A diagnosis system and method based on transformer oil state multi-parameter monitoring

Abstract

The application discloses a diagnosis system and method based on transformer oil state multi-parameter monitoring, which mainly utilizes a transformer oil state multi-parameter sensor array to monitor the following multi-parameters of the original signals of the transformer oil in the transformer in real time: oil level, oil temperature, characteristic gas concentration and air pressure, and sends the processed data to an edge computing terminal in a wireless or wired manner; the edge computing terminal performs real-time diagnosis and early warning on the sudden fault type and severity of the transformer based on a diagnosis algorithm engine, so that the precise prediction of the sudden fault state of the transformer is realized.

Description

Technical Field

[0001] This invention relates to the field of transformer fault diagnosis technology, and in particular to a diagnostic system and method based on multi-parameter monitoring of transformer oil condition. Background Technology

[0002] With the integration of numerous new energy sources into the power system, power electronic devices such as inverters and frequency converters introduce significant high-order harmonic components. These harmonics can cause transformer temperatures to rise and increase the probability of discharge faults. As core equipment in the power system, the stable operation of transformers is crucial for ensuring grid security. External faults such as surface flashover, foreign object discharge, and bushing rupture can often be quickly diagnosed through direct observation. Whether it's the main transformer or the distribution transformer, when internal discharge or overheating faults occur, the transformer oil inside will decompose to generate specific fault gases, such as hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), and acetylene (C2H2). H2 is the most prevalent gas in transformer faults. Unlike other hydrocarbon gases, H2 has relatively low solubility and diffusion coefficient in transformer oil. When it is generated, it floats in the oil and forms tiny bubbles, accumulating in large quantities above the oil surface, affecting the safe operation of the equipment. Currently, all main transformers are equipped with oil chromatographs before leaving the factory. However, the chromatographs have long testing cycles and lack the ability to monitor the gas production status of transformer oil in real time, making it impossible to predict sudden and transient faults. Furthermore, distribution transformers are relatively inexpensive, and currently only mechanical float-type oil level gauges are used to observe the internal oil level; other detection methods are lacking, resulting in a lack of monitoring data on the transformer's condition. Previous research using Pd alloys as the sensitive material to develop gas sensors suffers from high costs, hydrogen adsorption poisoning, and severe environmental drift, making it difficult to accurately measure the insulation status and thus unable to accurately identify sudden transformer faults.

[0003] The information disclosed in the background section is only intended to enhance the understanding of the background of the present invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] To address the shortcomings or defects of the existing technology, a diagnostic system and method based on multi-parameter monitoring of transformer oil condition is provided. From the perspective of reducing condition monitoring costs and improving the timeliness of transformer sudden fault diagnosis, the system increases the monitoring dimensions of transformer oil condition parameters, combines a real-time diagnostic algorithm for transformer sudden faults, and embeds it into an edge computing terminal, thereby achieving accurate prediction of transformer sudden fault conditions.

[0005] The objective of this invention is achieved through the following technical solutions.

[0006] A diagnostic system based on multi-parameter monitoring of transformer oil condition includes:

[0007] The outer casing is a hollow cylindrical shell installed in the transformer oil tank, and the outer casing is provided with a glass window;

[0008] A transformer oil condition multi-parameter sensor array is used to monitor the following parameters of the transformer oil inside the transformer in real time: oil level, oil temperature, characteristic gas concentration and gas pressure. The transformer oil condition multi-parameter sensor array is integrated into the housing and the digital display signal of the transformer oil condition multi-parameter sensor array is observed through the glass window.

[0009] The data collection and transmission board is connected to the transformer oil status multi-parameter sensor array to collect and process the raw signals of oil level, oil temperature, characteristic gas concentration and gas pressure, and transmit the processed data to the edge computing terminal wirelessly or via wired means.

[0010] The edge computing terminal uses a diagnostic algorithm engine to perform real-time diagnosis and early warning of the types and severity of sudden faults in transformers.

[0011] In the diagnostic system based on multi-parameter monitoring of transformer oil condition, the multi-parameter sensor array for transformer oil condition includes:

[0012] The oil level sensor measures the oil level by detecting changes in the magnetic field through a Hall sensor that works in conjunction with a magnet on the float linkage.

[0013] Oil temperature sensor, which uses thermocouple or electronic temperature sensor to measure the oil temperature of transformer oil;

[0014] Gas sensor used for selective detection of hydrogen and selective detection of acetylene;

[0015] A pressure sensor measures the air pressure above the transformer oil surface using the piezoresistive or capacitive effect principle.

[0016] In the diagnostic system based on multi-parameter monitoring of transformer oil condition, the tin oxide type includes any of the following: SnO2, Sn3O4, Au-SnO2, Pt-SnO2.

[0017] In the diagnostic system based on multi-parameter monitoring of transformer oil condition, the outer shell is installed at the opening on the upper surface of the transformer oil tank cover, and the bottom is connected to the transformer oil tank cover by threads and flange studs. A rubber gasket is provided between the threads and the opening of the transformer oil tank cover to maintain a seal. There are four flange studs, which are evenly distributed on the bottom circumference of the outer shell at 90-degree intervals.

[0018] The aforementioned diagnostic system based on multi-parameter monitoring of transformer oil condition:

[0019] A float is provided directly below the bottom opening of the outer casing. The float is connected by a float linkage rod, which is fixed to the inner wall of the outer casing by a snap fastener.

[0020] As the oil level changes, the float drives the connecting rod to move up and down. A two-color indicator buoy is installed on the float connecting rod, which is level with the glass window. The two-color indicator buoy integrates a digital display screen to display the signals of the transformer oil status multi-parameter sensor array in real time.

[0021] In the aforementioned diagnostic system based on multi-parameter monitoring of transformer oil condition, the data collection and transmission board includes:

[0022] The signal acquisition module periodically monitors the transformer oil inside the transformer and outputs data, which is then connected to the system's main control module.

[0023] The system's main control module connects to the signal acquisition module to perform analog-to-digital conversion and signal modulation on the data output from the signal acquisition front-end, and then analyzes and processes the modulated signal.

[0024] The data transmission module is used to transmit the multi-parameter data processed by the system main control module to the edge computing terminal. The data transmission methods include wired and wireless methods, as well as Timer I and Timer II.

[0025] The control method for the data collection and transmission board includes:

[0026] After the main control module of the system is powered on, it enters the stop mode, all sensors are powered off, and only the real-time clock timer I is retained;

[0027] When Timer I reaches the set time T0, it wakes up the main control module, first powering the gas sensor to preheat it, and then starting Timer II;

[0028] After Timer II reaches the set time T1, it wakes up the main control module again to power the oil level, oil temperature, and air pressure sensors and collect data.

[0029] After the data is packaged and sent, if the host computer verifies it successfully, it returns to stop mode; if the verification fails, it re-collects and re-sends the data.

[0030] In the aforementioned diagnostic system based on multi-parameter monitoring of transformer oil condition, the diagnostic algorithm engine performs the following steps:

[0031] Obtain a multi-parameter monitoring matrix Φ=[H,T,C,P] consisting of oil level H, oil temperature T, characteristic gas concentration C, and gas pressure P;

[0032] Based on the oil level signal X1 and the oil temperature signal X2, the true oil level is analyzed by the function H=f1(X1,X2), where f1 is obtained by least squares fitting.

[0033] Based on the selective detection of hydrogen and acetylene gas signals X3 from the gas sensor, the oil temperature signal X2 from the oil temperature sensor, and the gas pressure signal X4 from the gas pressure sensor, the characteristic gas concentration C is analyzed by the function C=f3(X2,X3,X4), and environmental drift compensation is achieved by a deep convolutional neural network model based on the Siamese network framework.

[0034] Based on multi-parameter time series data of oil level, oil temperature, characteristic gas concentration, and gas pressure, statistical features are extracted and normalized. The maximum information coefficient (MIC) method is used to calculate the correlation between each feature and the fault type, and feature weights are generated.

[0035] A machine learning model is constructed. The machine learning model takes a weighted feature vector as input and outputs a sudden failure state matrix GZ=[GZ1,GZ2,GZ3,GZ4,GZ5]. In the sudden failure state matrix, each component from the first to the fifth item represents the classification and severity assessment results of local overheating, discharge, oil leakage, moisture, and oil quality deterioration, respectively.

