentration is defined as an exclusion factor that is not included in the coding calculation; if the gas volume fraction in the denominator of the first, second, and third ratios is <5μL / L, the ratio is recorded as 0; if the molecular gas in the first, second, and third ratios is not detected, the ratio is recorded as 0. Define the relationship between codes and transformer faults; overheat fault codes are 000, 001, 001, 011, 002, 012, where 000 and 001 represent low-temperature overheating; 011 and 002 represent medium-temperature overheating; and 002 and 012 represent high-temperature overheating; discharge fault codes are 100, 101, 112, 122, 202, 212, 222, where 100 and 101 represent partial discharge; 112 and 122 represent spark discharge; and 202, 212, 222 represent arc discharge; overheating combined with discharge fault codes are 111, 121, 211; other code calculation results indicate no fault.

[0010] Preferably, the transformer oil chromatography model adopts the LightGBM model; the model sequentially outputs preliminary diagnostic results through feature weight matching and gradient boosting classification reasoning, including normal operation, overheating fault, discharge fault, mixed fault and corresponding diagnostic confidence, the confidence level represents the probability of the model's judgment of the diagnostic result, and the value range is [0,1].

[0011] Preferably, during the training of the transformer oil chromatography model, the LightGBM basic model framework is initialized, the gradient update of some basic hyperparameters of the model by default is turned off, and the trainability of parameters strongly correlated with oil chromatography fault diagnosis is retained; the model input feature dimension is set to 7, and the feature dimension represents the number of core features of the oil chromatography data processed by the model. A hierarchical feature weighting training strategy is adopted, and trainable feature weight vectors are added to the feature input layer of the LightGBM model. The weight vectors are divided into three categories according to the transformer fault type: overheating fault, discharge fault, and mixed fault, which correspond to the three types of oil chromatographic feature templates designed respectively. The initial value of the weight vector is set to be uniformly distributed, and the feature weight parameters are iteratively optimized by gradient boosting.

[0012] Preferably, the transformer fault voiceprint recognition model adopts a CNN+LSTM hybrid model, which sequentially extracts spatial features through the CNN layer, captures temporal features through the LSTM layer, fuses features, and performs classification reasoning to output recognition results, including normal operation, winding fault, core fault, cooling system fault, and corresponding recognition confidence. The confidence represents the probability of the model's judgment on the recognition result, and the value range is [0,1].

[0013] Preferably, the process of constructing the training dataset for the transformer fault acoustic signature recognition model includes: One microphone is deployed at each of the transformer's vibration and noise-generating locations, including the tank sidewall, core support, and winding lead terminals, to collect vibration and abnormal noise audio signals during transformer operation. Audio data length is standardized. Data shorter than 10 seconds is padded with zeros, and data longer than 10 seconds is truncated to the first 10 seconds. After standardization, each segment of audio is collected as a single audio data sample. Normalize individual audio data samples, and calculate the audio data amplitude A using a 0-1 distribution. Data processing results The calculation formula is as follows: ; In the formula This represents the minimum value in the audio data sample; This represents the maximum value in the audio data sample.

[0014] Preferably, during the degradation period, if the output results of both the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model are "normal operation", the combined output result of the two models is "normal operation", and the confidence output result = transformer oil chromatography model × 0.6 + transformer fault acoustic fingerprint recognition model × 0.4; If both the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model output "fault state", the combined output of the two models will also be "fault state"; the confidence level will be output according to specific calculation rules. If the transformer oil chromatography model outputs "normal operation" and the transformer fault acoustic fingerprint recognition model outputs "fault status", the combined output of the two models is "fault status", and the confidence output is = transformer fault acoustic fingerprint recognition model × 0.8; If the transformer oil chromatography model outputs "fault state" and the transformer fault acoustic fingerprint recognition model outputs "normal operation", the combined output of the two models is "fault state", and the confidence output is = transformer oil chromatography model × 0.8.

