A comprehensive monitoring method, system and storage medium for transformer oil quality

Abstract

The present invention discloses a comprehensive transformer oil quality monitoring method, system and storage medium, which relate to the field of transformer oil quality monitoring and solve the technical problem that the existing technology cannot monitor transformer oil quality in real time, resulting in difficulty in timely detecting oil quality changes. The present invention obtains transformer oil quality parameters in real time and preprocesses the oil; performs temperature compensation on the preprocessed refractive index based on temperature to obtain the actual refractive index and construct an oil quality feature vector; extracts a number of oil quality feature vectors from historical data to train an LSTM-Attention model to obtain an oil quality assessment model; the oil quality assessment model assesses the warning level of the transformer oil to be tested; the present invention can monitor the physical state of transformer oil in real time, issue a warning of oil quality deterioration in a timely manner, and ensure the safe operation of the power system.

Description

Technical Field

[0001] The present invention belongs to the field of transformer oil quality monitoring, and in particular relates to a transformer oil quality comprehensive monitoring method, system and storage medium. Background Art

[0002] As a core medium in power systems, transformer oil plays a crucial role in insulation, cooling, and arc protection. Its quality is directly related to the proper operation of transformers and the safety of power systems. Transformer oil's physical properties, such as refractive index, moisture content, acidity, and viscosity, can reveal its aging, contaminant content, and potential failure risks. Therefore, real-time and accurate monitoring of transformer oil is crucial for ensuring the safe operation and extending the life of power equipment.

[0003] Traditional oil quality testing typically requires shutting down the transformer to collect oil samples, which are then sent to a laboratory for chemical analysis. This offline testing method not only causes transformer downtime and increases downtime costs, but also lacks real-time monitoring, making it difficult to promptly detect changes in oil quality and potential equipment problems. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a comprehensive transformer oil quality monitoring method, system and storage medium, which are used to solve the technical problem that the prior art cannot monitor the transformer oil quality in real time, resulting in difficulty in timely detecting changes in oil quality.

[0005] To achieve the above objectives, a first aspect of the present invention provides a method for comprehensive monitoring of transformer oil quality, comprising:

[0006] Obtain transformer oil quality parameters in real time and perform pre-processing; oil quality parameters include refractive index, temperature, and oil flow pressure;

[0007] The oil quality parameters are acquired in real time by a fiber optic refractive index sensor; the fiber optic refractive index sensor is a dual-parameter composite probe comprising a fiber Bragg grating and a tilted fiber Bragg grating; the fiber Bragg grating is used to detect refractive index and temperature, and the tilted fiber Bragg grating is used to monitor oil flow pressure;

[0008] Performing temperature compensation on the pre-processed refractive index based on the temperature to obtain the actual refractive index;

[0009] The actual refractive index n of the transformer oil is obtained through the temperature compensation algorithm. The calculation formula is as follows:

[0010] n=n raw -α×ΔT1;

[0011] Among them, n rawis the pre-processed refractive index collected in real time, α = 2.3 × 10 -5 RIU / ℃, ΔT1 is the difference between the current measured temperature and the standard temperature;

[0012] The refractive index of transformer oil in this invention changes with temperature. Without temperature compensation, the refractive index data collected in real time will contain temperature-induced errors. This algorithm effectively eliminates the effect of temperature on refractive index measurements by subtracting the refractive index deviation (α × ΔT) caused by temperature variation, thereby obtaining a value closer to the true refractive index of the transformer oil and significantly improving measurement accuracy. This reduces the potential for misdiagnosis of equipment failures and unnecessary maintenance work caused by inaccurate measurements due to temperature errors, thereby reducing transformer maintenance costs and operational risks.

[0013] Constructing the oil quality feature vector based on the actual refractive index, pre-processed temperature and oil flow pressure;

[0014]

[0015] Where Δt is the analysis period, Δn is the difference between the highest and lowest actual refractive indexes within the analysis period, is the actual refractive index change rate, ΔT is the difference between the highest temperature and the lowest temperature during the analysis period, f is the oil flow pulsation frequency, is the refractive index-temperature coupling parameter, ∫(Δn) 2 dt is the cumulative change in refractive index, is the oil flow-refractive index dynamic coupling, Var(Δn) is the refractive index sliding variance, j is the number of historical alarms, and the number of historical alarms is the cumulative number of times the refractive index change rate of the transformer oil exceeds the preset safety threshold during continuous monitoring;