[0036] In the aforementioned diagnostic system based on multi-parameter monitoring of transformer oil condition, the function H=f1(X1,X2) is implemented using the following quadratic polynomial surface model:

[0037] ,

[0038] The actual oil level coefficients a, b, c, d, e, f were obtained by fitting the data from laboratory isothermal calibration using the least squares method.

[0039] In the diagnostic system based on multi-parameter monitoring of transformer oil condition, the solution method for the function C=f3(X2,X3,X4) includes:

[0040] A DCNN model based on Siamese network is constructed, using the time series of gas signals X3 (selectively detecting hydrogen and acetylene) output by the gas sensor and the standard concentration time series of the corresponding characteristic gases as dual inputs;

[0041] Negative cosine similarity is used as the pre-training loss function to enable the model to learn the amplitude similarity between gas signals and concentration changes.

[0042] After pre-training, a small number of network parameters are fine-tuned to adapt to different oil temperature and air pressure environments;

[0043] By mapping network parameters to a low-dimensional space and performing linear transformation, cross-domain generalization of characteristic gas concentrations and environmental drift compensation can be achieved.

[0044] In the diagnostic system based on multi-parameter monitoring of transformer oil condition, the machine learning model adopts the random forest algorithm. The input is a feature vector weighted by the maximum information coefficient method (MIC), and the output is a five-dimensional fault severity vector. Each dimension corresponds to local overheating, discharge, oil leakage, moisture, and oil quality deterioration, respectively, and the value range is [0,1]. When any dimension exceeds the preset threshold, a fault warning is triggered.

[0045] Furthermore, this invention also discloses a diagnostic method based on multi-parameter monitoring of transformer oil condition, which includes:

[0046] S100: Collects raw signals of oil level, oil temperature, characteristic gas concentration and gas pressure through a transformer oil condition multi-parameter sensor array;

[0047] S200: In the edge computing terminal, the actual oil level H is analyzed based on the oil temperature compensation oil level signal;

[0048] S300: Performs environmental drift compensation for gas sensors and analyzes the concentration C of characteristic gases. The environmental drift compensation for gas sensors using a deep convolutional neural network based on Siamese network includes two stages: pre-training and fine-tuning. In the pre-training stage, the time series of gas signals that selectively detect hydrogen and acetylene output by the gas sensor and the standard concentration time series of the corresponding characteristic gases are used as dual inputs to optimize the negative cosine similarity loss function. In the fine-tuning stage, only some network parameters are adjusted in the field environment to achieve low-dimensional linear transformation of the parameter space.

[0049] S400: Time series normalization and feature extraction of multiple parameters such as oil level, oil temperature, characteristic gas concentration and gas pressure;

[0050] S500: The Maximum Information Coefficient (MIC) method is used to calculate the correlation between each feature and the fault type, and feature weights are generated.

[0051] S600: Input weighted features into the machine learning model and output the sudden failure state matrix GZ to achieve the classification and severity assessment of sudden failures.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention achieves real-time diagnosis of overheating, discharge type and severity inside the transformer by monitoring multiple parameters such as oil level, oil temperature, characteristic gas concentration and gas pressure of the transformer oil inside the transformer, combined with a diagnostic algorithm engine.

[0053] The description provided is merely an overview of the technical solution of this invention. In order to make the technical means of this invention clearer and more understandable, so that those skilled in the art can implement it according to the contents of the specification, and to make the described and other objects, features and advantages of this invention more obvious and understandable, specific embodiments of this invention are described below. Attached Figure Description

[0054] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0055] In the attached diagram:

[0056] Figure 1 This is a topology diagram of the diagnostic system in one embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the structure of a transformer oil status multi-parameter sensor array in one embodiment of the present invention;

[0058] Figure 3 This is a topology diagram of the data collection and transmitter board functions in one embodiment of the present invention;

[0059] Figure 4 This is a flowchart of the control algorithm for data collection and low-power operation of the transmitter board in one embodiment of the present invention;

[0060] Figure 5 This is a timing diagram of the control for data collection and low-power operation of the transmitter board in one embodiment of the present invention;

[0061] Figure 6 This is a schematic diagram of the power conditioning module in the data collection and transmission board in one embodiment of the present invention;

[0062] Figure 7 This is a schematic diagram of a transformer oil level analysis algorithm in one embodiment of the present invention;

[0063] Figure 8 This is a schematic diagram of a pre-training method based on a Siamese network framework in one embodiment of the present invention;

[0064] Figure 9 This is a schematic diagram of a gas recognition model based on a DCNN structure in one embodiment of the present invention;

[0065] Figure 10 This is a schematic diagram of a domain generalization framework optimization method driven by the Koopmam operator in one embodiment of the present invention;

[0066] Figure 11 This is a schematic diagram of a real-time transformer fault diagnosis algorithm in one embodiment of the present invention;

[0067] Figure 12 This is a schematic diagram illustrating the data interaction principle of an edge computing terminal in one embodiment of the present invention.

[0068] Figure 13 This is a schematic diagram illustrating the parameter changes during a discharge fault in one embodiment of the present invention;

[0069] Figure 14 This is a schematic diagram of the nonlinear correction output of hydrogen concentration and oil level signal in one embodiment of the present invention;

[0070] Figure descriptions: 1-Pressure relief port, 2-Top pressure relief valve, 3-Antenna, 4-Control unit, 5-Pressure sensor, 6-Gas sensor, 7-Strong magnet, 8-Glass window, 9-Red indicator buoy, 10-Blue indicator buoy, 11-Housing, 12-Float linkage rod, 13-Oil level gauge connection thread or flange stud, 14-Locking screw, 15-Power supply board, 16-Main control board, 17-Temperature and humidity sensor, 18-Sensor board, 19-Hall sensor, 20-K-type thermocouple, 21-Transformer tank cover opening, 22-Transformer tank cover, 23-Float.

[0071] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0072] The following will refer to the appendix. Figures 1 to 14 Specific embodiments of the invention will be described in more detail below. While specific embodiments of the invention are shown in the accompanying drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0073] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0074] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments, and the accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0075] To better understand, such as Figures 1 to 12 As shown, a diagnostic system based on multi-parameter monitoring of transformer oil condition includes:

[0076] The outer casing is a hollow cylindrical shell installed in the transformer oil tank, and the outer casing is provided with a glass window;

[0077] A transformer oil condition multi-parameter sensor array is used to monitor the following parameters of the transformer oil inside the transformer in real time: oil level, oil temperature, characteristic gas concentration and gas pressure. The transformer oil condition multi-parameter sensor array is integrated into the housing and the digital display signal of the transformer oil condition multi-parameter sensor array is observed through the glass window.

[0078] The data collection and transmission board is connected to the transformer oil status multi-parameter sensor array to collect and process the raw signals of oil level, oil temperature, characteristic gas concentration and gas pressure, and transmit the processed data to the edge computing terminal wirelessly or via wired means.

[0079] The edge computing terminal uses a diagnostic algorithm engine to perform real-time diagnosis and early warning of the types and severity of sudden faults in transformers.

[0080] In a preferred embodiment of the diagnostic system based on multi-parameter monitoring of transformer oil condition, the multi-parameter sensor array for transformer oil condition includes:

[0081] The oil level sensor measures the oil level by detecting changes in the magnetic field through a Hall sensor that works in conjunction with a magnet on the float linkage.

[0082] Oil temperature sensor, which uses thermocouple or electronic temperature sensor to measure the oil temperature of transformer oil;

[0083] The gas sensor uses Sn3O4 material for selective detection of hydrogen and Au-SnO2 semiconductor material for selective detection of acetylene.

[0084] A pressure sensor measures the air pressure above the transformer oil surface using the piezoresistive or capacitive effect principle.

[0085] In a preferred embodiment of the diagnostic system based on multi-parameter monitoring of transformer oil condition, such as... Figure 2 As shown, the outer shell is installed at the opening on the upper surface of the transformer tank cover, and the bottom is connected to the transformer tank cover by threads and flange studs. A rubber gasket is provided between the threads and the opening of the transformer tank cover to maintain a seal. There are four flange studs, which are evenly distributed on the bottom circumference of the outer shell at 90-degree intervals.

[0086] In a preferred embodiment of the diagnostic system based on multi-parameter monitoring of transformer oil condition, a float is provided directly below the bottom opening of the housing. The float is connected by a float linkage rod, which is fixed to the inner wall of the housing by a snap fastener. As the oil level changes, the float drives the linkage rod to move up and down. A bicolor indicator float, positioned flush with the glass window, is installed on the float linkage rod. The bicolor indicator float integrates a digital display screen to display the signals from the multi-parameter monitoring sensor of transformer oil condition in real time.