[0015] Preferably, the confidence level is output according to a specific calculation rule, specifically: If the transformer oil chromatography model results indicate an overheating fault, the confidence weight is 0.7. If the transformer oil chromatography model result indicates a discharge fault, the confidence weight is 0.3. If the transformer oil chromatography model results indicate a mixed fault, the confidence weight is 0.6. If the transformer fault acoustic fingerprint identification model results in a winding fault, the confidence weight is 0.8. If the transformer fault acoustic fingerprint identification model results in a core fault, the confidence weight is 0.7. If the transformer fault acoustic fingerprint identification model results in a cooling system fault, the confidence weight is 0.3. The result of the transformer oil chromatography model is multiplied by the confidence weight of the transformer oil chromatography model, and is represented by flag1; the result of the transformer fault acoustic fingerprint recognition model is multiplied by the confidence weight of the transformer fault acoustic fingerprint recognition model, and is represented by flag2. If flag1 ≥ flag2, the combined output of the two models is "fault state", and the confidence output is equal to flag1. If flag1 < flag2, the combined output of the two models is "fault state", and the confidence output is equal to flag2.

[0016] Compared with the prior art, the present invention has the following innovative features: (1) Divide the entire life cycle of the transformer into different operation and maintenance stages: Unlike the fixed mode of dividing the operation and maintenance stages of the transformer according to a fixed duration in the existing technology, it is divided into three stages: break-in period, prime period and decline period based on the proportion threshold of the transformer's design service life (10%, 80%), and transformer fault early warning is executed according to the characteristics of the stages. (2) Construct a phased and dynamically adaptable fault monitoring model application strategy: Based on the fault patterns of different life stages of transformers and the maintenance cost of the online oil chromatography monitoring system, the transformer three-ratio model is used to realize fault monitoring during the break-in period, the transformer fault acoustic fingerprint recognition model is used alone during the prime period, and the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model are used in synergistic monitoring during the decline period, so as to realize the on-demand activation and dynamic switching of the monitoring model. (3) Training the transformer oil chromatography model and fault acoustic fingerprint recognition model of the target transformer during the break-in period: Gas concentration data and audio data are collected simultaneously during the break-in period. The gas concentration data is used as the input for the transformer three-ratio model. The gas concentration data is used as a whole to train the transformer oil chromatography model, and the audio data is classified into positive and negative samples according to the fault type to specifically train the fault acoustic fingerprint recognition model of the target transformer.

[0017] The beneficial effects brought about by the innovation of this invention include: (1) Scientific and universal division of the operation and maintenance stages of the transformer throughout its entire life cycle: This avoids the problem of inaccurate stage operation and maintenance strategies caused by the mismatch between the fixed life cycle division and the actual design life of the equipment, and makes the division of the entire life cycle stages more in line with the actual operating rules of the equipment, with stronger universality and adaptability. (2) A phased and dynamically adaptable fault monitoring model improves the overall cost-effectiveness of fault monitoring: The online oil chromatography monitoring system requires calibration every six months, resulting in high maintenance costs. This method enables on-demand activation and dynamic switching of the monitoring model, significantly reducing the maintenance cost of the online oil chromatography monitoring system; (3) The integrated pre-design of "data acquisition + model training" during the equipment break-in period enables the fault identification model to be specifically adapted to the target transformer: the fault identification model of the target transformer is trained in a targeted manner, so that the sample data for model training comes from the equipment's own operation process, which solves the problem of poor adaptability between general models and specific equipment in the existing technology and avoids the disconnect between model training and equipment monitoring in the later stage. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0019] Figure 2 This is a schematic diagram showing the stages of a transformer's entire life cycle.

[0020] Figure 3 This is a graph showing the results from the gas data base H2.

[0021] Figure 4 This is a diagram of the LightGBM model architecture.

[0022] Figure 5 This is a diagram of the CNN+LSTM model architecture.

[0023] Figure 6 The figure shows the simulation results of the fault during the decay period. Detailed Implementation

[0024] The present invention proposes a predictive maintenance method for transformers, the overall technical route flowchart of which is shown below. Figure 1 As shown, the specific steps are as follows: S1 divides the design service life of a transformer into three stages: the break-in period is when the operating time after commissioning is less than 10% of the design service life; the prime period is when the operating time is 10% of the design service life but less than 80% of the design service life; and the decline period is when the operating time is 80% of the design service life.

[0025] S2. During the transformer break-in period, an online oil chromatography monitoring system (calculating H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 gas concentration data) and a microphone are deployed to collect transformer operating status data. The H2, CH4, C2H6, C2H4, and C2H2 gas concentration data, after preprocessing, are used as input to the transformer's three-ratio model, which monitors transformer faults. At the end of the break-in period, the H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 gas concentration data from the break-in period are used as a training set to train the transformer oil chromatography model. The collected audio data is divided into two categories: audio data corresponding to faults are used as negative samples, and others are used as positive samples, to train a transformer fault acoustic signature recognition model.