[0016] This method analyzes the factors influencing transformer oil quality. By incorporating the relevant features of these factors into vectors, it can fully account for the interactions between various factors. The actual refractive index change rate and the cumulative refractive index change can reflect the dynamic changes in transformer oil refractive index in real time. The former reflects the rate of change of refractive index per unit time, while the latter reflects the overall accumulation of refractive index changes within the analysis period, helping to promptly identify short-term fluctuations and long-term trends in oil quality. The refractive index sliding variance measures the degree of refractive index fluctuation within a certain time window and can reflect the stability of oil quality. A larger variance indicates more severe refractive index fluctuations and less stable oil quality; conversely, a smaller variance indicates relatively stable oil quality. This provides an important basis for evaluating transformer oil operational reliability. The refractive index-temperature coupling parameter and the oil flow-refractive index dynamic coupling reveal the intrinsic connection between transformer oil refractive index, temperature, and oil flow. By analyzing these coupling relationships, we can gain an in-depth understanding of how oil quality changes under different working conditions, providing clues for further research into the mechanism of oil degradation. The historical alarm count also records the cumulative number of times the refractive index change rate of transformer oil exceeds the preset safety threshold during continuous monitoring. When the number of historical alarms is large, it indicates that the oil quality may be abnormal and requires special attention and further inspection.

[0017] Several oil quality feature vectors are extracted from historical data to train an LSTM-Attention model, resulting in an oil quality assessment model. The LSTM-Attention model is used to predict the transformer oil quality warning level based on the transformer oil quality feature vectors.

[0018] The oil quality feature vector of the transformer oil to be tested is input into the oil quality assessment model, and the warning level of the transformer oil quality is output.

[0019] Preferably, the oil quality parameter pre-processing process includes:

[0020] Fitting the oil quality parameters collected during the analysis period to obtain curves of the oil quality parameters; placing the curves of the oil quality parameters and the corresponding preset standard parameters in the same coordinate system;

[0021] Obtain abnormal areas from several areas enclosed by the oil quality curve and the preset standard curve, calculate the area of ​​the abnormal area, and determine whether the area of ​​the abnormal area is smaller than the preset area; if so, remove the oil quality parameters in the abnormal area; if not, retain the oil quality parameters in the abnormal area;

[0022] The calculation formula for the area of ​​the abnormal region is as follows:

[0023]

[0024] Where i is the serial number of each oil quality parameter, i = 0, 1, 2, representing temperature, refractive index, and oil flow pressure, respectively; fi(t) is the variation curve of oil quality parameter i within the analysis period, gi(t) is the preset standard curve of oil quality parameter i; t1 and t2 are the times corresponding to the intersection of the two curves, respectively.

[0025] Preferably, the abnormal area is the area enclosed by the oil quality curve above a preset standard curve, which is obtained by fitting based on several historical normal oil quality parameters; wherein the normal oil quality parameters are oil quality parameters within a preset range.

[0026] It should be noted that the abnormal area refers to the oil quality parameters that exceed the normal upper limit.

[0027] The present invention fits each oil quality parameter into a curve and places it in the same coordinate system as the corresponding preset standard parameter. This visualization method can intuitively present the changing trends of oil quality parameters over time or other variables. The area enclosed by the oil quality curve and the preset standard curve is extracted and calculated. This area is used as the criterion for determining abnormal areas. Compared with simple numerical threshold judgment, this method more comprehensively considers the changes in oil quality parameters over time. For example, some oil quality parameters may experience small fluctuations over a short period of time, but the overall trend remains within the normal range. Using the curve-enclosed area can prevent such cases from being misclassified as abnormal. When the area of ​​the abnormal area is smaller than the preset area, the abnormality is likely caused by accidental factors such as measurement error and environmental interference. Removing the oil quality parameters from this abnormal area effectively prevents the impact of irrelevant interference data on subsequent analysis, improving data accuracy and reliability. If the area of ​​the abnormal area is not smaller than the preset area, the oil quality parameters within the abnormal area are retained. This ensures that truly valuable abnormal data is not missed, providing an important basis for in-depth analysis of the causes of oil quality anomalies and the development of appropriate treatment measures.

[0028] Preferably, the training process of the oil quality assessment model is as follows:

[0029] Several groups of historical oil quality feature vectors are collected from historical data and labeled with warning labels. The labeled historical oil quality feature vectors are divided into a training dataset and a validation dataset. The LSTM-Attention model is trained using the labeled historical oil quality feature vectors to obtain an oil quality assessment model.

[0030] Preferably, the LSTM-Attention model includes an input layer, an LSTM layer, an Attention layer and an output layer;

[0031] The input layer is used to receive the oil quality feature vector;

[0032] The LSTM layer includes a first layer and a second layer; the first layer is used to extract short-term dynamic features of the oil quality feature vector with a span of less than N time steps; the second layer is used to extract long-term dynamic features of the oil quality feature vector with a span of more than 3N time steps; wherein N is a positive integer;

[0033] The Attention layer is used to determine the weight coefficient of the time step;

[0034] The output layer performs weighted fusion on the short-term dynamic features and long-term dynamic features extracted by the LSTM layer based on the time step weight coefficient determined by the Attention layer to generate a comprehensive feature vector. The comprehensive feature vector is mapped to a preset transformer oil quality warning level classification space according to the comprehensive feature vector, and finally the warning level of the transformer oil quality is output.

