Monitoring and diagnosis method, device and equipment of power transformer and medium

Abstract

The invention relates to a monitoring and diagnosing method, device and equipment of a power transformer and a medium. The method comprises the following steps: acquiring multi-dimensional monitoring data acquired by a sensor array in a full state of a transformer, and denoising to obtain a clean data sequence; inputting the cleaning data into a long short-term memory network to extract oil deposition time sequence dynamic characteristics, and combining a trend quantification model to study and judge a deposition accumulation trend and generate a trend curve; if the accumulation amount or the rate in the curve exceeds a preset threshold value calibrated based on historical fault data and industry standards, multi-source deposition risk key influence factors are collected, and after preprocessing, an improved weighted evidence theory is adopted for fusion to obtain deposition risk dominant factors; and inputting the factor into a convolutional neural network to locate a potential fault type, generating an adaptive diagnosis report, and if a high-risk fault exists, triggering and outputting a real-time intervention instruction. According to the method, early recognition, accurate diagnosis and timely intervention of oil deposition associated faults are realized, and the operation safety and stability of a power system are improved.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and in particular relates to a monitoring and diagnosis method, device, equipment and medium for power transformers. Background Technology

[0002] As a core hub in the power generation, transmission, and distribution chain, the operating status of power transformers directly affects the safety, stability, and continuity of power supply. With the development of power systems towards higher voltage, larger capacity, and greater intelligence, transformer operating conditions are becoming increasingly complex. Internal insulation aging, core wear, and partial discharge are easily affected by a combination of factors, including oil accumulation, temperature field distortion, and uneven electric field distribution. Failure to identify these issues promptly and accurately can lead to equipment shutdowns or even large-scale power outages. Existing monitoring and diagnostic technologies largely rely on manual inspections, offline testing, or online monitoring of single parameters. These technologies suffer from problems such as significant data noise interference, insufficient fusion of multi-source heterogeneous information, weak fault trend prediction capabilities, and insufficient diagnostic accuracy. They struggle to provide early warning and precise location of transformer oil accumulation-related faults, failing to meet the intelligent and refined health management requirements of power systems throughout the entire equipment lifecycle. Summary of the Invention

[0003] Therefore, it is necessary to provide a monitoring and diagnostic method, device, equipment, and medium for power transformers that can achieve early identification, accurate diagnosis, and timely intervention of transformer oil sludge-related faults, thereby improving the safety and stability of power system operation, in response to the above-mentioned technical problems.

[0004] Firstly, this application provides a monitoring and diagnostic method for power transformers, comprising:

[0005] Multidimensional monitoring data collected by sensor arrays under all operating and shutdown conditions of power transformers are acquired. Adaptive multi-scale filtering is used to process the multidimensional monitoring data to obtain a clean data sequence.

[0006] Based on clean data sequences, the temporal dynamic features of oil accumulation are extracted through long short-term memory networks, and the trend of oil accumulation is judged by combining trend quantification models.

[0007] If the accumulation characteristic value in the oil siltation accumulation trend curve exceeds the preset threshold, then multiple key influencing factors related to siltation risk are collected, and the key influencing factors are integrated using an improved weighted evidence theory fusion algorithm to obtain the dominant factor of siltation risk.

[0008] By inputting the dominant factor of siltation risk into a convolutional neural network to locate potential fault types, diagnostic reports and real-time intervention instructions adapted to the fault scenarios are generated based on the potential fault types.

[0009] In one embodiment, the clean data sequence is input into a long short-term memory network to extract the temporal dynamic features of oil sludge accumulation, and combined with a trend quantification model to obtain an oil sludge accumulation trend curve, including:

[0010] The clean data sequence is divided into preset time windows and then input into a long short-term memory network to extract the temporal dynamic features of oil accumulation.

[0011] Based on the time-series dynamic characteristics of oil sludge, the accumulation rate index is calculated using a preset rate calculation formula; the accumulation rate index quantitatively reflects the rate of change of particle concentration in the oil over time.

[0012] The accumulation rate index is input into the trend quantification model. The model performs time-series fitting and trend extrapolation on the index to obtain the oil accumulation trend curve. The trend quantification model is constructed by a time-series attention mechanism.

[0013] In one embodiment, the accumulation rate metric is calculated using the following formula:

[0014]

[0015] in, express The rate of oil accumulation at any given time. Indicates the duration of the preset time window. This represents the working condition correction factor. , Indicates the real-time load of the transformer. Indicates the rated load. Indicates the LSTM extraction of the first Time-series dynamic characteristics of oil accumulation The corresponding attention weights , They represent Time and Oil particle concentration at any given time Represents the characteristic attenuation coefficient, with values ​​ranging from 1 to 2. Calibration based on historical data fitting. Indicates the LSTM output of the first... Dimensional time-series dynamic characteristics.

[0016] In one embodiment, if the accumulation feature value in the oil sludge accumulation trend curve exceeds a preset threshold, then multiple key influencing factors related to sludge accumulation risk are collected, and an improved weighted evidence theory fusion algorithm is used to integrate the key influencing factors to obtain the dominant factor of sludge accumulation risk, including:

[0017] If the amount of oil accumulation or the real-time accumulation rate represented by the oil accumulation trend curve exceeds the preset threshold, multi-source heterogeneous information is collected. The multi-source heterogeneous information includes oil physicochemical parameters, equipment operating load, temperature field distribution, electric field strength, and ambient temperature and humidity data. The preset threshold is based on transformer historical fault data and insulation withstand standard calibration.

[0018] After standardizing and preprocessing multi-source heterogeneous information and removing outliers, a fusion preparation dataset containing feature identifiers, collection timestamps, and parameter values ​​is constructed.