[0026] S3. During the prime of the transformer equipment, the transformer fault acoustic fingerprint recognition model is obtained through the second step of training to monitor the transformer's working status. The microphone collects audio data as the input of the transformer fault acoustic fingerprint recognition model. At this time, the H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 gas sensors deployed when the transformer was put into operation have reached their designed service life, so the transformer oil chromatography model is no longer used to monitor transformer faults.

[0027] S4. During the transformer equipment degradation period, the online oil chromatography monitoring system is recalibrated and enabled. At the same time, the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model are used to perform transformer fault early warning, reducing the probability of missed or false transformer faults.

[0028] The specific implementation process of the present invention will be described in detail below with reference to specific embodiments.

[0029] S1. Divide the design service life of the transformer into 3 stages. S1-1 Obtain basic transformer parameters and extract core data on design service life. Retrieve the rated design service life (in years) of the target transformer from the transformer equipment's factory technical documents and equipment ledger management system as the data benchmark for subsequent stage divisions.

[0030] S1-2 Based on the performance degradation law and fault occurrence characteristics of transformers throughout their entire life cycle, the proportional thresholds for stage division are set: using the designed service life as a benchmark, 10% is the threshold for dividing the break-in period and the prime period, and 80% is the threshold for dividing the prime period and the decline period; the stage determination rules are clarified: if the actual operating time after commissioning is less than 10% of the designed service life, it is determined to be the break-in period; if 10% of the designed service life ≤ the actual operating time < 80% of the designed service life, it is determined to be the prime period; if the actual operating time ≥ 80% of the designed service life, it is determined to be the decline period. Figure 2 As shown.

[0031] S2. Fault monitoring during the break-in period and the construction of two models An online oil chromatography monitoring system was deployed. The collected H2, CH4, C2H6, C2H4, and C2H2 gas concentration data, after data preprocessing, were used as input to the transformer three-ratio model, which was then used to monitor transformer faults. At the end of the break-in period, the H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 gas concentration data collected by the online oil chromatography monitoring system during the break-in period were used as the training set to train the transformer oil chromatography model. A microphone was deployed to collect audio data. At the end of the break-in period, the collected audio data was divided into two categories: audio data corresponding to faults were used as negative samples, and others as positive samples. At the end of the break-in period, a transformer fault acoustic signature recognition model was trained.

[0032] The specific steps for deploying the online oil chromatography monitoring system in S2-1 are as follows: (1) Install the special sampling probe at the sampling point of the transformer oil circulation pipeline. The sampling point is selected from the oil circulation branch at the bottom of the transformer oil tank to ensure that the special sampling probe is in full contact with the transformer oil. Continuously extract a small amount of transformer oil sample. The pipeline is specially treated to prevent adsorption and blockage, and impurities / water in the oil are filtered out. (2) Using polymer membrane degassing technology, when the oil sample flows through the permeable membrane, the dissolved gas in the oil can pass through the membrane and enter the detection gas path, and the oil phase is returned to the transformer, achieving lossless and continuous degassing without manual intervention; (3) The degassed gas enters the micro online gas chromatograph, where it is separated by a micro chromatographic column and detected by a micro detector (TCD / FID), and the whole process is automated; (4) The system has built-in calibration curves and concentration conversion programs to directly calculate and display the real-time concentrations of dissolved gases (H2, CH4, C2H6, C2H4, C2H2, CO, CO2) in the oil. The data can be uploaded to the background monitoring system through the communication interface.

[0033] S2-2 performs preprocessing on the collected gas concentration data to obtain the gas data base; the specific steps are as follows: (1) The sampling frequency is set to 1 hour / time, and the data format is set to timestamp + gas type + concentration value, where the concentration value unit is μL / L; (2) Imputing missing values ​​in the collected gas data. Taking H2 gas concentration data as an example, the number of missing H2 gas concentration data is counted on a daily basis, and the percentage of missing values ​​relative to the theoretical total data collected is calculated. If the percentage of missing values ​​exceeds 10% of the total collected data, the online oil chromatography monitoring system is maintained and the data collected that day is discarded. If the percentage of missing values ​​is less than 10% of the total collected data, imputation is performed for the missing values. H2 gas data at time If missing, then , express H2 gas data at specific times; (3) Normalize the gas concentration data after imputation of missing values. Taking H2 gas concentration data as an example, the data is calculated using a 0-1 distribution on a daily basis. H2 gas data processing results at time 10:00 The calculation formula is as follows: ; In the formula This indicates the minimum value of H2 gas data for that day; This indicates the maximum value of H2 gas data for that day; Similarly, we can obtain 、 、 、 、 、 . express The result of interpolating and normalizing the CH4 gas concentration data at any given time. express The result of interpolating and normalizing the C2H6 gas concentration data at any given time. express The result of interpolating and normalizing the C2H4 gas concentration data at any given time. express The result of interpolating and normalizing the C2H2 gas concentration data at any given time. express time CO The result of imputing missing values ​​and normalizing the gas concentration data. express The result of interpolating and normalizing the CO2 gas concentration data at any given time. Taking H2 gas data as an example, the gas data base data is as follows: Figure 3 As shown.