[0035] Preferably, the weight coefficient of the time step is obtained by the following methods, including:

[0036] By formula α t =softmax[u T ×tanh(W×h t +b)] calculate the time step weight coefficient α t ; Where t is the time step, softmax is the activation function, tanh is the hyperbolic tangent activation function, h t is the hidden state at time step t, W is the weight matrix, b is the bias vector, u is the attention vector, and T is the transpose of the attention vector.

[0037] The LSTM-Attention model of this invention consists of an input layer, an LSTM layer, an Attention layer, and an output layer, each with a clear division of labor. The input layer receives oil quality feature vectors, providing a data foundation for subsequent processing; the LSTM layer extracts short-term and long-term dynamic features, capturing the temporal changes in oil quality; the Attention layer determines the weight coefficients of time steps, enhancing attention to important temporal information; and the output layer performs feature fusion and outputs warning levels. The entire model is compact and logically clear.

[0038] The LSTM layer is divided into the first and second layers, which respectively extract short-term dynamic features spanning less than N time steps and long-term dynamic features spanning more than 3N time steps. This hierarchical feature extraction approach can simultaneously consider both short-term fluctuations and long-term trends in oil quality, enabling the model to gain a more comprehensive and in-depth understanding of changes in oil quality.

[0039] By calculating the weight coefficients of time steps, the Attention layer automatically focuses on the time steps that are more important for oil quality assessment, highlighting the impact of key features. This mechanism avoids the drawback of traditional models that treat all time steps equally, improving the model's sensitivity to important information and its efficiency.

[0040] The output layer performs weighted fusion of the short-term and long-term dynamic features extracted by the LSTM layer based on the time step weight coefficient determined by the Attention layer to generate a comprehensive feature vector. The weighted fusion method can fully utilize the advantages of different features, making the comprehensive feature vector more accurately reflect the true state of oil quality, thereby improving the accuracy of warning level classification.

[0041] Preferably, the second aspect of the present invention provides a transformer oil quality comprehensive monitoring system, comprising a data acquisition module, a data processing module, an oil quality assessment module and a blockchain trusted management module;

[0042] The data acquisition module is used to collect the refractive index, temperature and oil flow pressure of the transformer oil in real time;

[0043] The data processing module is used to process and analyze the detection data;

[0044] The oil quality assessment module evaluates and warns of oil quality degradation based on the LSTM-Attention model;

[0045] The blockchain trusted management module is used to perform trusted management of detection data.

[0046] Preferably, the data upload process of the blockchain trusted management module is that the edge device uploads the monitored data to the side chain and generates a unique identifier. The smart contract in the main chain records the unique identifier and the corresponding timestamp, and adopts a dynamic consensus mechanism to verify and store data.

[0047] The edge device of the present invention uploads monitoring data to the sidechain and generates a unique identifier. This identifier acts as a data "identity card" and is tightly bound to the data once generated. During subsequent transmission and storage, if the data is maliciously tampered with, the corresponding unique identifier will not match. This allows for rapid detection of data anomalies, effectively preventing data tampering and ensuring data integrity.

[0048] Using a combined sidechain and mainchain approach, the sidechain receives data uploaded by edge devices, reducing the burden on the mainchain while also providing an additional layer of security. The mainchain, through smart contracts, records unique identifiers and timestamps, further ensuring data immutability and security. Even if the sidechain is attacked, the records on the mainchain remain reliable proof of data authenticity.

[0049] Preferably, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements a method for comprehensive monitoring of transformer oil quality when executed by a processor.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention uses a dual-parameter composite probe combined with a matrix decoupling algorithm to achieve high-precision simultaneous measurement of refractive index, temperature, and oil flow pressure parameters, overcoming the limitation of traditional sensors that only monitor a single parameter. Furthermore, based on the LSTM-Attention multidimensional analysis model, through 8-dimensional feature vectors such as refractive index change rate, temperature gradient, oil flow pulsation frequency, and sliding window time-frequency analysis, combined with a double-layer LSTM with a 128-node short-term feature layer and a 64-node long-term feature layer and an attention mechanism, dynamic prediction of oil quality degradation is achieved. Furthermore, based on the blockchain's main chain-side chain dual-chain architecture and dynamic consensus mechanism, the monitoring data is ensured to be tamper-proof throughout the process, improving evidence storage efficiency, supporting dynamic consensus mechanisms, and increasing system throughput. The present invention combines fiber optic refractive index sensing, deep learning time series prediction, and blockchain evidence storage to address the core issues of traditional oil quality monitoring, such as poor real-time performance, low data credibility, and delayed warnings. This ensures that the system can achieve high-precision, real-time monitoring of transformer oil quality, while ensuring reliable data management and early warning of oil quality degradation, providing strong protection for the safe operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 Schematic diagram of the process of the present invention;

[0054] Figure 2 This is a flow chart of the oil quality assessment model training method of the present invention;

[0055] Figure 3 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0056] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0057] See also Figure 1 The first embodiment of the present invention provides a method for comprehensive monitoring of transformer oil quality, comprising:

[0058] A fiber optic refractive index sensor is used to collect transformer oil quality parameters in real time, including refractive index, temperature, and oil flow pressure. The collected refractive index is temperature compensated to obtain the actual refractive index.