[0087] For example, the data collection and transmission board is installed on the dual-color indicator buoy. The data collection and transmission board is connected to a battery compartment and its battery. The data collection and transmission board includes a power supply board, a main control board, and a sensor board. A signal antenna is fixedly connected to the upper surface of the power supply board. An electromagnetic compatibility module is integrated on the upper surface of the power supply board. A power conditioning module is integrated on the lower surface of the power supply board. A control unit is fixed on the main control board. A sensor module is fixed on the sensor board. The main control board is connected to the power supply board via a connector. The sensor board is connected to the main control board via a connector below the main control board.

[0088] In another, more preferred embodiment, such as Figure 3 As shown, the data collection and transmission board includes:

[0089] The signal acquisition module periodically monitors the transformer oil inside the transformer and outputs data, which is then connected to the system's main control module.

[0090] The system's main control module connects to the signal acquisition module to perform analog-to-digital conversion and signal modulation on the data output from the signal acquisition front-end, and then analyzes and processes the modulated signal.

[0091] The data transmission module is used to transmit the multi-parameter data processed by the system main control module to the edge computing terminal. The data transmission methods include wired and wireless methods, as well as Timer I and Timer II.

[0092] See further Figure 4 The data collection and transmission board control method includes:

[0093] After the main control module of the system is powered on, it enters the stop mode, all sensors are powered off, and only the real-time clock timer I is retained;

[0094] When Timer I reaches the set time T0, it wakes up the main control module, first powering the gas sensor to preheat it, and then starting Timer II;

[0095] After Timer II reaches the set time T1, it wakes up the main control module again to power the oil level, oil temperature, and air pressure sensors and collect data.

[0096] After the data is packaged and sent, if the host computer verifies it successfully, it returns to stop mode; if the verification fails, it re-collects and re-sends the data.

[0097] In another embodiment, such as Figure 11 As shown, in a preferred embodiment of the diagnostic system based on multi-parameter monitoring of transformer oil condition, the diagnostic algorithm engine performs the following steps:

[0098] Obtain a multi-parameter monitoring matrix Φ=[H,T,C,P] consisting of oil level H, oil temperature T, characteristic gas concentration C, and gas pressure P;

[0099] Based on the oil level signal X1 and the oil temperature signal X2, the true oil level is analyzed by the function H=f1(X1,X2), where f1 is obtained by least squares fitting.

[0100] Based on the selective detection of hydrogen and acetylene gas signals X3 from the gas sensor, the oil temperature signal X2 from the oil temperature sensor, and the gas pressure signal X4 from the gas pressure sensor, the characteristic gas concentration C is analyzed by the function C=f3(X2,X3,X4), and environmental drift compensation is achieved by a deep convolutional neural network model based on the Siamese network framework.

[0101] The deep convolutional neural network model is based on the Siamese network framework, and its core consists of an encoder and a fine-tuning network. The encoder adopts a weight-sharing dual-path architecture. Path A receives the gas signals X3 output by the gas sensor, which selectively detect hydrogen and acetylene. Path B receives the standard concentration time series C of the characteristic gas corresponding to the gas signals. refBoth pathways have the same network structure, including an input layer, two convolutional layers and a pooling layer, as well as a flattening layer and an output layer; the input layer has 128 nodes to receive one-dimensional signal sequences (for example, the gas signal X3 of pathway A is processed into 128 segments, and correspondingly, the standard concentration time series C of pathway B is processed into 128 segments). ref It was also processed into 128 segments, which can be understood as segments corresponding to time series (a time series is a sequence formed by taking time as the independent variable and the corresponding dependent variable as the value). The first convolutional layer uses 32 5×1 convolutional kernels, and the second convolutional layer uses 64 3×1 convolutional kernels, with a stride of 1 for both layers and the ReLU activation function. The size of the max pooling layer after the convolutional layers is 2. The final output is a 128-dimensional feature vector.

[0102] The fine-tuned network employs a three-layer fully connected structure, with 64, 32, and 1 neurons in each layer, respectively. The first two layers use the ReLU activation function, while the output layer uses a linear activation function to directly generate predicted gas concentration values.

[0103] Model training consists of two stages: pre-training and fine-tuning. Pre-training is conducted in a laboratory environment, using a large amount of gas sensor response data, such as the gas signal X3 and the standard concentration time series of the corresponding feature gas, to train the encoder. The negative cosine similarity between the two feature vectors is calculated as the loss function.

[0104] The fine-tuning phase was conducted in the field environment, using a small amount of data about (X2,X3,X4) and the actual characteristic gas concentration time series C. true To adapt to specific scenarios, this stage freezes encoder parameters, connects to a fine-tuning network, obtains predicted concentration values ​​through forward propagation, and jointly optimizes the mean squared error loss and the Koopman operator reconstruction loss. The optimization employs a hierarchical optimization strategy, where dynamic learning rate adjustment and gradient pruning are implemented during the pre-training phase. The learning rate is reduced when the validation loss fails to decrease for 10 consecutive rounds, and training is terminated after 20 consecutive rounds without improvement. The Koopman operator maps a nonlinear system to an approximately linear space and has a theoretical guarantee of convergence.

[0105] In another embodiment, based on multi-parameter time series data of oil level, oil temperature, characteristic gas concentration, and gas pressure, statistical features are extracted and normalized. The maximum information coefficient (MIC) method is used to calculate the correlation between each feature and the fault type, generating feature weights. This includes the following steps:

[0106] Statistical features were extracted and normalized from the multi-parameter time series of oil level, oil temperature, characteristic gas concentration, and gas pressure.

[0107] Meanwhile, based on historical fault records, a sudden fault state matrix GZ=[GZ1,GZ2,GZ3,GZ4,GZ5] is constructed, and the normalized feature parameters and fault states are used as variables and discretely distributed in a two-dimensional space.

[0108] The joint probability distribution is calculated using a dynamic grid partitioning method, and the nonlinear correlation between each feature and each type of fault is quantified using the maximum information coefficient (MIC) method. The closer the MIC result is to 1, the stronger the positive correlation; the closer it is to -1, the stronger the negative correlation; and the closer it is to 0, the more independent they are.

[0109] For each type of fault, the MIC results of the fault and each feature are sorted in descending order, and the corresponding weight coefficients are obtained by normalization.

[0110] Multiply the original feature vector with the calculated weight coefficients to generate a feature vector weighted by the Maximum Information Coefficient (MIC) method, ensuring that highly relevant features have a higher weight in fault diagnosis.

[0111] A machine learning model is constructed. The machine learning model takes a weighted feature vector as input and outputs a sudden failure state matrix GZ=[GZ1,GZ2,GZ3,GZ4,GZ5]. In the sudden failure state matrix, each component from the first to the fifth item represents the classification and severity assessment results of local overheating, discharge, oil leakage, moisture, and oil quality deterioration, respectively.

[0112] In a preferred embodiment of the diagnostic system based on multi-parameter monitoring of transformer oil condition, the function H=f1(X1,X2) is implemented using the following quadratic polynomial surface model:

[0113]

[0114] The actual oil level coefficients a, b, c, d, e, f were obtained by fitting the data from laboratory isothermal calibration using the least squares method.

[0115] like Figure 8 As shown, in a preferred embodiment of the diagnostic system based on multi-parameter monitoring of transformer oil condition, the method for solving the function C=f3(X2,X3,X4) includes:

[0116] A DCNN model based on Siamese network is constructed, with the gas signals X3 of selective detection of hydrogen and selective detection of acetylene output by the gas sensor and the standard concentration time series of the corresponding characteristic gases as dual inputs;

[0117] Negative cosine similarity is used as the pre-training loss function to enable the model to learn the amplitude similarity between gas signals and concentration changes;

[0118] After pre-training, a small number of network parameters are fine-tuned to adapt to different oil temperature and air pressure environments;

[0119] By mapping network parameters to a low-dimensional space and performing linear transformation, cross-domain generalization of characteristic gas concentrations and environmental drift compensation can be achieved.