[0019] An improved weighted evidence theory fusion algorithm is adopted to use temperature field distribution data and electric field intensity data in the fusion preparation dataset as the core evidence source. Combined with the credibility weight of each parameter, a multi-dimensional comprehensive state vector is obtained.

[0020] Based on the multi-dimensional comprehensive state vector, the information gain of each feature component is calculated using the entropy weight method to determine the weight distribution of key influencing factors.

[0021] The evidence conflict coefficient is dynamically adjusted based on the weight distribution of key influencing factors to obtain an optimized set of evidence.

[0022] Based on the DS evidence synthesis rules, the fusion confidence of each state category is recalculated using an optimized evidence body set to generate a refined comprehensive state vector; the state categories include normal, mild silo fault, and severe silo fault.

[0023] If the value of the feature component of the refined integrated state vector exceeds the secondary threshold, the variance contribution rate method is used to extract the dominant feature component in the refined integrated state vector to obtain the dominant factor of siltation risk; the secondary threshold is set based on the rated operating parameters of the equipment and the insulation safety threshold.

[0024] In one embodiment, the fusion confidence of each state category is calculated using the following formula:

[0025]

[0026] in, Indicates the first The fusion confidence of each state category, Indicates the preset state category. This indicates the number of evidence sources in the optimized evidence set. Indicates the first The weights of key influencing factors for each source of evidence are calculated using the entropy weight method. Indicates the first Each optimized evidence pair Basic probability assignment of class states This represents the dynamically adjusted coefficient of evidence, with a range of values. , , Indicates the first The initial conflict coefficients of each evidence source are obtained from data consistency checks. Indicates the first Conflict probability allocation of evidence sources This represents the empty set.

[0027] In one embodiment, the siltation risk-dominant factor is input into a convolutional neural network to locate potential fault types. Based on the potential fault types, a diagnostic report and real-time intervention instructions adapted to the fault scenario are generated, including:

[0028] By inputting the dominant factor of siltation risk into a convolutional neural network for preliminary feature extraction, a risk feature vector focusing on the risk-causing characteristics of siltation is obtained.

[0029] A deep learning algorithm based on an attention mechanism is used to classify multi-category faults on the risk feature vector to determine the potential fault types.

[0030] Based on the potential fault type, the system calls the preset fault diagnosis rule library and historical fault case library to generate fault diagnosis report data containing fault category, risk level and scope of impact.

[0031] If the fault diagnosis report data contains a fault category that meets the preset high-risk level, then a real-time intervention instruction adapted to the fault scenario will be generated.

[0032] Secondly, this application also provides a monitoring and diagnostic device for a power transformer, the device comprising:

[0033] The data acquisition module is used to acquire multi-dimensional monitoring data collected by the sensor array under the full operation and shutdown conditions of the power transformer. Adaptive multi-scale filtering is used to process the multi-dimensional monitoring data to obtain a clean data sequence.

[0034] The trend analysis module is used to extract the time-series dynamic features of oil accumulation based on clean data sequences through a long short-term memory network, and combine the trend quantification model to analyze the oil accumulation trend.

[0035] The risk assessment module is used to collect multiple key influencing factors related to siltation risk if the accumulation characteristic value in the oil siltation accumulation trend curve exceeds a preset threshold. The module then integrates these key influencing factors using an improved weighted evidence theory fusion algorithm to obtain the dominant factor of siltation risk.

[0036] The fault diagnosis module is used to input the dominant factors of siltation risk into a convolutional neural network to locate potential fault types, and generate diagnostic reports and real-time intervention instructions adapted to the fault scenarios based on the potential fault types.

[0037] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0038] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method.

[0039] The aforementioned monitoring and diagnostic method, device, computer equipment, and storage medium for power transformers acquire multi-dimensional monitoring data collected by a sensor array under all operating and shutdown conditions of the power transformer. This multi-dimensional monitoring data includes information such as oil physicochemical parameters, equipment operating load, temperature field distribution, electric field strength, and ambient temperature and humidity. Adaptive multi-scale filtering is used to suppress noise in the multi-dimensional monitoring data, outputting a high-fidelity clean data sequence. Based on this clean data sequence, it is divided into preset time windows and input into a long short-term memory network to extract the temporal dynamic features of oil accumulation. These temporal dynamic features are then input into a trend quantification model. Through temporal fitting and trend extrapolation, the oil accumulation trend is determined, and an accumulation trend curve is generated. If the accumulation amount or rate in the accumulation trend curve exceeds a preset threshold (based on historical transformer fault data and industry insulation withstand standards), multiple key influencing factors related to accumulation risk are collected. These factors undergo standardized preprocessing and outlier removal, and are then integrated using an improved weighted evidence theory fusion algorithm to obtain the dominant factor for accumulation risk. The method inputs the dominant factor of oil sludge risk into a convolutional neural network, and locates potential fault types through feature extraction and fault classification. Combining the risk level and impact range of the potential fault types, a diagnostic report adapted to the fault scenario is generated. If a high-risk fault category exists in the diagnostic report, a real-time intervention command generation module is triggered, outputting the corresponding real-time intervention command. This method achieves noise reduction and purification of multi-dimensional monitoring data under all conditions through adaptive multi-scale filtering, ensuring data reliability. It improves the effectiveness of oil sludge time-series feature extraction and the accuracy of accumulation trend prediction by combining a long short-term memory network with a trend quantification model. An improved weighted evidence theory fusion algorithm integrates multiple key influencing factors, solving the problem of insufficient fusion of multi-source heterogeneous information and improving the accuracy of the dominant factor of oil sludge risk. The convolutional neural network achieves precise location of potential faults associated with oil sludge, and diagnostic reports and real-time intervention commands are generated based on the fault type. This method overcomes the limitations of existing technologies such as single-parameter monitoring, insufficient trend prediction, and low diagnostic accuracy, enabling early identification, accurate diagnosis, and timely intervention of transformer oil sludge-related faults. It provides technical support for the full life-cycle health management of equipment and improves the safety and stability of power system operation. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating a monitoring and diagnostic method for a power transformer, provided as an embodiment of the present invention;