[0059] Specifically, the optical fiber refractive index sensor adopts a dual-parameter composite probe, which includes a fiber Bragg grating and a fiber Bragg grating tilted grating;

[0060] The fiber Bragg grating (FBG) is inscribed using a phase mask method. During the inscription process, parameters such as the UV laser power, exposure time, and phase mask period are precisely controlled. For example, the UV laser power is stabilized at 50mW, the exposure time is set to 30s, and the phase mask period is set to 1060nm. This ensures the accuracy and stability of the fiber Bragg grating's reflection wavelength, enabling it to accurately sense changes in the transformer oil's refractive index and temperature.

[0061] The fiber Bragg grating (FBG) tilted grating precisely adjusts the grating's tilt angle to 8° to detect transformer oil flow pressure. This angle has been verified through extensive experiments to ensure sensitivity to oil flow shear forces while maintaining the ability to monitor refractive index. The specific detection process is as follows:

[0062] A clear oil flow direction arrow is marked on the surface of the fiber optic probe to guide alignment during installation, ensuring that the probe maintains a preset angle (such as parallel or perpendicular) to the oil flow direction. The arrow direction is aligned with the fiber's deformation-sensitive axis, allowing lateral pressure changes (caused by oil flow impact) to be accurately captured by the fiber Bragg grating tilt grating. When the transformer oil flows in the direction of the arrow, the oil flow exerts lateral shear force on the probe, causing micron-level deformation of the fiber (such as bending or stretching). The grating period changes due to deformation, and the characteristic wavelength of the reflection spectrum shifts, with the offset proportional to the oil flow pressure. The relationship between the direction marked by the arrow and the deformation direction can be used to distinguish between forward and reverse oil flow.

[0063] Furthermore, each oil quality parameter collected during the analysis period is fitted to obtain a curve of each oil quality parameter; each oil quality parameter curve is placed in the same coordinate system as the corresponding preset standard parameter;

[0064] Obtain abnormal areas from several areas enclosed by the oil quality curve and the preset standard curve, calculate the area of ​​the abnormal area, and determine whether the area of ​​the abnormal area is smaller than the preset area; if so, remove the oil quality parameters in the abnormal area; if not, retain the oil quality parameters in the abnormal area;

[0065] The abnormal area is the area enclosed by the oil quality curve above the preset standard curve; the preset standard curve is obtained by fitting based on several historical normal oil quality parameters; and the normal oil quality parameters are oil quality parameters within the preset range.

[0066] For example, within an analysis cycle (e.g., 24 hours), the oil quality parameters, including actual refractive index, temperature, and oil flow pressure, are collected once every hour through sensors installed on the transformer. A total of 24 sets of data are collected, and the analysis cycle can be set according to actual conditions.

[0067] Using curve fitting methods such as the least squares method, 24 sets of data for actual refractive index, temperature, and oil flow pressure were fitted to obtain curve equations for each oil quality parameter. For example, for the actual refractive index data, the curve equation obtained was n = f1(t) (where t is time), the temperature curve equation was T = f2(t), and the oil flow pressure curve equation was P = f3(t).

[0068] Assume that the oil quality parameters of the transformer under normal operating conditions in the past few years are collected. According to the preset normal value range of each oil quality parameter, the data within the normal range is screened out to obtain several sets of historical normal oil quality parameters;

[0069] Using the filtered historical normal oil quality parameters, curve fitting is performed for the actual refractive index, temperature, and oil flow pressure to obtain the preset standard curve equations. Assume that the standard curve equation for the actual refractive index is ns = g1(t), the standard curve equation for temperature is Ts = g2(t), and the standard curve equation for oil flow pressure is Ps = g3(t).

[0070] Plot the actual refractive index curve n = f1(t) against the standard curve ns = g1(t), the temperature curve T = f2(t) against the standard curve Ts = g2(t), and the oil flow pressure curve P = f3(t) against the standard curve Ps = g3(t) in the same coordinate system. For each oil quality parameter, identify the area enclosed by the oil quality curve above the preset standard curve; these areas are considered abnormal regions.