[0120] In another embodiment, during a 30-minute test in which the hydrogen concentration increased from 0 ppm to 100 ppm, the sensor resistance gradually decreased from 115.82 kΩ to 16.85 kΩ. The twin DCNN model segmented the continuous signal into 128 steps of gas sensor outputs, selectively detecting hydrogen and selectively detecting acetylene, as gas signals X3, and used the corresponding standard concentration time series Cref as dual inputs.

[0121] During the pre-training phase, the model extracts features through a weight-sharing convolutional network, encoding the resistance and concentration time series into 128-dimensional feature vectors. The negative cosine similarity loss function is used to calculate feature similarity, which increases from an initial 0.15 to 0.82 after training, enabling the model to learn the correlation between decreasing resistance and increasing concentration. When the resistance drops to 45.3 kΩ, the model accurately identifies the corresponding concentration as 62 ppm. Regarding environmental adaptability, when the temperature rises from 25℃ to 35℃, the model fine-tunes the parameters of the last two layers to compensate. The Koopman operator projects the network parameters into a 16-dimensional subspace, eliminating the influence of temperature drift through a low-dimensional linear transformation of the parameter space, maintaining a stable resistance-concentration mapping. Finally, at a resistance of 16.85 kΩ, it accurately outputs a concentration value of 100 ppm, with an error controlled within ±2 ppm.

[0122] In a preferred embodiment of the diagnostic system based on multi-parameter monitoring of transformer oil condition, the machine learning model employs the random forest algorithm, and the input is a feature vector X weighted by the Maximum Information Coefficient (MIC) method. i ·w i The output is a five-dimensional fault severity vector, with each dimension corresponding to local overheating, discharge, oil leakage, moisture, and oil quality deterioration, and the value range is [0,1]. When any dimension exceeds the preset threshold, a fault warning is triggered.

[0123] In another embodiment, the present invention also discloses a diagnostic method based on multi-parameter monitoring of transformer oil condition, comprising:

[0124] S100: Collects raw signals of oil level, oil temperature, characteristic gas concentration and gas pressure through a transformer oil condition multi-parameter sensor array;

[0125] S200: In the edge computing terminal, the actual oil level H is analyzed based on the oil temperature compensation oil level signal;

[0126] S300: Environmental drift compensation is performed on the gas sensor to analyze the characteristic gas concentration C. Specifically, a deep convolutional neural network based on Siamese network is used to compensate for environmental drift of the gas sensor. This step also includes two stages: pre-training and fine-tuning. In the pre-training stage, the time series of gas signals that selectively detect hydrogen and acetylene output by the gas sensor and the standard concentration time series of the corresponding characteristic gases are used as dual inputs to optimize the negative cosine similarity loss function. In the fine-tuning stage, only some network parameters are adjusted in the field environment, and the low-dimensional linear transformation of the parameter space is achieved by combining the Koopman operator.

[0127] S400: Time series normalization and feature extraction of multiple parameters such as oil level, oil temperature, characteristic gas concentration and gas pressure;

[0128] S500: The Maximum Information Coefficient (MIC) method is used to calculate the correlation between each feature and the fault type, and feature weights are generated.

[0129] S600: Input weighted features into the machine learning model and output the sudden failure state matrix GZ to achieve the classification and severity assessment of sudden failures.

[0130] In another embodiment, a diagnostic system based on multi-parameter monitoring of transformer oil status includes an array of multi-parameter sensors for monitoring the casing, oil level, oil temperature, characteristic gas concentration, and gas pressure; a data collection and transmission board; a real-time transformer fault diagnosis algorithm engine; and an edge computing terminal, wherein:

[0131] The outer casing 11 is identical to that of the tubular oil level gauge, cylindrical in shape and hollow inside. A glass window 8 is provided on the casing for observing the digital display signals of the multi-parameter sensor array. A top pressure relief valve 2 is fixedly connected to the top, and a pressure relief port 1 is fixedly connected to the top of the pressure relief valve. The casing can be installed on the transformer tank cover opening 21, and the bottom is connected to the transformer tank cover 22 via threads and flange studs 13. A rubber gasket is provided between the threads and the transformer tank cover opening. There are four flange studs, evenly distributed at 90-degree intervals on the bottom circumference of the casing. Further, a data acquisition and transmission board is fixed inside the casing, and the data acquisition and transmission board is connected to a battery compartment and its battery. The data acquisition and transmission board includes a power supply board 15, a main control board 16, and a sensor board 18.

[0132] Furthermore, a signal antenna 3 is fixedly connected to the upper surface of the power supply board, an electromagnetic compatibility module is integrated on the upper surface of the power supply board, a power conditioning module is integrated on the lower surface of the power supply board, a control unit 4 is fixedly mounted on the main control board, and a sensor module is fixedly mounted on the sensor board. The power supply board is connected to the main control board via a connector at its lower end, and the sensor board is connected to the main control board via a connector at its lower end.

[0133] Furthermore, a float 23 is connected to the lower part of the outer casing via a float linkage rod 12. Two-color indicator floats 9 and 10 are fixedly installed on the upper part of the linkage rod, positioned flush with the glass window. A digital display screen is integrated on each two-color indicator float to display the signals from the transformer oil status multi-parameter monitoring sensor in real time.

[0134] The transformer oil condition multi-parameter sensor array includes functions for monitoring oil level, oil temperature, characteristic gas concentration, and gas pressure. Further, a cylindrical strong magnet 7 is fixedly installed at the top of the connecting rod, with the axis of the cylindrical strong magnet aligned with the Z-axis of the bottom Hall sensor 19. Using the Hall ranging principle, a relationship is established between the oil level signal X1 output by the Hall sensor and the actual oil level H, enabling real-time measurement of the oil level. Preferably, a capacitive sensor can also be used for oil level monitoring. Further, a thermocouple is used for oil temperature measurement, or an electronic temperature sensor can be selected, establishing a relationship between the output oil temperature signal X2 and the transformer oil temperature T.

[0135] For example, the gas sensor, based on the redox principle, employs gas sensors based on gas-sensitive semiconductor materials such as Sn3O4 and Au-SnO2 to selectively detect the characteristic gases hydrogen (H2) and acetylene (C2H2) of transformer oil, respectively, establishing a relationship between the gas signal X3 and the concentration C of the characteristic gas. In this case, the change in gas signal X3 essentially reflects the resistance change of the Sn3O4 material and the Au-SnO2 gas-sensitive semiconductor material.

[0136] Furthermore, for air pressure monitoring, sensors with piezoresistive or capacitive effects are selected, and the relationship between the air pressure signal X4 output by the sensor and the air pressure P is established.

[0137] In another embodiment, the data aggregation and transmission board includes: a signal acquisition module, a system main control module, a power conditioning module, a wireless data transmission module, an electromagnetic compatibility module, and a power supply module, wherein:

[0138] The signal acquisition module periodically monitors the transformer oil inside the transformer and outputs relevant data. It includes an oil level sensor, an oil temperature sensor, an air pressure sensor, a temperature and humidity sensor, and a gas sensor, which are connected to the system's main control module.

[0139] The system's main control module (MCU) performs analog-to-digital conversion and signal modulation on the data output from the signal acquisition front-end, and analyzes and processes the modulated signal.

[0140] When any abnormality is detected in the signal corresponding to any factor, the system main control module issues a command, and the signal acquisition module shortens the interval or continuously acquires multiple parameters.

[0141] The data transmission module is used by the system main control module to transmit multiple parameters to the edge computing terminal. The data transmission methods include wired and wireless methods. The wired method includes RS485, and the wireless method includes LoRa and 4G.

[0142] Furthermore, the data aggregation and transmission board has a low-power operation mode. Two timers are set, and the operation is performed as follows:

[0143] Step S201: The system main control module is powered on, all sensors are not working, and timer I starts timing T0;

[0144] Step S202: Timer I wakes up the system's main control module MCU, the gas sensor is powered on and preheated, and Timer II starts timing T1;

[0145] Step S203: Timer II wakes up the MCU, the oil level, oil temperature, and air pressure sensors are powered on and initialized, test data is generated and sent;

[0146] Step S204: The host computer verifies the data status. If the verification bit is incorrect, proceed to step S203; otherwise, proceed to step S201.

[0147] In addition, the power conditioning module, which includes a lithium battery and power conditioning circuitry, can convert and process the input voltage to provide voltage and current that meet the requirements of electronic devices and control overall power consumption.

[0148] Electromagnetic compatibility (EMC) modules are used to suppress electromagnetic interference emissions and enhance anti-interference capabilities, ensuring that the device operates stably and reliably under electromagnetic interference such as pulse groups, surges, and static electricity.