[0034] S2-3 deploys a microphone to collect audio data and performs data preprocessing on the collected audio data to obtain an audio data base; the specific steps are as follows: (1) At the locations where the transformer vibrates and produces noise: one microphone is deployed on each of the tank sidewall, core support, and winding lead terminals to collect audio signals such as vibration and abnormal noise during transformer operation; (2) The audio data acquisition frequency is set to once every 5 minutes, the sampling rate is set to 20kHz, the single audio segment acquisition duration is 10s, and the audio format is saved as WAV; (3) Audio data length standardization processing: padding with zeros for data shorter than 10 seconds, and truncating the first 10 seconds of data for data longer than 10 seconds. After standardization, each segment of audio data is collected as a single audio data sample. (4) Normalize the individual audio data sample and perform 0-1 distribution calculation on the audio data amplitude A to obtain Data processing results The calculation formula is as follows: ; In the formula This represents the minimum value in the audio data sample; This represents the maximum value in the audio data sample.

[0035] S2-4 During the break-in period, transformer faults are monitored using a three-ratio model; the specific steps are as follows: (1) Define the calculation formula for the three-ratio model of the transformer. The first ratio is C2H2 / C2H4, the second ratio is CH4 / H2, and the third ratio is C2H4 / C2H6; (2) Define the relationship between the first, second, and third ratios and the code. If the first, second, and third ratios are less than 0.1, the code is 0; if 0.1 is less than the first, second, and third ratios and less than 1, the code is 1; if the first, second, and third ratios are greater than or equal to 1, the code is 2. The first, second, and third ratios are then combined into a single code. (3) Define the exclusion factors that gas concentration does not participate in the coding calculation. If the gas volume fraction in the denominator of the first, second, and third ratios is <5μL / L, the ratio is recorded as 0; if the molecular gas in the first, second, and third ratios is not detected, the ratio is recorded as 0. (4) Define the relationship between codes and transformer faults. Overheat fault codes are 000, 001, 001, 011, 002, and 012, where 000 and 001 represent low-temperature overheating; 011 and 002 represent medium-temperature overheating; and 002 and 012 represent high-temperature overheating. Discharge fault codes are 100, 101, 112, 122, 202, 212, and 222, where 100 and 101 represent partial discharge; 112 and 122 represent spark discharge; and 202, 212, and 222 represent arc discharge. Overheat combined with discharge fault codes are 111, 121, and 211. Other code calculation results indicate no fault.

[0036] At the end of the S2-5 break-in period, a transformer oil chromatographic model was trained based on the gas data base of S2-1; the specific steps are as follows: (1) Model selection: LightGBM model. Initialize the LightGBM basic model framework, disable the default gradient updates of some basic hyperparameters, and retain the trainability of parameters strongly related to oil chromatography fault diagnosis (number of leaves, learning rate); set the model input feature dimension to 7 (H2, CH4, C2H6, C2H4, C2H2, CO, CO2 in the S2-1 gas data base). The feature dimension represents the number of core features processed by the model in the oil chromatography data. The model architecture diagram is as follows. Figure 4 As shown; (2) A hierarchical feature weighting training strategy is adopted. Trainable feature weight vectors (with the same dimension as the input feature dimension) are added to the feature input layer of the LightGBM model. The weight vectors are divided into three categories according to the transformer fault type: overheating fault, discharge fault, and mixed fault, which correspond to the three types of oil chromatography feature templates designed respectively. The initial value of the weight vector is set to be uniformly distributed (0.1~0.9). Gradient boosting is used to iteratively optimize the feature weight parameters to realize the adaptation of the model to the transformer oil chromatography fault diagnosis task. (3) Split the datasets of H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 in the S2-1 gas data base into training and validation sets in an 8:2 ratio.