[0071] The area of ​​each abnormal region is calculated using a numerical integration method (such as the trapezoidal method). For example, for an abnormal region of actual refractive index, its area Sn can be calculated using the following formula:

[0072]

[0073] Where i is the serial number of each oil quality parameter, assuming that i = 0 represents temperature, i = 1 represents refractive index, and i = 2 represents oil flow pressure; fi(t) is the variation curve of oil quality parameter i within the analysis period, gi(t) is the preset standard curve of oil quality parameter i; t1 and t2 are the times corresponding to the intersection of the two curves, respectively.

[0074] Preset area setting: According to actual requirements and experience, preset the areas of abnormal regions of the actual refractive index, temperature, and oil flow pressure as Sn, ST, and SP respectively;

[0075] For the area S0 of the temperature abnormal region, if S0 < ST, remove the temperature data at each time point within this abnormal region; if S0 ≥ ST, retain the temperature data at each time point within this abnormal region;

[0076] For the area S1 of the actual refractive index abnormal region, if S1 < Sn, remove the actual refractive index data at each time point within this abnormal region; if S1 ≥ Sn, retain the actual refractive index data at each time point within this abnormal region;

[0077] For the area S2 of the oil flow pressure abnormal region, if S2 < SP, remove the oil flow pressure data at each time point within this abnormal region; if S2 ≥ SP, retain the oil flow pressure data at each time point within this abnormal region.

[0078] Precisely measure the wavelength offset through a demodulator, and combine it with a temperature compensation algorithm to obtain the actual refractive index value of the transformer oil.

[0079] Since temperature changes will affect the refractive index of the optical fiber, the reference grating method is adopted for temperature compensation in this system. By measuring the wavelength drift caused by temperature and differentiating it from the refractive index change, the interference of temperature is eliminated. The calculation formula for the actual refractive index n is as follows:

[0080] n = n raw -α × ΔT1;

[0081] where, n raw is the refractive index after preprocessing of real-time acquisition, α = 2.3 × 10 -5 RIU / ℃, and ΔT1 is the difference between the current measured temperature and the standard temperature.

[0082] Furthermore, for the mixed signal collected by the dual-parameter composite probe, a matrix decoupling algorithm is used for decoupling. By establishing a mathematical relationship model among the refractive index, temperature, and oil flow pressure signals, the corresponding decoupling algorithm program is written. During the program implementation process, matrix operations are performed on the collected signals to separate the independent refractive index, temperature, and oil flow pressure signals, providing accurate data for subsequent analysis. To remove the noise in the data, a digital filtering algorithm is adopted. For example, the Kalman filtering algorithm is used to process the data, which can perform optimal estimation on the data containing noise according to the state equation and observation equation of the system, effectively improving the quality of the data.

[0083] It should be noted that when installing the fiber optic refractive index sensor, a location in the transformer oil circulation pipeline with stable oil flow and uniform temperature should be selected. A specially designed mounting fixture secures the dual-parameter composite probe firmly inside the pipeline, ensuring full contact between the probe and the oil flow and preventing displacement or damage due to oil flow impact. During installation, high-precision optical alignment equipment is used to ensure accurate connection between the optical fiber and the probe, minimizing optical signal loss. After the connection is complete, the sensor is calibrated using a standard refractive index oil sample, environments with different temperatures, and a simulated oil flow device. Using an oil sample with a known refractive index, the sensor's output signal is measured, and the sensor's measurement results are calibrated and corrected based on the theoretical relationship between wavelength offset and refractive index change. For temperature calibration, the sensor's output is recorded under different temperature environments, and temperature compensation algorithms (such as the reference grating method) are used to compensate for temperature-induced wavelength drift to ensure accurate and reliable measured refractive index values. According to the oil flow parameters, a simulated oil flow device is used to generate oil flows with different flow rates and pressures. The fiber Bragg grating tilt grating is used to monitor the lateral pressure changes caused by the oil flow, and the function of monitoring the shear force of the oil flow is calibrated to determine the sensitivity and response characteristics of the sensor.

[0084] Several oil quality feature vectors are extracted from historical data to train the LSTM-Attention model, and an oil quality assessment model is obtained through training.

[0085] Specifically, before extracting oil quality feature vectors, the collected oil quality parameters are preprocessed. First, the oil quality parameters are denoised to remove outliers caused by factors such as sensor noise and electromagnetic interference. Data smoothing is performed using methods such as median filtering to improve data quality. Then, the oil quality parameters are normalized, mapping parameter values ​​in different ranges to the interval [0, 1]. This ensures that different features have equal weights, facilitating subsequent model training and analysis.

[0086] Extract multi-dimensional oil quality feature vectors based on the needs of oil quality degradation assessment. Calculate the refractive index change rate, refractive index-temperature coupling parameter, cumulative refractive index change, oil flow-refractive index dynamic coupling, and refractive index sliding variance based on the collected oil quality parameters.