[0149] The power supply module includes a solar panel and a battery.

[0150] The diagnostic algorithm engine executes three steps: a sudden fault state matrix, a transformer oil state multi-parameter monitoring matrix, and mapping relationship parsing. Specifically:

[0151] Conducting transformer sudden fault simulation experiments, constructing the sudden fault state matrix GZ, and recording the multi-parameter monitoring matrix of transformer oil status. =[H,T,C,P], establish GZ and The mapping relationship F between them is:

[0152]

[0153] The transformer sudden fault status matrix GZ includes, but is not limited to: local overheating, discharge, oil leakage, moisture, and oil deterioration.

[0154] Multi-parameter monitoring signal matrix There are relationships between the elements, specifically:

[0155] The oil level H is affected by both the oil level signal X1 and the oil temperature signal X2, that is:

[0156] H = f1(X1, X2)

[0157] The oil temperature signal X2 can directly reflect the oil temperature T, that is:

[0158] T=f2(X2)

[0159] The gas concentration C is jointly affected by the oil temperature signal X2, the gas signals X3 (selectively detecting hydrogen and acetylene) output by the gas sensor, and the gas pressure signal X4 output by the pressure sensor, that is:

[0160] C = f3(X2, X3, X4)

[0161] The air pressure signal X4 can directly reflect the air pressure P, that is:

[0162] P=f4(X4)

[0163] To achieve real-time and accurate diagnosis of sudden transformer faults, it is first necessary to solve for f1 and f3, and then solve for F. Specifically:

[0164] For example, signal preprocessing and physical quantity analysis are as follows:

[0165] Since the transformer oil level signal is stable and is mainly affected by temperature, this invention uses a fitting method to solve for f1, specifically:

[0166] By setting up a constant temperature test environment, the oil level signal X1 and oil temperature signal X2 are obtained at different oil temperatures T. With X2 kept constant, the least squares method is used to fit the relationship between the oil level H and the oil level signal X1, obtaining multiple fitted curves at different oil temperatures T, thereby solving for f1.

[0167] Furthermore, the fitting function can include, but is not limited to, linear functions, multinomial functions, exponential functions, and power functions.

[0168] Furthermore, a quadratic surface fitting method is adopted to directly fit the relationship between oil level H and oil level signal X1 and oil temperature signal X2. The fitting function includes, but is not limited to, plane fitting, polynomial surface fitting, Gaussian function fitting, and exponential function fitting.

[0169] For example, the accurate analysis of gas concentration based on Siamese networks and domain generalization is as follows:

[0170] like Figure 9 As shown, a deep convolutional neural network (DCNN) model based on the Siamese network framework is established to achieve accurate detection of the concentrations of characteristic gases H2 and C2H2 and accurate solution of f3.

[0171] Multiple convolutional layers are introduced to extract features from the raw signals of oil temperature, characteristic gas concentration, and gas pressure. Fully connected layers are then introduced to concatenate and aggregate the multiple features.

[0172] Under fixed oil temperature and air pressure, gas sensor data is input into a pre-trained DCNN. The fully connected layers in the DCNN summarize the signal features extracted by the convolutional layers and predict the concentration of the characteristic gas at the current moment. By fine-tuning the network in the DCNN that plays the role of feature summarization, the DCNN can automatically compensate for gas sensor drift caused by ambient oil temperature and air pressure, and accurately predict the concentration of the characteristic gas in the environment.

[0173] Furthermore, the gas signal and its corresponding concentration time series (e.g., the aforementioned 128 segments) are used as the dual inputs to the encoder. This encoder consists of a DCNN and a multilayer sensing architecture, and can be any other neural network used for feature extraction from the data. To enable the encoder to extract the similarity in amplitude variation between the gas sensor response and gas concentration changes, the following negative cosine similarity is used as the loss function in the pre-training process:

[0174] ,

[0175] ,

[0176] ,

[0177] In the formula, C represents the concentration corresponding to the gas signal X3; b() represents the encoder transformation; Z2 represents the result obtained from the encoder transformation; h(f()) represents the multilayer perceptron transformation; P1 represents the result obtained from the multilayer perceptron transformation; and D represents the negative cosine similarity of P1 relative to Z2.

[0178] For example, this invention implements the gas sensor as an array to monitor as comprehensively as possible. During training, the time series of the response signals in the gas sensor array are compared with the corresponding concentration time series of the characteristic gas to minimize the loss function L. For example, the loss function L is:

[0179]

[0180] Among them, P2 and P1 are both results obtained from the multilayer perceptron transformation, but P2 corresponds to another path; Z1 and Z2 are both results obtained from the encoder transformation, but Z1 corresponds to another path. As mentioned above, the encoder adopts a weight-sharing dual-path architecture. Path A receives the gas signals X3 output by the gas sensor, which selectively detect hydrogen and acetylene. Path B receives the standard concentration time series C of the characteristic gas corresponding to the gas signal.ref Both pathways have the same network structure, which includes an input layer, two convolutional layers and a pooling layer, as well as a flattening layer and an output layer.

[0181] For example, Z1=b(C ref Z2 = b(X3), or vice versa. Used to force Z2 to be aligned with P1 in order to preserve the valid features of the corresponding input (e.g., X3); Used to force Z1 and P2 to be in the same direction, so as to preserve the corresponding input (e.g., C). ref The standard features of the gas signal X3 are shown in the formula above. The reason why the loss function L is as described is to align Z1 and Z2 of the same sample in different modalities after projection. Thus, its contribution to the existing technology is: to extract the features of the two input modalities through a weight-sharing encoder, and then to map the features to the contrast space through a projection head, calculate the similarity loss of the cross-modalities, and thus, through contrastive learning, make the representation of the gas signal X3 consistent with the standard concentration sequence C. ref The representations are aligned to improve the accuracy of gas concentration estimation.

[0182] It should be noted that the total loss value for one training iteration is the average of the loss values ​​of the data samples, and the minimum value that this loss function can reach is -1. During training, gradient stopping plays a crucial role. This is achieved by ignoring the backpropagation gradient calculations generated during the computation of P1 and Z2 when optimizing the encoder using gradient descent.

[0183] The pre-trained DCNN has locked most of the trainable network parameters. When the environment of the gas sensor changes, such as when the oil temperature rises or the gas pressure increases, only a small amount of data is needed to fine-tune a small number of network parameters of the DCNN to achieve environmental drift compensation of the gas sensor under different environmental conditions.

[0184] Specifically, a learnable function is constructed and optimized in the following way:

[0185] ,

[0186] In the formula: This represents the initial network parameters of the DCNN under humidity condition r; This represents the learnable function that has been constructed; The network parameters represent the learnable function h; This indicates that it has been fine-tuned to adapt to ambient humidity. The parameters of the DCNN network; similar; Indicates adaptation to environmental humidity The DCNN network parameters are generated using a learnable function h to adapt to ambient humidity. The parameters of the DCNN network.

[0187] Furthermore, the Koopman space is used to simplify the learning process of fine-tuning the model parameters, mapping the high-dimensional network parameter space to the low-dimensional Koopman space. The Koopman operator performs a linear transformation on the low-dimensional representation of the network parameters in the Koopman space, mapping the low-dimensional representation generated in the Koopman space back to the high-dimensional network parameter space, i.e.:

[0188]

[0189] In the formula: z represents the network parameters in low dimension via encoder The resulting representation in Koopman space, for In Koopman space, K represents the Koopman operator. Indicates through decoder The obtained DCNN network parameters.

[0190] To ensure that the above process accurately captures the drift characteristics of the gas sensor signal as humidity changes, the process needs to be jointly optimized across multiple processes, including: a finely tuned DCNN network, the encoder and decoder corresponding to the mapping from parameter space to Koopman space and the inverse mapping, and the Koopman operator. The following loss function needs to be minimized:

[0191]

[0192] In the formula: Yi represents the DMC concentration label in the i-th humidity domain; g represents the finely tuned DCNN network; This represents an encoder that maps the parameters of a DCNN fine-tunable network to the Koopman space; This represents a decoder that maps low-dimensional parameter representations in the Koopman space to high-dimensional network parameter spaces. Indicates the loss of the internal triplet; This represents the integral consistency loss; Indicates the reconstruction loss; This represents the loss of dynamic consistency. This indicates a loss of consistency.