[0042] Figure 2 This is a structural block diagram of a monitoring and diagnostic device for a power transformer provided in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0044] In one embodiment, such as Figure 1 As shown, this application provides a monitoring and diagnostic method for power transformers, which may include the following steps:

[0045] Step S101: Acquire multi-dimensional monitoring data collected by the sensor array under the full operation and shutdown states of the power transformer, and use adaptive multi-scale filtering to process the multi-dimensional monitoring data to obtain a clean data sequence.

[0046] Specifically, a sensor array deployed at key parts of the power transformer synchronously collects multi-dimensional monitoring data under all operating and shutdown conditions. The sensor array includes oil sensors, temperature sensors, electric field sensors, load sensors, and environmental sensors. The corresponding monitoring data covers physicochemical parameters such as oil particle concentration and dielectric loss factor; electrical parameters such as real-time operating load, voltage, and current; winding and tank temperature field distribution data; internal electric field strength data; and environmental temperature and humidity data. An adaptive multi-scale filtering algorithm is used to process the collected multi-dimensional monitoring data. This algorithm performs multi-scale wavelet decomposition on the data, adaptively sets filtering thresholds based on the differences in noise and effective signal characteristics at different scales, and reconstructs the data after threshold processing of each scale component, ultimately outputting a clean data sequence that removes random noise and retains key features.

[0047] Step S102: Based on the clean data sequence, extract the time-series dynamic features of oil sludge accumulation through a long short-term memory network, and combine the trend quantification model to judge the oil sludge accumulation trend.

[0048] Furthermore, the obtained clean data sequence is divided into several time-series data segments according to a preset time window (set to 5-30 minutes / window based on monitoring accuracy requirements), which are then input into a Long Short-Term Memory (LSTM) network. Utilizing the LSTM network's ability to capture long-term dependencies in time-series data, gradient propagation and weight updates in the input, hidden, and output layers are used to extract time-series dynamic features related to oil accumulation. These features include key indicators such as particle concentration change rate, concentration fluctuation variance, and peak frequency. The extracted time-series dynamic features are then input into a trend quantification model. This model, based on an improved Gaussian process regression-temporal attention mechanism fusion algorithm, performs time-series fitting on the feature data to eliminate the influence of data dispersion. Through trend extrapolation, an oil accumulation trend curve is calculated, containing core accumulation feature values ​​such as accumulation amount and real-time accumulation rate at different time points, enabling a quantitative assessment of the oil accumulation development trend.

[0049] Step S103: If the accumulation characteristic value in the oil siltation accumulation trend curve exceeds the preset threshold, then collect multi-source key influencing factors related to siltation risk, and use an improved weighted evidence theory fusion algorithm to integrate the key influencing factors to obtain the dominant factor of siltation risk.

[0050] Preferably, the preset threshold is based on historical fault statistics of transformers, rated operating parameters of equipment, and industry insulation tolerance standards such as GB / T7595 "Quality of Transformer Oil in Operation". Through interval statistics and risk level mapping, if the accumulation amount or real-time accumulation rate in the oil accumulation trend curve obtained in step two exceeds the preset threshold, the multi-source key influencing factor collection process is triggered. The collected multi-source key influencing factors include oil physicochemical parameters (particle size distribution, moisture content), equipment operating status parameters (load fluctuation amplitude, operating time), environmental parameters (ambient temperature change rate, humidity), and structural parameters (winding arrangement, tank structure). The collected multi-source data undergoes standardized preprocessing (unifying dimensions to the [0,1] interval) and outlier removal (using the 3σ criterion). An improved weighted evidence theory fusion algorithm is used, with entropy weight method to determine the credibility weight of each influencing factor, dynamically adjusting the evidence conflict coefficient to resolve information contradictions, and integrating the effective information of all key influencing factors through evidence synthesis rules, ultimately obtaining the dominant factor of accumulation risk, which includes the core risk factor category, influence weight, and risk contribution.

[0051] Step S104: Input the siltation risk dominant factor into the convolutional neural network to locate the potential fault type, and generate a diagnostic report and real-time intervention instructions adapted to the fault scenario based on the potential fault type.

[0052] The obtained siltation risk dominance factors are input into a pre-trained convolutional neural network (CNN). This CNN extracts and reduces the dimensionality of deep features from the risk dominance factors through convolutional and pooling layers, maps them to the fault category space through a fully connected layer, and uses a Softmax classifier to accurately locate potential fault types. These potential fault types include insulation aging caused by oil contamination, partial discharge caused by particle siltation, and accelerated core wear, which are directly related to oil siltation. Based on the located potential fault types, combined with the corresponding risk level (based on the fault's impact range and evolution rate), affected components, and possible consequences, a structured fault diagnosis report is generated. If the diagnosis report clearly shows a fault category that meets the preset high-risk level, the real-time intervention command generation module is automatically triggered, outputting intervention commands adapted to the fault scenario. These commands include equipment load adjustment suggestions, maintenance and repair priorities, and emergency handling procedures.