[0087] The calculation formula for the refractive index change rate is: The calculation formula of the refractive index-temperature coupling parameter is: The calculation formula for the cumulative change in refractive index is ∫(Δn) 2 dt; The calculation formula of oil flow-refractive index dynamic coupling is: The calculation formula of the refractive index sliding variance is Var(Δn).

[0088] Furthermore, the oil quality feature vector is constructed as follows:

[0089]

[0090] Wherein, Δt is the analysis period, Δn is the difference between the highest actual refractive index and the lowest actual refractive index within the analysis period, ΔT is the difference between the highest temperature and the lowest temperature within the analysis period, f is the oil flow pulsation frequency, and j is the number of historical alarms, which is the cumulative number of times that the refractive index change rate of the transformer oil exceeds the preset safety threshold during continuous monitoring.

[0091] See also Figure 2 , obtain several historical oil quality feature vectors and corresponding warning labels from historical data; wherein the warning labels correspond to the warning levels one by one; the warning labels can be set as natural numbers;

[0092] The oil quality feature vectors and warning labels are integrated into several sets of training data and test data. The LSTM-Attention model is trained using the training data. The trained LSTM-Attention model is tested using the test data and adjusted based on the test results. Finally, an oil quality assessment model is obtained, whose input is the oil quality feature vector and whose output is the warning label.

[0093] For example, the extracted historical data is divided into training data and test data in a ratio of 7:3. The LSTM-Attention model includes an input layer, an LSTM layer, an Attention layer, and an output layer.

[0094] During the training process, the input layer is used to receive the oil quality feature vector;

[0095] The two-layer network structure of LSTM is set as follows:

[0096] The first layer extracts short-term features and is configured as a short-time window processing unit with high forget gate sensitivity. By dynamically adjusting the weight ratio of the input gate to the forget gate (the recommended range is 1:1.5-2.0), it prioritizes capturing short-term dynamic features in the input sequence that span less than N time steps. These short-term features include but are not limited to: instantaneous signal fluctuation patterns, local gradient change characteristics, and high-frequency component correlation characteristics. There are 128 nodes in total.

[0097] The second layer extracts long-term features: It forms a cascade structure with the first layer, with its memory unit capacity expanded to 1.5-2 times that of the first layer. It adopts a decreasing learning rate strategy (the initial value is set to 0.6-0.8 times that of the first layer). By strengthening the propagation path of the historical state vector, it extracts long-term dependency features spanning more than 3N time steps in the input sequence. These long-term features include: periodic regularity features, trend evolution features, and cross-time correlation features. There are 64 nodes in total; where N is a positive integer;

[0098] The Attention layer determines the weight coefficient of the time step, through the formula α t =softmax[u T ×tanh(W×h t +b)] calculate the weight coefficient α of the time step t ; Where t is the time step, softmax is the activation function, tanh is the hyperbolic tangent activation function, h t is the hidden state at time step t, W is the weight matrix, b is the bias vector, u is the attention vector, and T is the transpose of the attention vector.

[0099] W, b, and u are all trainable parameters, that is, variables that the model automatically learns and adjusts through training data, and their values ​​are updated by optimization algorithms such as gradient descent during the backpropagation process; W and u jointly determine which hidden state features are more important for the current task, and b allows the model to apply a baseline offset to the features before nonlinear transformation, avoiding sensitive dependence on zero input.

[0100] The model's training parameters were set, such as a learning rate of 0.001, a training batch size of 32, and 100 training rounds. During training, the model was validated using validation data, and the model parameters were adjusted based on the loss function value of the validation data to prevent overfitting. After training, the oil quality feature vector obtained from real-time oil quality parameters was input into the model, which then output a prediction result. Based on the prediction result and the preset warning level rules, the corresponding warning level was generated.

[0101] The oil quality characteristic parameters of the transformer oil to be tested are input into the oil quality assessment model, and the warning level of the transformer oil quality is output.

[0102] The LSTM model proposed in this paper is a deep learning model capable of processing time series data, making it suitable for handling nonlinear and time-varying data. Combining LSTM with the Attention mechanism enables dynamic prediction of oil quality data, proactively identifying trends in oil quality degradation and providing early warning of abnormal oil quality conditions. Blockchain, as a distributed ledger technology, can provide reliable data storage and auditing capabilities for oil quality monitoring data. Through blockchain's decentralized nature and encryption algorithms, the security and immutability of oil quality data during collection, storage, transmission, and auditing are ensured, meeting the power system's requirements for data transparency, reliability, and security.