[0193] like Figure 10 As shown, the fine-tunable DCNN network, encoder, decoder, and Koopman transformation matrix in the domain generalization framework are described. The joint optimization process can be expressed as the following formula:

[0194]

[0195] In the formula: , , This represents an adjustable weight coefficient, used to ensure that no single factor dominates during training.

[0196] The deep convolutional neural network (DCNN) model based on the twin network framework established by this invention can correct the signal drift of gas sensors caused by oil temperature and gas pressure, solve for f3, and achieve accurate monitoring of characteristic gas concentrations H2 and C2H2.

[0197] Compared with the pre-trained model proposed in this invention, the neural network trained by supervised learning needs to be retrained with data from the field environment, which increases the amount of training data required by several times.

[0198] For example, the intelligent diagnosis process for sudden faults based on feature weighting is as follows:

[0199] Regarding the aforementioned GZ and Solving for the mapping relationship F between them involves the following steps:

[0200] Step S1: Based on the transformer oil multi-parameter sensor array, acquire four types of features of the transformer: oil level H, oil temperature T, gas temperature C, and gas pressure P. Within one sampling period, combine the multi-feature parameter data into a time series. For example, the oil level time series H=[h1,h2,...h... n The gas time series C = [c1, c2, ... c2] n ] , where n is a positive integer representing the number of samples.

[0201] Step S2: Normalize the different time series data to eliminate dimensional differences. The normalization method can be min-max scaling, standard deviation standardization, maximum absolute value scaling, or robust scaling.

[0202] Step S3: Extract statistical features from the time series, including mean, variance, and extreme values. Based on historical fault records, construct a sudden fault state matrix GZ from the samples, including five types of faults: overheating, discharge, oil leakage, moisture, and oil quality deterioration, and convert them into quantifiable labels.

[0203] Furthermore, the sudden failure state matrix In, each element The value range is [0,1], j=1~5. The closer the value is to 1, the more severe the sudden failure.

[0204] Step S4: Using the Maximum Information Coefficient (MIC) method, the feature parameter X and the sudden fault GZ are discretely distributed in a two-dimensional space and represented using a scatter plot. The joint probability is calculated by observing the occurrence of different scatter points in the two-dimensional space. Specifically:

[0205]

[0206] Where a and b represent the number of grid cells in two directions of the two-dimensional space, respectively, and the value of B depends on the data volume raised to the power of 0.6. The MIC value ranges from -1 to 1. When two variables are independent, they are unrelated; the closer to 1, the stronger the positive correlation; the closer to -1, the stronger the negative correlation.

[0207] Step S5: Analyze different sudden faults in the transformer (GZ) j The corresponding characteristic parameter X i The relevance coefficients calculated using the maximum information coefficient method are sorted from largest to smallest. After normalization, the weight coefficients w are calculated. i .

[0208] Step S600: Establish a machine learning-based transformer sudden fault diagnosis model, with input X. i ·w i The output is GZ j This involves solving for the mapping relationship F. The machine learning algorithms used are not limited to neural networks, decision trees, or support vector machines.

[0209] By using the method proposed in this invention, transformer oil condition parameters are monitored in multiple dimensions, and combined with a diagnostic algorithm engine, abnormal operating conditions of the transformer can be diagnosed in real time, thereby achieving early warning of sudden faults.

[0210] The edge computing terminal can receive multi-parameter monitoring signals from data aggregation and remote transmission from the transmitter board, and has the functions of data display, storage, analysis, and uploading. Furthermore, this terminal can analyze oil level, oil temperature, characteristic gas concentration, and gas pressure parameters, and achieve transformer oil condition assessment by executing a series of algorithms proposed in this invention. Based on the transformer oil condition parameters, sudden transformer fault diagnosis can be achieved.

[0211] In one embodiment, the transformer oil status multi-parameter sensor array includes a signal acquisition front-end section that periodically monitors the transformer oil inside the distribution transformer and outputs relevant data, which is then connected to the system's main control module. It includes an oil level sensor, an oil temperature sensor, a pressure sensor, a temperature and humidity sensor, and a gas sensor, as shown in the table below.

[0212]

[0213] The gas sensor exemplified in this invention is based on a SnO2 thin film prepared by reactive magnetron sputtering. It exhibits high sensitivity, high selectivity, low detection limit, and excellent stability, making it suitable for real-time monitoring of hydrogen concentration in transformer oil. This sensor is prepared under conditions of an oxygen-argon ratio of 0.7, a substrate temperature of 100°C, and a reaction vacuum of 1 Pa. Its SnO2 thin film is dense and exhibits optimal hydrogen-sensing performance, achieving a linear response within the range of 0–2000 ppm with a nonlinear error of only 2.69%. It boasts a sensitivity as high as 3.55% / ppm and a detection limit as low as 12 ppb, far superior to existing commercial sensors.

[0214] Furthermore, even in the presence of typical interfering gases in various transformer oils (such as CO, CO2, CH4, C2H2, C2H4, and C2H6), the sensor's response value to hydrogen remains seven times that of the interfering gases, demonstrating excellent gas selectivity and meeting the technical requirements of DL / T1432.2-2016 for hydrogen monitoring devices in transformer oil. Since the gas sensor is based on a SnO2 thin film, it can be mass-produced at low cost using MEMS technology, providing a reliable and economical hydrogen sensing solution for distribution transformer condition monitoring.

[0215] It should be noted that commercially available gas sensors can also be used, instead of the SnO2 thin film-based sensors mentioned above, if cost is not a concern.

[0216] For example:

[0217] The system's main control module includes a main control chip and a code download interface, which can meet the sampling requirements of a single gas sensor, as well as the requirements for SPI communication, I2C and serial communication. It also has a built-in ADC analog-to-digital converter to receive, process and output sensor data.

[0218] The power conditioning module, including a lithium battery and power conditioning circuitry, converts and processes the input voltage to provide voltage and current that meet the requirements of the electronic device and controls overall power consumption. The power management module includes a 12V to 6V circuit, a 6V to 3.3V circuit, a 6V to 5V intermittent circuit, a 6V to 3.3V intermittent circuit, a 6V to 1.2V intermittent circuit, and a filter circuit composed of several capacitors. The system main control module uses a continuous 3.3V power supply, while the oil level sensor, oil temperature sensor, air pressure sensor, and temperature and humidity sensor use an intermittent 3.3V power supply, and the wireless signal transmission module uses an intermittent 5V power supply.

[0219] The wireless data transmission module is provided. The system main control module receives and analyzes the modulated signal, and then transmits the signal wirelessly to the backend server via the wireless communication module. The wireless data transmission method can be either 4G or LoRa. When using 4G communication, either TCP or UDP protocol can be selected to transmit the signal to the backend server. When using LoRa communication, a wireless signal receiving device is required. The wireless signal receiving device includes a main control module, a wireless signal receiving module, and a data reporting module, which is used to convert the LoRa wireless signal into a 485 signal and then transmit it to the backend server.

[0220] Furthermore, the data is packaged using the standard TLV format, consisting of a frame header, packet length, destination ID, local ID, data type, specific data, and checksum. The packet length is the total length of the packet, including the frame header and checksum. The remote ID is the destination ID, the local ID is the local ID, identifying different sensors, and the checksum is a cumulative sum.

[0221] The electromagnetic compatibility module suppresses electromagnetic interference emissions and enhances anti-interference capabilities, ensuring that the device operates stably and reliably under electromagnetic interference such as pulse groups, surges, and static electricity.

[0222] Optionally, it further includes a solar power module, which comprises a solar panel and a lead-acid battery;

[0223] Optionally, it further includes a voltage regulator board, the output of which is connected to the electromagnetic compatibility module; the voltage regulator board integrates a battery capacity monitoring module and a comparator module, which can monitor the input voltage and the remaining battery capacity in real time; when the solar power supply voltage is lower than the threshold, the comparator outputs a digital signal, which, after passing through an inverter, controls the start of the 220V DC power supply and the shutdown of the solar power supply, thereby realizing the automatic switching of the power supply mode;

[0224] Optionally, the system main control module compares the real-time sensor signals horizontally and, if it finds any abnormality in the signal corresponding to any factor, sends a collection command to the signal collection module to shorten the interval or continuously collect multiple monitoring signals of the distribution transformer.

[0225] Optionally, the system main control module can receive a configuration signal from the backend server within a short period of time after reporting through the wireless data transmission module, in order to configure the reporting cycle and sensor preheating time.