[0053] The aforementioned monitoring and diagnostic method for power transformers acquires multi-dimensional monitoring data, including oil physicochemical properties, operating load, temperature field, electric field, and ambient temperature and humidity, collected by a sensor array under all transformer conditions. After adaptive multi-scale filtering and noise reduction, a clean data sequence is obtained. The clean data is then divided into preset time windows and input into a long short-term memory network to extract the temporal dynamic features of oil sludge accumulation. A trend quantification model is used to determine the sludge accumulation trend and generate a trend curve. If the accumulation amount or rate in the curve exceeds a preset threshold based on historical fault data and industry standards, key influencing factors of multi-source sludge risk are collected. After preprocessing, an improved weighted evidence theory is used to fuse these factors to obtain the dominant factor of sludge risk. This factor is then input into a convolutional neural network to locate potential fault types and generate an adaptive diagnostic report. If a high-risk fault exists, a real-time intervention command is triggered and output. This method optimizes data reliability, feature extraction effectiveness, and multi-source information fusion accuracy through multi-stage technology optimization. It overcomes the limitations of existing technologies such as single-parameter monitoring, insufficient trend prediction, and low diagnostic accuracy. It enables early identification, accurate diagnosis, and timely intervention of oil sludge-related faults, providing support for the full life cycle health management of equipment and improving the safety and stability of power system operation.

[0054] In one embodiment, inputting the clean data sequence into a long short-term memory network to extract the temporal dynamic features of oil sludge accumulation, and combining this with a trend quantification model to obtain an oil sludge accumulation trend curve, may include the following steps:

[0055] Step S201: The cleaning data sequence is divided into preset time windows and then input into a long short-term memory network to extract the time-series dynamic features of oil sludge accumulation.

[0056] Step S202: Based on the time-series dynamic characteristics of oil sludge, the accumulation rate index is calculated using a preset rate calculation formula; the accumulation rate index quantitatively reflects the rate of change of particle concentration in the oil over time.

[0057] Step S203: Input the accumulation rate index into the trend quantification model, and perform time-series fitting and trend extrapolation on the index through the model to obtain the oil siltation accumulation trend curve; the trend quantification model is constructed by the time-series attention mechanism.

[0058] Specifically, the clean data sequence is divided into several time-series data segments according to a preset time window. This division method is suitable for the processing requirements of Long Short-Term Memory (LSTM) networks for time-series data, ensuring the temporal continuity of feature extraction. The divided time-series data segments are input into the LSTM, utilizing its ability to capture long-term dependencies to extract the temporal dynamic features of oil accumulation. These features cover core information such as particle concentration variation patterns, fluctuation amplitudes, and peak occurrence intervals. Based on the extracted oil accumulation temporal dynamic features, an accumulation rate index is calculated using a preset rate calculation formula. This formula quantifies the rate of change of particle concentration in the oil over time by associating the temporal dynamic features with the original particle concentration data, establishing a quantitative mapping relationship between features and accumulation rate. The calculated accumulation rate index is input into a trend quantification model. This model is constructed using a temporal attention mechanism, which can adaptively focus on the importance of the accumulation rate index at different time points. By eliminating the influence of data discreteness through temporal fitting processing, and combined with a trend extrapolation algorithm, an oil accumulation trend curve is obtained. The curve can intuitively present the changes in accumulation rate and the evolution of cumulative accumulation at different time points.

[0059] This embodiment combines LSTM with time window segmentation to effectively improve the extraction of time-series dynamic features of oil sludge, providing high-quality input for calculating the accumulation rate index. The preset rate calculation formula quantifies the accumulation rate based on time-series dynamic features, avoiding calculation biases caused by solely relying on particle concentration differences and improving the accuracy of particle concentration change rate representation over time. The trend quantification model introduces a time-series attention mechanism, which strengthens the influence weight of rate indicators in key periods, improving the accuracy of time-series fitting and the reliability of trend extrapolation, making the obtained oil sludge accumulation trend curve more closely match the actual sludge development trend. The overall process solves the problems of insufficient time-series feature capture, large rate quantification deviation, and low trend prediction reliability in traditional processing, providing accurate and effective data support for subsequent oil sludge risk identification and trend early warning.

[0060] In one embodiment, the accumulation rate metric can be calculated using the following formula:

[0061]

[0062] in, express The rate of oil accumulation at any given time. Indicates the duration of the preset time window. This represents the working condition correction factor. , Indicates the real-time load of the transformer. Indicates the rated load. Indicates the LSTM extraction of the first Time-series dynamic characteristics of oil accumulation The corresponding attention weights , They represent Time and Oil particle concentration at any given time Represents the characteristic attenuation coefficient, with values ​​ranging from 1 to 2. Calibration based on historical data fitting. Indicates the LSTM output of the first... Dimensional time-series dynamic characteristics.

[0063] This embodiment effectively improves the accuracy and scenario adaptability of rate quantification. The introduction of the operating condition correction coefficient k solves the problem that traditional calculation methods do not consider the impact of load changes on the accumulation rate, making the rate index adaptable to different operating load scenarios. The attention weight, combined with the time-series dynamic features extracted by LSTM, strengthens the contribution of key features and avoids the loss of feature information caused by relying solely on particle concentration differences. The exponential decay term considers the dynamic correction of historical particle concentrations by time-series features, reducing the interference of data discreteness on rate calculation. Overall, this formula breaks through the simple calculation logic of traditional "concentration difference / time". Through multi-parameter collaborative correction, it improves the reliability of the accumulation rate index, providing high-quality data support for subsequent oil accumulation trend judgment based on this index, and helping to improve the timeliness and accuracy of accumulation risk identification.