[0103] See also Figure 3 , a second aspect of the present invention provides a transformer oil quality comprehensive monitoring system, including a data acquisition module, a data processing module, an oil quality assessment module and a blockchain trusted management module;

[0104] The data acquisition module collects the refractive index, temperature and oil flow pressure of the transformer oil in real time;

[0105] The data processing module processes and analyzes the collected data;

[0106] It should be noted that the data acquisition module utilizes a high-speed, high-precision data acquisition card. With a sampling frequency exceeding 10kHz, this card can rapidly capture the weak optical signals output by the sensor and convert them into digital signals. Its 16-bit resolution ensures precise data acquisition and accurately captures minute changes in refractive index, temperature, and oil flow parameters.

[0107] The data acquisition card is connected to the sensor via a high-speed data cable. To reduce electromagnetic interference, the data cable uses well-shielded and well-grounded cables. A combination of wired and wireless data transmission is employed. Inside the transformer, due to the complex electromagnetic environment, shielded twisted-pair cables are used to ensure data is protected from electromagnetic interference. Outside the transformer, a 5G communication module is used to enable remote data transmission, transmitting collected data in real time to a data processing center.

[0108] During data transmission, data verification and encryption technologies are used to ensure data integrity and accuracy. For example, a cyclic redundancy check algorithm is used to verify data to ensure that no errors occur during data transmission; the Advanced Encryption Standard encryption algorithm is used to encrypt data to prevent data theft or tampering.

[0109] The oil quality assessment module provides early warning of oil quality based on the LSTM-Attention model;

[0110] The blockchain trusted management module performs trusted management of detection data.

[0111] In one embodiment, the blockchain architecture is constructed as follows: The blockchain trusted management module adopts a dual-chain architecture: a main chain and a side chain. The main chain is built on Hyperledger Fabric, with multiple servers with stable performance and strong computing power selected as main chain nodes. Smart contracts are deployed on the main chain to store hash values ​​of key events, such as early warning triggers and operation and maintenance operations. Smart contracts are written in Go and follow Hyperledger Fabric's smart contract development specifications to ensure their security and reliability. The side chain is built using IPFS (InterPlanetary File System), leveraging IPFS's distributed storage features to store raw sensor data streams. Within the IPFS network, multiple nodes are deployed across different geographical locations to improve data storage security and access efficiency. Data stored on the IPFS side chain is integrity verified using a Merkle Tree structure. During data storage, the data is divided into multiple small blocks, and a hash value is calculated for each small block. The root hash value of the Merkle tree is then calculated layer by layer and stored on the main chain to verify data integrity.

[0112] Data upload and consensus mechanism: A data upload program is written on the edge device. When new oil quality monitoring data is collected, it is uploaded to the IPFS sidechain. The IPFS sidechain returns a unique identifier, and the edge device sends the unique identifier and corresponding timestamp to the main chain smart contract for recording. During the data upload process, the data is compressed to reduce data transmission volume and storage space.

[0113] To improve system throughput and ensure data credibility, a dynamic consensus mechanism is employed. Network load varies with the seasons; for example, the number of nodes is set to 15 in summer and 10 in winter. The zk-SNARK (zero-knowledge succinct non-interactive argument) algorithm is used during consensus verification. When verifying data, nodes generate proofs using the zk-SNARK algorithm, allowing other nodes to verify the integrity and authenticity of the data without knowing the specific data content. This approach improves verification efficiency while protecting data privacy and enhancing the credibility of the entire system.

[0114] An embodiment of the present application also provides a computer-readable storage medium, comprising instructions, which, when executed on a computer, enables the computer to execute any of the above methods.

[0115] Some of the data in the above formula are calculated by removing the dimensions and taking their numerical values. The formula is a formula that is closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0116] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A comprehensive monitoring method for transformer oil quality, characterized in that: include: Obtain transformer oil quality parameters in real time and perform pre-processing; oil quality parameters include refractive index, temperature, and oil flow pressure; The oil quality parameters are acquired in real time by a fiber optic refractive index sensor; the fiber optic refractive index sensor is a dual-parameter composite probe comprising a fiber Bragg grating and a tilted fiber Bragg grating; the fiber Bragg grating is used to detect refractive index and temperature, and the tilted fiber Bragg grating is used to monitor oil flow pressure; Performing temperature compensation on the pre-processed refractive index based on the temperature to obtain the actual refractive index; The actual refractive index n of the transformer oil is obtained through the temperature compensation algorithm. The calculation formula is as follows: n=n raw -α×ΔT1; Among them, n raw is the pre-processed refractive index collected in real time, α = 2.3 × 10 -5 RIU / ℃, ΔT1 is the difference between the current measured temperature and the standard temperature; Constructing the oil quality feature vector based on the actual refractive index, pre-processed temperature and oil flow pressure; Where Δt is the analysis period, Δn is the difference between the highest and lowest actual refractive indexes within the analysis period, is the actual refractive index change rate, ΔT is the difference between the highest temperature and the lowest temperature during the analysis period, f is the oil flow pulsation frequency, is the refractive index-temperature coupling parameter, ∫(Δn) 2 dt is the cumulative change in refractive index, is the oil flow-refractive index dynamic coupling, Var(Δn) is the refractive index sliding variance, j is the number of historical alarms, and the number of historical alarms is the cumulative number of times the refractive index change rate of the transformer oil exceeds the preset safety threshold during continuous monitoring; Several oil quality feature vectors are extracted from historical data to train an LSTM-Attention model, resulting in an oil quality assessment model. The LSTM-Attention model is used to predict the transformer oil quality warning level based on the transformer oil quality feature vectors. The oil quality feature vector of the transformer oil to be tested is input into the oil quality assessment model, and the warning level of the transformer oil quality is output.