[0226] Furthermore, the data aggregation and transmission board features an intelligent low-power operation mode, achieving precise power consumption control by setting the RTC clock. Specific operation steps include:

[0227] Step S201: After the main control module of the system is powered on and initialized, it immediately enters the stop mode; all sensor power supply circuits are cut off, and only the RTC real-time clock is retained. The set duration is T0, which can be set to several minutes to several hours according to actual needs. At this time, the overall power consumption of the system is reduced to the microampere level.

[0228] Step S202: When Timer I reaches T0, an interrupt signal is generated to wake up the MCU; the MCU first controls the power management module to supply power to the gas sensor to enter the preheating state to ensure measurement accuracy; at the same time, the RTC timing duration is set to T1, usually 2-5 minutes to meet the time required for the gas sensor to stabilize; during this period, the MCU re-enters the stop mode and only maintains the necessary basic operation;

[0229] Step S203: After the RTC real-time clock reaches T1, it wakes up the MCU again and controls the power management module to power on and initialize the oil level sensor, oil temperature sensor, air pressure sensor, etc. After each sensor completes its self-test, it quickly collects data and performs real-time calculation and packaging through the data processing algorithm. The data packet is sent to the remote monitoring center through the wireless transmission module. After completion, the power supply to all sensors is immediately cut off.

[0230] Step S204: After transmission is completed, the system waits for the host computer to send a data reception confirmation signal. If the received check bit is correct, it indicates that the data transmission was successful. The system immediately clears the buffer, resets all flag bits, and returns to step S201 to enter the next sleep cycle. If the check bit is incorrect or no confirmation is received after timeout, step S203 is re-executed to re-acquire and retransmit the data to ensure data reliability. At the same time, the system records the number of retransmissions. When the number of consecutive failures exceeds the set threshold, an abnormal alarm mechanism is triggered.

[0231] Furthermore, the steps of the real-time diagnosis method for sudden transformer faults are consistent with those described in Part III of the invention, and will not be repeated here.

[0232] In one embodiment, the function f1 is implemented using the following quadratic polynomial surface model:

[0233] ,

[0234] The coefficients a, b, c, d, e, and f were obtained by fitting the data from laboratory constant temperature calibration using the least squares method. An exemplary set of coefficients is: a=0.05, b=-0.01, c=0.1, d=2.5, e=1.8, f=15.3.

[0235] As another specific embodiment, the encoder of the Siamese-based DCNN model has the following specific network structure:

[0236] Input layer: Receives sensor signal segments X3 with a length of 128.

[0237] Convolutional layer 1: The number of convolutional kernels is 32, the kernel size is 5, the stride is 1, and the activation function is ReLU.

[0238] Pooling layer 1: Max pooling, pooling size is 2.

[0239] Convolutional layer 2: The number of convolutional kernels is 64, the kernel size is 3, the stride is 1, and the activation function is ReLU.

[0240] Pooling layer 2: Max pooling, with a pooling size of 2.

[0241] Fully connected layer: The feature map is flattened and then fed into a 128-dimensional fully connected layer to output the final feature vector.

[0242] Pre-training parameters: The Adam optimizer is used, with an initial learning rate of 1e-4, a batch size of 32, and 100 training epochs. The loss function is negative cosine similarity.

[0243] As another specific embodiment, in step 4.3, the machine learning-based fault diagnosis function GZ=F(H,T,C,P) is implemented using the random forest algorithm in this embodiment.

[0244] The input features are the feature vectors weighted by the MIC weights. Taking the localized overheating fault GZ1 as an example, an example of its weighting coefficients is as follows:

[0245] ,

[0246] The random forest has 100 trees and a maximum depth of 10. The model output is a 5-dimensional vector [GZ1, GZ2, GZ3, GZ4, GZ5], representing the probability values ​​of the severity of five faults: "local overheating," "discharge," "oil leakage," "moisture absorption," and "oil degradation," with values ​​ranging from [0,1]. When any GZ... j When the value exceeds the preset threshold, a warning for the corresponding fault is triggered.

[0247] In another embodiment, the edge computing terminal is implemented using an industrial-grade processor based on the ARM Cortex-A53 core. Its specific workflow is as follows:

[0248] Data packets from the data aggregation and transmission board are received via TCP / IP protocol through a 4G module or LoRa gateway.

[0249] The built-in parsing algorithm is invoked to extract the original X1, X2, X3, and X4 signal values ​​from the data packet;

[0250] Calculate H, T, C, P, then perform fault diagnosis to obtain the sudden fault state matrix GZ;

[0251] The diagnostic results (GZ) and raw data were uploaded to the cloud platform via the MQTT protocol.

[0252] The connected touchscreen displays real-time information on oil level, temperature, gas concentration, and fault alarms.

[0253] Based on the constructed transformer fault diagnosis model, this invention establishes a comprehensive intelligent diagnostic logic table for multiple transformer fault types, as shown below. The parameter thresholds for various fault types are calibrated according to the historical data distribution characteristics output by the diagnostic model. The calibration process is closely related to the model's training and validation. Specifically, in the model establishment phase, multi-dimensional sensor data is collected and combined with known fault cases to train a fault classification model using machine learning algorithms. Subsequently, based on the model's analysis of massive amounts of operational data, the typical numerical ranges and variation patterns of each parameter under different fault types are statistically summarized, thereby completing the threshold calibration of the comparison table. The calibration results directly reflect the fault characteristic patterns learned by the model and can serve as an intuitive interpretation of the model's diagnostic results and a rapid reference for on-site operation and maintenance.

[0254]

[0255] The model uses dynamic baseline comparison, which does not rely on fixed thresholds. Instead, it compares real-time data with dynamic normal baselines learned from historical data. The algorithm not only looks at numerical values, but also pays attention to the trend, rate of change and acceleration of parameters. For example, "oil level drop" refers to the historical normal oil level of the transformer under similar load and environment.

[0256] By employing multi-parameter coordination and weight allocation, this invention dynamically assigns weights to the contributions of different parameters under different fault modes. For example, in the early stage of partial discharge, gas characteristics have the highest weight; in overheating faults, the persistence of oil temperature has the highest weight.

[0257] The algorithm employs fault mode matching and incorporates a fault mode library trained from historical fault case data. It matches real-time data streams against these patterns to provide the most probable diagnostic result.

[0258] In another embodiment, to verify the effectiveness and reliability of this diagnostic system, a transformer oil multi-parameter monitoring system was developed and has been installed and deployed on-site in 10kV distribution transformers of the power grid. It has successfully provided early warning of potential faults and provided solid data support for accurate maintenance based on transformer oil status.

[0259] This system integrates a multi-parameter sensor array for oil level, oil temperature, hydrogen, gas pressure, and humidity, enabling comprehensive sensing of the transformer oil condition. Deployed in transformer #1 (model: S13-M-400 / 10), this system successfully captured and diagnosed a typical developing discharge fault, such as... Figure 13 As shown in the diagram, during this fault, the sensor detected a real-time increase in oil temperature from approximately 65°C to 75°C, along with slight increases in air pressure and humidity. The gas sensor resistance plummeted from a reference value of 1000kΩ to 90kΩ. The data aggregation board converted the raw signals into digital values, packaged them, and sent them to the edge computing terminal.

[0260] The edge terminal first performs environmental compensation calculations. Based on a pre-trained DCNN model and considering the current air pressure conditions, it performs nonlinear correction on the gas sensor output resistance and oil level signal. The correction result is as follows: Figure 14 As shown.

[0261] The weighted feature vectors were input into the random forest model to calculate the sudden fault state matrix GZ. The discharge fault dimension GZ2 was significantly higher than other fault types and exceeded the warning threshold. The system immediately generated a spark discharge fault warning, uploaded the diagnostic results to the cloud platform via a 4G wireless module, and verified the results using a partial discharge detector. The measured partial discharge quantity confirmed the floating potential discharge at the high-voltage winding end, ultimately confirming a potential floating potential discharge inside the transformer. Timely shutdown was implemented, preventing a possible equipment damage and power outage.

[0262] This embodiment demonstrates through actual field operation data that the diagnostic system based on multi-parameter monitoring of transformer oil can achieve early, accurate, and real-time diagnosis of sudden transformer faults, upgrading the operation and maintenance mode from "periodic maintenance" to "condition-based maintenance," significantly improving the operational reliability and economic efficiency of transformers.