[0064] In one embodiment, if the accumulation feature value in the oil sludge accumulation trend curve exceeds a preset threshold, then multiple key influencing factors related to sludge accumulation risk are collected, and an improved weighted evidence theory fusion algorithm is used to integrate the key influencing factors to obtain the dominant factor of sludge accumulation risk. This may include the following steps:

[0065] Step S301: If the amount of oil sludge accumulated or the real-time accumulation rate represented by the oil sludge accumulation trend curve exceeds the preset threshold, then collect multi-source heterogeneous information; the multi-source heterogeneous information includes oil physicochemical parameters, equipment operating load, temperature field distribution, electric field strength and ambient temperature and humidity data; the preset threshold is based on transformer historical fault data and insulation withstand standard calibration.

[0066] Step S302: After standardizing and preprocessing the multi-source heterogeneous information and removing outliers, construct a fusion preparation dataset containing feature identifiers, collection timestamps, and parameter values.

[0067] Step S303: An improved weighted evidence theory fusion algorithm is used to integrate temperature field distribution data and electric field intensity data in the fusion preparation dataset as the core evidence source, and combined with the credibility weight of each parameter to obtain a multi-dimensional comprehensive state vector.

[0068] Preferably, the fusion preparation dataset includes multiple parameters such as oil physicochemical parameters, equipment operating load, temperature field distribution, electric field strength, and ambient temperature and humidity. Considering the direct correlation between abnormal temperature field distribution and electric field strength distortion and oil sludge accumulation, and their higher weighting in influencing sludge accumulation risk, these two are designated as core evidence sources. The improved weighted evidence theory fusion algorithm first initializes the various parameters in the fusion preparation dataset as evidence sources, determining the basic probability allocation for each parameter. Then, combining the credibility weights of each parameter obtained through pre-data consistency verification, it weights the core and non-core evidence sources, strengthening the information contribution of core evidence sources and weakening the interference of low-credibility parameters. Through information integration operations, the algorithm transforms multi-source heterogeneous parameters into feature components of a unified dimension, ultimately outputting a multi-dimensional comprehensive state vector containing fused information of various influencing factors.

[0069] Step S304: Based on the multi-dimensional comprehensive state vector, the information gain of each feature component is calculated using the entropy weight method to determine the weight distribution of key influencing factors.

[0070] Furthermore, the values ​​of each feature component are first normalized to eliminate dimensional differences. Then, based on the definition of information entropy, the information entropy of each feature component is calculated. The smaller the information entropy, the greater the numerical difference of the feature component, the more effective information it carries, and the higher the corresponding information gain. Finally, according to the correspondence between information entropy and information gain, the information gain is converted into the weights of each feature component, forming a weight distribution of key influencing factors. This weight distribution objectively reflects the degree of influence of each feature component on the risk of oil sludge accumulation.

[0071] Step S305: Dynamically adjust the evidence conflict coefficient according to the weight distribution of key influencing factors to obtain an optimized evidence set.

[0072] Step S306: Based on the DS evidence synthesis rules, the fusion confidence of each state category is recalculated using the optimized evidence body set to generate a refined comprehensive state vector; the state categories include normal, mild siltation fault, and severe siltation fault.

[0073] The DS evidence synthesis rules are used to combine multiple evidence sources in the set. During the operation, based on the core logic of DS evidence synthesis, the basic probability assignments of each optimized evidence body to different state categories are fused, and the fusion confidence scores of the three state categories—normal, mild siltation fault, and severe siltation fault—are recalculated. The fusion confidence score directly represents the credibility of each state category. Combining the fusion confidence score and the original feature information of the multi-dimensional comprehensive state vector, a refined comprehensive state vector is constructed. This vector not only contains the specific confidence scores of the three state categories but also associates the key feature component information corresponding to each state category, achieving a precise quantitative representation of the oil siltation state.

[0074] Step S307: If the value of the feature component of the refined integrated state vector exceeds the secondary threshold, the dominant feature component in the refined integrated state vector is extracted using the variance contribution rate method to obtain the dominant factor of siltation risk; the secondary threshold is set based on the rated operating parameters of the equipment and the insulation safety threshold.

[0075] Specifically, if the oil accumulation amount or real-time accumulation rate, as represented by the oil accumulation trend curve, exceeds a preset threshold, the multi-source heterogeneous information collection process is triggered. The preset threshold is based on historical transformer fault statistics and industry insulation withstand standards. The multi-source heterogeneous information specifically covers oil physicochemical parameters, equipment operating load, temperature field distribution, electric field strength, and ambient temperature and humidity data. The collected multi-source heterogeneous information undergoes standardized preprocessing (unifying dimensions to a preset range) and outlier removal (using the 3σ criterion). After completion, a fusion preparation dataset containing feature identifiers, collection timestamps, and parameter values ​​is constructed. An improved weighted evidence theory fusion algorithm is used, with temperature field distribution data and electric field strength data from the fusion preparation dataset as core evidence sources. Combined with the credibility weights of each parameter, a multi-dimensional comprehensive state vector is obtained. Based on this multi-dimensional comprehensive state vector, the information gain of each feature component is calculated using the entropy weight method, thereby determining the weight distribution of key influencing factors. Then, the evidence conflict coefficient is dynamically adjusted according to this weight distribution to form an optimized evidence set. Based on the DS evidence synthesis rules, the fusion confidence of each state category is recalculated using an optimized evidence set to generate a refined comprehensive state vector. The state categories are explicitly divided into normal, mild siltation fault, and severe siltation fault. If the feature component values ​​of the refined comprehensive state vector exceed a secondary threshold (set based on the equipment's rated operating parameters and insulation safety threshold), the variance contribution rate method is used to extract the dominant feature components in the refined comprehensive state vector, ultimately obtaining the dominant factor of siltation risk.