2. A transformer oil quality comprehensive monitoring method according to claim 1, characterized in that: The oil quality parameter preprocessing process includes: Fitting the oil quality parameters collected during the analysis period to obtain curves of the oil quality parameters; placing the curves of the oil quality parameters and the corresponding preset standard parameters in the same coordinate system; Obtain abnormal areas from several areas enclosed by the oil quality curve and the preset standard curve, calculate the area of ​​the abnormal area, and determine whether the area of ​​the abnormal area is smaller than the preset area; if so, remove the oil quality parameters in the abnormal area; if not, retain the oil quality parameters in the abnormal area; The calculation formula for the area of ​​the abnormal region is as follows: Where i is the serial number of each oil quality parameter, i = 0, 1, 2, representing temperature, refractive index, and oil flow pressure, respectively; fi(t) is the variation curve of oil quality parameter i within the analysis period, gi(t) is the preset standard curve of oil quality parameter i; t1 and t2 are the times corresponding to the intersection of the two curves, respectively.

3. A transformer oil quality comprehensive monitoring method according to claim 2, characterized in that: The abnormal area is the area enclosed by the oil quality curve above the preset standard curve, which is obtained by fitting based on several historical normal oil quality parameters; wherein the normal oil quality parameters are oil quality parameters within a preset range.

4. A transformer oil quality comprehensive monitoring method according to claim 1, characterized in that: The training process of the oil quality assessment model is as follows: Several sets of historical oil quality feature vectors and corresponding warning level labels are obtained from historical data. The historical oil quality feature vectors are divided into a training dataset and a validation dataset. The LSTM-Attention model is trained using the training dataset and the validation dataset to obtain an oil quality assessment model.

5. A transformer oil quality comprehensive monitoring method according to claim 4, characterized in that: The LSTM-Attention model includes an input layer, an LSTM layer, an Attention layer, and an output layer; The input layer is used to receive the oil quality feature vector; The LSTM layer includes a first layer and a second layer; the first layer is used to extract short-term dynamic features of the oil quality feature vector with a span of less than N time steps; the second layer is used to extract long-term dynamic features of the oil quality feature vector with a span of more than 3N time steps; wherein N is a positive integer; The Attention layer is used to determine the weight coefficient of the time step; The output layer performs weighted fusion on the short-term dynamic features and long-term dynamic features extracted by the LSTM layer based on the time step weight coefficient determined by the Attention layer to generate a comprehensive feature vector. The comprehensive feature vector is mapped to a preset transformer oil quality warning level classification space according to the comprehensive feature vector, and finally the warning level of the transformer oil quality is output.

6. A transformer oil quality comprehensive monitoring method according to claim 5, characterized in that: The weight coefficient of the time step is obtained by the following methods, including: By formula α t =softmax[u T ×tanh(W×h t +b)] calculate the time step weight coefficient α t ; Where t is the time step, softmax is the activation function, tanh is the hyperbolic tangent activation function, h t is the hidden state at time step t, W is the weight matrix, b is the bias vector, u is the attention vector, and T is the transpose of the attention vector.

7. A transformer oil quality comprehensive monitoring system, operating based on a transformer oil quality comprehensive monitoring method according to any one of claims 1 to 6, characterized in that: It includes data acquisition module, data processing module, oil quality assessment module and blockchain trusted management module; The data acquisition module is used to collect the refractive index, temperature and oil flow pressure of the transformer oil in real time; The data processing module is used to process and analyze the detection data; The oil quality assessment module is used to provide early warning of oil quality; The blockchain trusted management module is used to perform trusted management of detection data.

8. A transformer oil quality comprehensive monitoring system according to claim 7, characterized in that: The data upload process of the blockchain trusted management module is that the edge device uploads the monitored data to the side chain and generates a unique identifier. The smart contract in the main chain records the unique identifier and the corresponding timestamp, and uses a dynamic consensus mechanism to verify and store data.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for comprehensive monitoring of transformer oil quality according to any one of claims 1 to 6 is implemented.

Images