[0037] Using the training set as input, the feature weight vector is concatenated with the oil chromatography sample features and then fed into the model to construct a multi-class training objective function (overheating fault / discharge fault uses cross-entropy loss; mixed fault uses mean squared error loss, with weight ratios of 0.4, 0.4, and 0.2, respectively); the native gradient boosting optimizer of LightGBM is used, with a learning rate of 0.05, a batch size of 64, and 50 training epochs.

[0038] After each training round, the model performance (fault classification accuracy, fault type F1 score, and feature fit) is evaluated using a validation set. Training is stopped and the optimal model weights are saved if the transformer overheating fault diagnosis accuracy (≥98%), discharge fault diagnosis F1 score (≥97%), and insulation aging trend fit (≥96%) are simultaneously satisfied.

[0039] The F1 score is the harmonic mean of precision and recall, expressed as follows: ; At the end of the S2-6 break-in period, a transformer fault acoustic signature recognition model was trained based on the audio data base of S2-2; the specific steps are as follows: (1) Model selection: CNN+LSTM hybrid model. Initialize the CNN+LSTM hybrid model framework, turn off the partial pre-trained weight gradient update of the CNN basic convolutional layer and the LSTM basic recurrent layer in the model, and retain the trainability of the core parameters of voiceprint feature extraction (CNN convolutional kernel size, LSTM hidden layer dimension); set the model input feature dimension to (128, 100), and the feature dimension represents the core spatiotemporal feature dimension of the model in processing voiceprint audio.

[0040] Model structure as follows Figure 5 As shown in the figure, the parameters are explained as follows: If the current time is time t (time t represents...) time), This indicates the hidden state at the previous moment; This indicates the hidden state at this moment; This indicates the hidden state to predict the next time step.

[0041] (2) A spatiotemporal feature weighted fusion strategy is adopted. A trainable spatiotemporal weight vector is added to the feature fusion layer of the CNN+LSTM model. The weight vector is divided into three categories according to the transformer fault type: winding fault, core fault and cooling system fault, which correspond to the three types of designed voiceprint spatiotemporal feature templates. The weight vector length is set to 64, and it is initialized with uniform distribution (0.05~0.95). The weight vector parameters are iteratively optimized through backpropagation to realize the adaptation of the model to the transformer fault voiceprint recognition task.

[0042] (3) Split the preprocessed voiceprint audio dataset into a training set and a validation set in a 7:3 ratio. Using the training set as input, the spatiotemporal weight vector and the spatiotemporal features of the voiceprint are concatenated and fed into the model to construct a multi-classification training objective function (cross-entropy loss is used for winding faults / core faults, and focus loss is used for cooling system faults, with weight ratios of 0.4, 0.3, and 0.3, respectively); the Adam optimizer is used, with a learning rate of 1e-4, a batch size of 32, and 20 training epochs.

[0043] After each training round, the model performance (fault identification accuracy, fault type F1 score, and voiceprint feature matching degree) is evaluated using a validation set. Training is stopped and the optimal model weights are saved if the transformer winding fault identification accuracy (≥96%), iron core fault identification F1 score (≥95%), and cooling system fault feature matching degree (≥94%) are simultaneously satisfied.

[0044] The F1 score is the harmonic mean of precision and recall, expressed as follows: ;

[0045] S3. Fault monitoring using a transformer fault soundprint recognition model for young and middle-aged adults. S3-1 saves the transformer fault acoustic signature recognition model trained by S2-6 to the transformer full life cycle operation and maintenance management platform, associates it with the basic operation and maintenance file of the target transformer, completes the communication adaptation between the model and the data acquisition terminal and the microphone, and ensures that the model can be directly called to carry out fault monitoring during the young and middle-aged and declining periods.

[0046] S3-2 data acquisition is performed according to steps (2)-(4) in S2-3 to obtain the voiceprint data acquisition and processing results. The processed voiceprint data is used as the input of the transformer fault voiceprint recognition model obtained in S2-6.

[0047] The S3-3 transformer fault acoustic signature recognition model sequentially performs spatial feature extraction at the CNN layer, temporal feature capture at the LSTM layer, feature fusion, and classification inference. It outputs the recognition result (normal operation / winding fault / core fault / cooling system fault) and the corresponding recognition confidence level. The confidence level represents the model's probability of judging the recognition result, with a value range of [0,1]. If the output fault recognition result has a confidence level ≥ 0.9, the system determines it as a valid result and executes fault information reporting.