[0076] This embodiment employs differentiated calibration of preset and secondary thresholds to adapt to the needs of trend warning and precise risk classification, enhancing the targeting of threshold settings. Preprocessing of multi-source heterogeneous information reduces the interference of data noise and outliers on the fusion results, ensuring the quality of the fusion preparation dataset. An improved weighted evidence theory fusion algorithm, combining core evidence sources and credibility weights, addresses the problem of insufficient fusion of multi-source heterogeneous information, improving the reliability of the multi-dimensional comprehensive state vector. The entropy weight method determines the weight distribution of key influencing factors, providing an objective basis for the dynamic adjustment of evidence conflict coefficients, effectively resolving contradictions between evidence and optimizing the quality of the evidence body. A refined comprehensive state vector generated based on DS evidence synthesis rules enables precise characterization of different siltation states. The variance contribution rate method extracts dominant feature components, focusing on core risk-causing factors and enhancing the directionality of siltation risk dominant factors. The overall process overcomes the limitations of traditional multi-source information fusion conflict handling and vague risk factor extraction, providing high-quality risk characterization data for the precise diagnosis of oil siltation-related faults, helping to improve the timeliness of fault warnings and the targeting of treatment.

[0077] In one embodiment, the fusion confidence of each state category can be calculated using the following formula:

[0078]

[0079] in, Indicates the first The fusion confidence of each state category, Indicates the preset state category. This indicates the number of evidence sources in the optimized evidence set. Indicates the first The weights of key influencing factors for each source of evidence are calculated using the entropy weight method. Indicates the first Each optimized evidence pair Basic probability assignment of class states This represents the dynamically adjusted coefficient of evidence, with a range of values. , , Indicates the first The initial conflict coefficients of each evidence source are obtained from data consistency checks. Indicates the first Conflict probability allocation of evidence sources This represents the empty set.

[0080] This embodiment integrates the confidence calculation formula to effectively improve the accuracy and scenario adaptability of confidence calculation. By introducing the weights of key influencing factors, it achieves differentiated weighting of the contribution of evidence sources, solving the problem of weakening high-value information caused by "equal weighting of all evidence" in traditional DS evidence synthesis. The dynamic conflict coefficient adopts the weighted synthesis of each evidence source weight, replacing the fixed conflict coefficient, avoiding the blindness of conflict handling, and can adaptively adjust the conflict suppression magnitude according to the importance of the evidence source. The conflict impact of high-weight evidence is reasonably weakened, and the invalid conflict of low-weight evidence is effectively suppressed. At the same time, the formula structure is simple, without complex nested operations, reducing the difficulty of engineering implementation. The output fused confidence can accurately represent the credibility of each state category, providing high-quality data support for the subsequent generation of refined comprehensive state vectors, and helping to improve the accuracy of oil sludge state judgment.

[0081] In one embodiment, inputting the siltation risk-dominant factor into a convolutional neural network to locate potential fault types, and generating diagnostic reports and real-time intervention instructions adapted to the fault scenario based on the potential fault types, may include the following steps:

[0082] Step S401: Input the dominant factor of siltation risk into a convolutional neural network for preliminary feature extraction to obtain a risk feature vector focusing on the risk-causing characteristics of siltation.

[0083] Step S402: Use a deep learning algorithm based on attention mechanism to classify multi-category faults on the risk feature vector to determine the potential fault types.

[0084] Step S403: Based on the potential fault type, call the preset fault diagnosis rule library and historical fault case library to generate fault diagnosis report data containing fault category, risk level and scope of impact.

[0085] Step S404: If the fault diagnosis report data contains a fault category that meets the preset high-risk level, then generate a real-time intervention instruction that is adapted to the fault scenario.

[0086] Specifically, the dominant risk factor of siltation is input into a pre-trained convolutional neural network. Through convolutional and pooling layers, the network performs preliminary feature extraction and dimensionality reduction on the deep siltation-related risk information within the dominant factor, filtering redundant information and outputting a risk feature vector focusing on the core features of siltation-related risks. Using this risk feature vector as input, a deep learning algorithm based on an attention mechanism is employed for multi-category fault classification. The attention mechanism adaptively focuses on feature components in the risk feature vector that contribute significantly to fault classification, strengthening the weight of key information, weakening interference from invalid information, and accurately determining the potential fault type. Based on the determined potential fault type, the system automatically calls upon a pre-set fault diagnosis rule base and a historical fault case base. Combining the diagnostic logic in the rule base with the experience in handling similar faults in the case base, it structurally integrates information such as fault category, risk level (based on fault evolution rate and impact range), and impact range (related to core transformer components and corresponding power supply areas) to generate standardized fault diagnosis report data. The fault diagnosis report data is verified for risk level. If there is a fault category that meets the preset high-risk level, then based on the fault type, scope of impact and real-time operating conditions of the equipment, a real-time intervention instruction adapted to the current fault scenario is generated. The instruction content covers core information such as operation and maintenance handling direction and equipment parameter adjustment suggestions.

[0087] This embodiment ensures the accuracy of fault diagnosis and the timeliness of handling. The convolutional neural network extracts preliminary features of the dominant factors of oil sludge accumulation risk, focusing on the risk-causing features and providing high-quality data input for subsequent classification. The deep learning algorithm based on the attention mechanism solves the problem of insufficient attention to key features in traditional classification algorithms, improving the accuracy of multi-category fault classification and reducing the probability of misjudgment and missed judgment. The use of a pre-set fault diagnosis rule base and a historical fault case database ensures the standardization, completeness, and reliability of fault diagnosis report data, avoiding the subjectivity and inconsistency of manually generated reports. The design of a high-risk level triggering real-time intervention command realizes automated connection from fault identification to handling command generation, shortening fault response time. The overall process overcomes the limitations of traditional diagnosis, such as unfocused feature extraction, insufficient classification accuracy, low report generation efficiency, and delayed intervention, providing technical support for the accurate diagnosis and timely handling of transformer oil sludge accumulation-related faults and ensuring equipment operational stability.

[0088] In one embodiment, such as Figure 2 As shown, this application also provides a monitoring and diagnostic device for power transformers, which may include:

[0089] The data acquisition module 501 is used to acquire multi-dimensional monitoring data collected by the sensor array under the full operation and shutdown conditions of the power transformer. The multi-dimensional monitoring data is processed by adaptive multi-scale filtering to obtain a clean data sequence.

[0090] The trend analysis module 502 is used to extract the time-series dynamic features of oil sludge accumulation based on clean data sequences through a long short-term memory network, and to analyze the oil sludge accumulation trend by combining it with a trend quantification model.

[0091] The risk assessment module 503 is used to collect multiple key influencing factors related to siltation risk if the accumulation characteristic value in the oil siltation accumulation trend curve exceeds a preset threshold. The module then integrates the key influencing factors using an improved weighted evidence theory fusion algorithm to obtain the dominant factor of siltation risk.

[0092] The fault diagnosis module 504 is used to input the dominant factor of siltation risk into the convolutional neural network to locate potential fault types, and generate a diagnostic report and real-time intervention instructions adapted to the fault scenario based on the potential fault types.

[0093] The aforementioned monitoring and diagnostic device for power transformers includes a data acquisition module, a trend analysis module, a risk assessment module, and a fault diagnosis module. These modules work together to achieve full-process monitoring and diagnosis of the oil sludge-related status of the power transformer. Specifically, the data acquisition module acquires multi-dimensional monitoring data collected by a sensor array under all operating and shutdown conditions of the power transformer. This multi-dimensional monitoring data covers information such as oil physicochemical parameters, equipment operating load, temperature field distribution, electric field strength, and ambient temperature and humidity. An adaptive multi-scale filtering algorithm is used to suppress noise in the multi-dimensional monitoring data, outputting a high-fidelity clean data sequence to provide reliable data input for subsequent processing. The trend analysis module takes the clean data sequence as input, divides it according to a preset time window, and inputs it into a long short-term memory network to extract the temporal dynamic features of oil sludge accumulation. These temporal dynamic features are then input into a trend quantification model. Through time series fitting and trend extrapolation, the oil sludge accumulation trend is determined, generating an oil sludge accumulation trend analysis model. The risk assessment module verifies the accumulated feature value in the accumulated trend curve in real time. If the feature value exceeds the preset threshold calibrated based on transformer historical fault data and industry insulation tolerance standards, it triggers the collection of multi-source key influencing factors. After standardization preprocessing and outlier removal of the collected factors, an improved weighted evidence theory fusion algorithm is used to integrate and process them to obtain the dominant factor of accumulated risk. The fault diagnosis module receives the dominant factor of accumulated risk, inputs it into a convolutional neural network, and locates the potential fault type through feature extraction and fault classification. Combining the risk level, impact range and other information of the potential fault type, it generates a diagnostic report adapted to the fault scenario. If there is a high-risk fault category in the report, it triggers the generation and output of real-time intervention instructions.

[0094] This embodiment effectively improves the accuracy and timeliness of transformer monitoring and diagnosis. The adaptive multi-scale filtering processing of the data acquisition module solves the noise interference problem in multi-dimensional monitoring data, ensuring data reliability. The trend analysis module combines long short-term memory networks and trend quantification models to enhance the effectiveness of extracting dynamic features of oil accumulation time series, improving the accuracy of accumulation trend prediction. The risk assessment module solves the problem of insufficient fusion of multi-source heterogeneous information through the fusion of multi-source key influencing factors and an improved weighted evidence theory algorithm, improving the accuracy of the dominant factors of accumulation risk. The fault diagnosis module uses convolutional neural networks to achieve precise location of potential faults, and generates diagnostic reports and real-time intervention instructions based on fault types, achieving automated connection between fault identification and handling suggestions. The overall device breaks through the limitations of single-module functions in existing technologies, realizing early identification, accurate diagnosis, and timely intervention of oil accumulation-related faults, providing technical support for the full life cycle health management of transformers, and improving the safety and stability of power system operation.

[0095] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0096] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a power transformer monitoring and diagnostic method as described above.

[0097] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0098] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0099] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A monitoring and diagnostic method for power transformers, characterized in that, The method includes: Multidimensional monitoring data collected by a sensor array under the full operation and shutdown states of a power transformer are acquired, and the multidimensional monitoring data are processed by adaptive multi-scale filtering to obtain a clean data sequence. Based on the clean data sequence, the time-series dynamic features of oil accumulation are extracted through a long short-term memory network, and the oil accumulation trend is judged by combining a trend quantification model. If the accumulation characteristic value in the oil sludge accumulation trend curve exceeds the preset threshold, then multiple key influencing factors related to sludge accumulation risk are collected, and the key influencing factors are integrated using an improved weighted evidence theory fusion algorithm to obtain the dominant factor of sludge accumulation risk. The siltation risk-dominant factor is input into a convolutional neural network to locate potential fault types, and diagnostic reports and real-time intervention instructions adapted to the fault scenarios are generated based on the potential fault types.