[0048] S4. Use two models in tandem for fault early warning during the decline period. During the degradation period, the transformer oil chromatography model trained by S2-5 and the transformer fault acoustic fingerprint recognition model trained by S2-6 are used together to perform transformer fault early warning, reducing the probability of missed or false detection of transformer faults.

[0049] S4-1 uses the transformer oil chromatographic model trained in S2-5; the specific steps are as follows: (1) Maintain and calibrate the online oil chromatography monitoring system deployed in S2-1; (2) Save the transformer oil chromatography model obtained by S2-5 training to the transformer full life cycle operation and maintenance management platform, associate it with the operation and maintenance basic file of the target transformer, complete the communication adaptation between the model and the data acquisition terminal and gas sensor, and ensure that the model can be called during the decay period; (3) Data Acquisition: Following step S2-2, the gas concentration data acquisition and processing results are obtained. The processed gas concentration data is then used as input to the transformer oil chromatography model obtained in S2-5. (4) The model sequentially completes feature weight matching and gradient boosting classification reasoning, and outputs preliminary diagnostic results (normal operation / overheating fault / discharge fault / mixed fault) and corresponding diagnostic confidence. The confidence represents the probability of the model's judgment of the diagnostic results, and the value range is [0,1].

[0050] S4-2 uses the transformer fault acoustic signature recognition model trained in S2-6; the specific steps are as follows: (1) Data acquisition is performed according to steps (2)-(4) in S2-3 to obtain the voiceprint data acquisition and processing results. The processed voiceprint data is used as the input of the transformer fault voiceprint recognition model obtained in S2-6; (2) The transformer fault soundprint recognition model outputs the recognition results (normal operation / winding fault / iron core fault / cooling system fault) and the corresponding recognition confidence. The confidence represents the model's judgment probability of the recognition result, and the value range is [0,1].

[0051] S4-3 establishes a dual-model collaborative working mechanism; the specific steps are as follows: (1) Decision input synchronization. Since the oil chromatography model acquisition interval is 1 hour and the transformer fault acoustic fingerprint recognition model acquisition interval is 5 minutes, it is necessary to align the transformer fault acoustic fingerprint recognition model results (normal operation / winding fault / core fault / cooling system fault) with the most recent transformer oil chromatography model results (normal operation / overheating fault / discharge fault / mixed fault) on the time axis; (2) Decision output consistency rule. If the output results of both the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model are "normal operation", the combined output result of the two models is "normal operation", and the confidence output result = transformer oil chromatography model × 0.6 + transformer fault acoustic fingerprint recognition model × 0.4.

[0052] If the output results of both the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model are “fault state”, the output result of the dual model collaboration is “fault state”, and the confidence calculation rules are explained in step (3).

[0053] If the transformer oil chromatography model outputs "normal operation" and the transformer fault acoustic fingerprint recognition model outputs "fault status", the combined output of the two models is "fault status", and the confidence output is = transformer fault acoustic fingerprint recognition model × 0.8.

[0054] If the transformer oil chromatography model outputs "fault state" and the transformer fault acoustic fingerprint recognition model outputs "normal operation", the combined output of the two models is "fault state", and the confidence output is = transformer oil chromatography model × 0.8.

[0055] (3) Define the weight relationship when locating faults. If the output results of the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model are both "fault state", the weight of the fault confidence score is calculated to satisfy the following relationship. If the transformer oil chromatography model results indicate an overheating fault, the confidence weight is 0.7. If the transformer oil chromatography model result indicates a discharge fault, the confidence weight is 0.3. If the transformer oil chromatography model results indicate a mixed fault, the confidence weight is 0.6. If the transformer fault acoustic fingerprint identification model results in a winding fault, the confidence weight is 0.8. If the transformer fault acoustic fingerprint identification model results in a core fault, the confidence weight is 0.7. If the transformer fault acoustic fingerprint identification model results in a cooling system fault, the confidence weight is 0.3. The confidence weight of the transformer oil chromatography model is calculated by multiplying the results by the model's confidence level, denoted as flag1. The confidence weight of the transformer fault acoustic fingerprint recognition model is calculated by multiplying the results by the model's confidence level, denoted as flag2.

[0056] If flag1 ≥ flag2, the combined output of the two models is "fault state", and the confidence output is equal to flag1. If flag1 < flag2, the combined output of the two models is "fault state", and the confidence output is equal to flag2. The results of the laboratory simulation of the recession period, based on real data and model output, are as follows: Figure 6 As shown in the figure, the experimental results indicate that the fault confidence level output by the model during the decay period fits the fault data well, and the model can effectively realize the function of transformer fault early warning.

[0057] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0058] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A predictive maintenance method for transformers, characterized in that, Includes the following steps: The service life of a transformer is divided into three stages: the break-in period, the prime period, and the decline period. During the break-in period, an online oil chromatography monitoring system was deployed to calculate the concentration data of H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 gases, and a microphone was used to collect transformer operating status data. The H2, CH4, C2H6, C2H4, and C2H2 gas concentration data were preprocessed and used as input to the transformer three-ratio model, which was used to monitor transformer faults. At the end of the break-in period, the concentration data of H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 gases during the break-in period were used as a training set to train a transformer oil chromatography model. The audio data collected by the microphone was divided into two categories: the audio data corresponding to the fault was used as a negative sample, and the others were used as positive samples to train a transformer fault soundprint recognition model. During its prime, the microphone collects audio data as input to the transformer fault acoustic fingerprint recognition model, and the transformer's operating status is monitored through the trained transformer fault acoustic fingerprint recognition model. During the decline period, the online oil chromatography monitoring system is activated, and transformer fault early warning is performed using the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model.

2. The predictive maintenance method for transformers as described in claim 1, characterized in that: Based on the designed service life, 10% is the threshold for dividing the break-in period and the prime period, and 80% is the threshold for dividing the prime period and the decline period; the stage determination rule is clearly defined: if the actual running time after commissioning is less than 10% of the designed service life, it is determined to be the break-in period; If the actual operating time is less than or equal to 10% of the designed service life, it is considered to be in its prime; if the actual operating time is greater than or equal to 80% of the designed service life, it is considered to be in its decline stage.

3. The predictive maintenance method for transformers as described in claim 1, characterized in that: The oil chromatography online monitoring system is as follows: A special sampling probe is installed at the sampling point of the transformer oil circulation pipeline. The sampling point is selected in the oil circulation branch at the bottom of the transformer oil tank, and a small amount of transformer oil sample is continuously extracted. Using polymer membrane degassing technology, when the oil sample flows through the permeable membrane, the dissolved gas in the oil can pass through the membrane and enter the detection gas path, while the oil phase is returned to the transformer; The degassed gas enters a miniature online gas chromatograph, where it is separated by a miniature chromatographic column and detected by a miniature detector. The system has built-in calibration curves and concentration conversion programs, which directly calculate and display the real-time concentrations of dissolved gases H2, CH4, C2H6, C2H4, C2H2, CO, and CO2 in the oil.

4. The predictive maintenance method for transformers as described in claim 1, characterized in that: The method of monitoring transformer faults using a three-ratio model specifically includes: Define the calculation formula for the three-ratio model of the transformer: the first ratio is C2H2 / C2H4, the second ratio is CH4 / H2, and the third ratio is C2H4 / C2H6; Define the relationship between the first, second, and third ratios and the code; if the first, second, and third ratios are less than 0.1, the code is 0; if 0.1 is less than the first, second, and third ratios and less than 1, the code is 1; if the first, second, and third ratios are greater than or equal to 1, the code is 2; and combine the first, second, and third ratios into one code. The gas concentration is defined as an exclusion factor that is not included in the coding calculation; if the gas volume fraction in the denominator of the first, second, and third ratios is <5μL / L, the ratio is recorded as 0; if the molecular gas in the first, second, and third ratios is not detected, the ratio is recorded as 0. Define the relationship between codes and transformer faults; overheat fault codes are 000, 001, 001, 011, 002, 012, where 000 and 001 represent low-temperature overheating; 011 and 002 represent medium-temperature overheating; and 002 and 012 represent high-temperature overheating; discharge fault codes are 100, 101, 112, 122, 202, 212, 222, where 100 and 101 represent partial discharge; 112 and 122 represent spark discharge; and 202, 212, 222 represent arc discharge; overheating combined with discharge fault codes are 111, 121, 211; other code calculation results indicate no fault.