2. The method according to claim 1, characterized in that, The step of inputting the clean data sequence into a long short-term memory network to extract the temporal dynamic features of oil accumulation, and combining this with a trend quantification model to obtain the oil accumulation trend curve, includes: The cleaning data sequence is divided into preset time windows and then input into a long short-term memory network to extract the temporal dynamic features of oil sludge accumulation. Based on the time-series dynamic characteristics of oil accumulation, an accumulation rate index is calculated using a preset rate calculation formula; the accumulation rate index quantitatively reflects the rate of change of particle concentration in the oil over time. The accumulation rate index is input into the trend quantification model, and the index is subjected to time series fitting and trend extrapolation through the model to obtain the oil accumulation trend curve; the trend quantification model is constructed by the time series attention mechanism.

3. The method according to claim 2, characterized in that, The accumulation rate index is calculated using the following formula: in, express The rate of oil accumulation at any given time. Indicates the duration of the preset time window. This represents the working condition correction factor. , Indicates the real-time load of the transformer. Indicates the rated load. Indicates the LSTM extraction of the first Time-series dynamic characteristics of oil accumulation The corresponding attention weights , They represent Time and Oil particle concentration at any given time Represents the characteristic attenuation coefficient, with values ​​ranging from 1 to 10. Calibration based on historical data fitting. Indicates the LSTM output of the first... Dimensional time-series dynamic characteristics.

4. The method according to claim 1, characterized in that, If the accumulation characteristic value in the oil sludge accumulation trend curve exceeds a preset threshold, then multiple key influencing factors related to sludge accumulation risk are collected, and the key influencing factors are integrated using an improved weighted evidence theory fusion algorithm to obtain the dominant factor of sludge accumulation risk, including: If the amount of oil accumulation or the real-time accumulation rate represented by the oil accumulation trend curve exceeds a preset threshold, multi-source heterogeneous information is collected; the multi-source heterogeneous information includes oil physicochemical parameters, equipment operating load, temperature field distribution, electric field strength, and ambient temperature and humidity data; the preset threshold is based on transformer historical fault data and insulation withstand standard calibration. After standardizing and removing outliers from the multi-source heterogeneous information, a fusion preparation dataset containing feature identifiers, collection timestamps, and parameter values ​​is constructed. An improved weighted evidence theory fusion algorithm is adopted, using temperature field distribution data and electric field intensity data in the fusion preparation dataset as core evidence sources, and combining the confidence weights of each parameter to integrate and obtain a multi-dimensional comprehensive state vector; Based on the multi-dimensional integrated state vector, the information gain of each feature component is calculated using the entropy weight method to determine the weight distribution of key influencing factors. The evidence conflict coefficient is dynamically adjusted based on the weight distribution of the key influencing factors to obtain an optimized set of evidence. Based on the DS evidence synthesis rules, the fusion confidence of each state category is recalculated using the optimized evidence body set to generate a refined comprehensive state vector; the state categories include normal, mild silo fault, and severe silo fault; If the value of the feature component of the refined integrated state vector exceeds the second threshold, the dominant feature component in the refined integrated state vector is extracted using the variance contribution rate method to obtain the dominant factor of siltation risk; the second threshold is set based on the rated operating parameters of the equipment and the insulation safety threshold.

5. The method according to claim 4, characterized in that, The fusion confidence score for each state category is calculated using the following formula: in, Indicates the first The fusion confidence of each state category, Indicates the preset state category. This indicates the number of evidence sources in the optimized evidence set. Indicates the first The weights of key influencing factors for each source of evidence are calculated using the entropy weight method. Indicates the first Each optimized evidence pair Basic probability assignment of class states This represents the dynamically adjusted coefficient of evidence, with a range of values. , , Indicates the first The initial conflict coefficients of each evidence source are obtained from data consistency checks. Indicates the first Conflict probability allocation of evidence sources This represents the empty set.

6. The method according to claim 1, characterized in that, The step of inputting the siltation risk-dominant factor into a convolutional neural network to locate potential fault types, and generating diagnostic reports and real-time intervention instructions adapted to the fault scenarios based on the potential fault types, includes: The dominant factor of siltation risk is input into a convolutional neural network for preliminary feature extraction, resulting in a risk feature vector focusing on the risk-causing characteristics of siltation. The risk feature vector is subjected to multi-class fault classification using a deep learning algorithm based on an attention mechanism to determine the potential fault type; Based on the potential fault type, a fault diagnosis report data containing fault category, risk level, and scope of impact is generated by calling the preset fault diagnosis rule library and historical fault case library. If the fault diagnosis report data contains a fault category that meets the preset high-risk level, then a real-time intervention instruction adapted to the fault scenario is generated.

7. A monitoring and diagnostic device for a power transformer, characterized in that, The device includes: The data acquisition module is used to acquire multi-dimensional monitoring data collected by the sensor array under the full operation and shutdown states of the power transformer, and to process the multi-dimensional monitoring data using adaptive multi-scale filtering to obtain a clean data sequence. The trend analysis module is used to extract the time-series dynamic features of oil accumulation based on the clean data sequence through a long short-term memory network, and to analyze the oil accumulation trend by combining the trend quantification model. The risk assessment module is used to collect multiple key influencing factors related to siltation risk if the accumulation feature value in the oil siltation accumulation trend curve exceeds a preset threshold, and integrate the key influencing factors using an improved weighted evidence theory fusion algorithm to obtain the dominant factor of siltation risk. The fault diagnosis module is used to input the siltation risk-dominant factor into a convolutional neural network to locate potential fault types, and generate a diagnostic report and real-time intervention instructions adapted to the fault scenario based on the potential fault types.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images