5. The predictive maintenance method for transformers as described in claim 1, characterized in that: The transformer oil chromatography model adopts the LightGBM model. The model outputs preliminary diagnostic results through feature weight matching and gradient boosting classification inference, including normal operation, overheating fault, discharge fault, mixed fault and corresponding diagnostic confidence. The confidence represents the probability of the model's judgment of the diagnostic result, and the value range is [0,1].

6. The predictive maintenance method for transformers as described in claim 5, characterized in that: During the training of the transformer oil chromatography model, the LightGBM basic model framework is initialized, the gradient update of some basic hyperparameters of the model by default is turned off, and the trainability of parameters strongly related to oil chromatography fault diagnosis is retained; the model input feature dimension is set to 7, and the feature dimension represents the number of core features of the oil chromatography data processed by the model. A hierarchical feature weighting training strategy is adopted, and trainable feature weight vectors are added to the feature input layer of the LightGBM model. The weight vectors are divided into three categories according to the transformer fault type: overheating fault, discharge fault, and mixed fault, which correspond to the three types of oil chromatographic feature templates designed respectively. The initial values ​​of the weight vector are set to be uniformly distributed, and the feature weight parameters are iteratively optimized using gradient boosting.

7. The predictive maintenance method for transformers as described in claim 1, characterized in that: The transformer fault voiceprint recognition model adopts a CNN+LSTM hybrid model, which sequentially extracts spatial features through the CNN layer, captures temporal features through the LSTM layer, fuses features, and performs classification reasoning to output recognition results, including normal operation, winding fault, core fault, cooling system fault, and corresponding recognition confidence. The confidence represents the probability of the model's judgment of the recognition result, and the value range is [0,1].

8. The predictive maintenance method for transformers as described in claim 1, characterized in that: The process of constructing the training dataset for the transformer fault acoustic signature recognition model includes: One microphone is deployed at each of the transformer's vibration and noise-generating locations, including the tank sidewall, core support, and winding lead terminals, to collect vibration and abnormal noise audio signals during transformer operation. Audio data length is standardized. Data shorter than 10 seconds is padded with zeros, and data longer than 10 seconds is truncated to the first 10 seconds. After standardization, each segment of audio is collected as a single audio data sample. Normalize individual audio data samples, and calculate the audio data amplitude A using a 0-1 distribution. Data processing results The calculation formula is as follows: ； In the formula This represents the minimum value in the audio data sample; This represents the maximum value in the audio data sample.

9. The predictive maintenance method for transformers as described in claim 1, characterized in that: During the degradation period, if the output results of both the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model are "normal operation", the combined output result of the two models is "normal operation", and the confidence output result = transformer oil chromatography model × 0.6 + transformer fault acoustic fingerprint recognition model × 0.4; If both the transformer oil chromatography model and the transformer fault acoustic fingerprint recognition model output "fault state", the combined output of the two models will also be "fault state"; the confidence level will be output according to specific calculation rules. If the transformer oil chromatography model outputs "normal operation" and the transformer fault acoustic fingerprint recognition model outputs "fault status", the combined output of the two models is "fault status", and the confidence output is = transformer fault acoustic fingerprint recognition model × 0.8; If the transformer oil chromatography model outputs "fault state" and the transformer fault acoustic fingerprint recognition model outputs "normal operation", the combined output of the two models is "fault state", and the confidence output result is = transformer oil chromatography model × 0.

8.

10. A predictive maintenance method for transformers as described in claim 9, characterized in that: The confidence level is output according to a specific calculation rule, specifically: If the transformer oil chromatography model results indicate an overheating fault, the confidence weight is 0.

7. If the transformer oil chromatography model result indicates a discharge fault, the confidence weight is 0.

3. If the transformer oil chromatography model results indicate a mixed fault, the confidence weight is 0.

6. If the transformer fault acoustic fingerprint identification model results in a winding fault, the confidence weight is 0.

8. If the transformer fault acoustic fingerprint identification model results in a core fault, the confidence weight is 0.

7. If the transformer fault acoustic fingerprint identification model results in a cooling system fault, the confidence weight is 0.

3. Calculate the transformer oil chromatography model result × the confidence weight of the transformer oil chromatography model, denoted by flag1; The result of the transformer fault acoustic fingerprint recognition model is multiplied by the confidence weight of the transformer fault acoustic fingerprint recognition model, and denoted by flag2. If flag1 ≥ flag2, the combined output of the two models is "fault state", and the confidence output is equal to flag1. If flag1 < flag2, the combined output of the two models is "fault state", and the confidence output is equal to flag2.

Images