[0263] This invention collects five parameters: oil level, oil temperature, gas pressure, temperature and humidity, and H2 content. H2 gas is a primary indicator gas for transformer faults and can directly reflect the internal operating status of distribution transformers. Compared to the aforementioned monitoring devices, this invention also adds a temperature and humidity sensor to ensure effective monitoring of the humidity conditions inside the transformer. This invention integrates random forest multi-classification for intelligent analysis of monitoring data, providing preliminary judgment of transformer fault types and more detailed information support for subsequent maintenance and repair work. The oil level monitoring device of this invention can also be applied to main transformers. For example, the sensor selection, installation location, and signal processing algorithm may need to be fine-tuned according to the specific characteristics of the main transformer to ensure optimal monitoring effect and reliability.

[0264] Furthermore, the present invention integrates a multi-parameter sensor array into a standard oil level gauge structure to achieve low-cost, no-modification deployment, and compatibility with new and old equipment. The twin network + Koopman + MIC + random forest achieves small-sample learning, cross-domain generalization, adaptive feature weighting, and multi-fault classification. The time-sharing wake-up low-power mechanism supports battery / solar power supply, is suitable for wide-area deployment in the field, and real-time edge-side diagnosis shortens the response time and enables very early warning of sudden faults.

[0265] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the aforementioned specific details.

[0266] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the invention to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A diagnostic system based on multi-parameter monitoring of transformer oil condition, characterized in that, It includes: The outer casing is a hollow cylindrical shell installed in the transformer oil tank, and the outer casing is provided with a glass window; A transformer oil condition multi-parameter sensor array is used to monitor the following parameters of the transformer oil inside the transformer in real time: oil level, oil temperature, characteristic gas concentration and gas pressure. The transformer oil condition multi-parameter sensor array is integrated into the housing and the digital display signal of the transformer oil condition multi-parameter sensor array is observed through the glass window. The data collection and transmission board is connected to the transformer oil status multi-parameter sensor array to collect and process the raw signals of oil level, oil temperature, characteristic gas concentration and gas pressure, and transmit the processed data to the edge computing terminal wirelessly or via wired means. The edge computing terminal uses a diagnostic algorithm engine to perform real-time diagnosis and early warning of the types and severity of sudden faults in transformers.

2. The diagnostic system based on multi-parameter monitoring of transformer oil condition as described in claim 1, characterized in that, Preferably, the transformer oil condition multi-parameter sensor array includes: The oil level sensor measures the oil level by detecting changes in the magnetic field through a Hall sensor that works in conjunction with a magnet on the float linkage. Oil temperature sensor, which uses thermocouple or electronic temperature sensor to measure the oil temperature of transformer oil; Gas sensor used for selective detection of hydrogen and selective detection of acetylene; A pressure sensor measures the air pressure above the transformer oil surface using the piezoresistive or capacitive effect principle.

3. The diagnostic system based on multi-parameter monitoring of transformer oil condition as described in claim 1, characterized in that, The outer casing is installed at the opening on the upper surface of the transformer tank cover. The bottom is connected to the transformer tank cover by threads and flange studs. A rubber gasket is provided between the threads and the opening of the transformer tank cover to maintain a seal. There are four flange studs, which are evenly distributed on the bottom circumference of the outer casing at 90-degree intervals.

4. The diagnostic system based on multi-parameter monitoring of transformer oil condition as described in claim 1, characterized in that, A float is provided directly below the bottom opening of the outer casing. The float is connected by a float linkage rod, which is fixed to the inner wall of the outer casing by a snap fastener. As the oil level changes, the float drives the connecting rod to move up and down. A two-color indicator buoy is installed on the float connecting rod, which is level with the glass window. The two-color indicator buoy integrates a digital display screen to display the signals of the transformer oil status multi-parameter sensor array in real time.

5. The diagnostic system based on multi-parameter monitoring of transformer oil condition as described in claim 1, characterized in that, The data collection and transmission board includes: The signal acquisition module periodically monitors the transformer oil inside the transformer and outputs data, which is then connected to the system's main control module. The system's main control module connects to the signal acquisition module to perform analog-to-digital conversion and signal modulation on the data output from the signal acquisition front-end, and then analyzes and processes the modulated signal. The data transmission module is used to transmit the multi-parameter data processed by the system main control module to the edge computing terminal. The data transmission methods include wired and wireless methods, as well as Timer I and Timer II. The control method for the data collection and transmission board includes: After the main control module of the system is powered on, it enters the stop mode, all sensors are powered off, and only the real-time clock timer I is retained; When Timer I reaches the set time T0, it wakes up the main control module, first powering the gas sensor to preheat it, and then starting Timer II; After Timer II reaches the set time T1, it wakes up the main control module again to power the oil level, oil temperature, and air pressure sensors and collect data. After the data is packaged and sent, if the host computer verifies it successfully, it returns to stop mode; if the verification fails, it re-collects and re-sends the data.

6. The diagnostic system based on multi-parameter monitoring of transformer oil condition as described in claim 1, characterized in that, The diagnostic algorithm engine is used to perform the following steps: Obtain a multi-parameter monitoring matrix Φ=[H,T,C,P] consisting of oil level H, oil temperature T, characteristic gas concentration C, and gas pressure P; Based on the oil level signal X1 and the oil temperature signal X2, the true oil level is analyzed by the function H=f1(X1,X2), where f1 is obtained by least squares fitting. Based on the selective detection of hydrogen and acetylene gas signals X3 from the gas sensor, the oil temperature signal X2 from the oil temperature sensor, and the gas pressure signal X4 from the gas pressure sensor, the characteristic gas concentration C is analyzed by the function C=f3(X2,X3,X4), and environmental drift compensation is achieved by a deep convolutional neural network model based on the Siamese network framework. Based on multi-parameter time series data of oil level, oil temperature, characteristic gas concentration, and gas pressure, statistical features are extracted and normalized. The maximum information coefficient (MIC) method is used to calculate the correlation between each feature and the fault type, and feature weights are generated. A machine learning model is constructed. The machine learning model takes a weighted feature vector as input and outputs a sudden failure state matrix GZ=[GZ1,GZ2,GZ3,GZ4,GZ5]. In the sudden failure state matrix, each component from the first to the fifth item represents the classification and severity assessment results of local overheating, discharge, oil leakage, moisture, and oil quality deterioration, respectively.

7. The diagnostic system based on multi-parameter monitoring of transformer oil condition as described in claim 6, characterized in that, The function H=f1(X1,X2) is implemented using the following quadratic polynomial surface model: ， The actual oil level coefficients a, b, c, d, e, f were obtained by fitting the data from laboratory isothermal calibration using the least squares method.

8. The diagnostic system based on multi-parameter monitoring of transformer oil condition as described in claim 6, characterized in that, The method for solving the function C=f3(X2,X3,X4) includes: A DCNN model based on Siamese network is constructed, using the time series of gas signals X3 (selectively detecting hydrogen and acetylene) output by the gas sensor and the standard concentration time series of the corresponding characteristic gases as dual inputs; Negative cosine similarity is used as the pre-training loss function to enable the model to learn the amplitude similarity between gas signals and concentration changes. After pre-training, a small number of network parameters are fine-tuned to adapt to different oil temperature and air pressure environments; By mapping network parameters to a low-dimensional space and performing linear transformation, cross-domain generalization of characteristic gas concentrations and environmental drift compensation can be achieved.

9. The diagnostic system based on multi-parameter monitoring of transformer oil condition as described in claim 6, characterized in that, The machine learning model uses the random forest algorithm. The input is a feature vector weighted by the maximum information coefficient method (MIC). The output is a five-dimensional fault severity vector. Each dimension corresponds to local overheating, discharge, oil leakage, moisture, and oil quality deterioration, with a value range of [0,1]. When any dimension exceeds a preset threshold, a fault warning is triggered.

10. A diagnostic method based on multi-parameter monitoring of transformer oil condition, characterized in that, It includes: S100: Collects raw signals of oil level, oil temperature, characteristic gas concentration and gas pressure through a transformer oil condition multi-parameter sensor array; S200: In the edge computing terminal, the actual oil level H is analyzed based on the oil temperature compensation oil level signal; S300: Performs environmental drift compensation for gas sensors and analyzes the concentration C of characteristic gases. The environmental drift compensation for gas sensors using a deep convolutional neural network based on Siamese network includes two stages: pre-training and fine-tuning. In the pre-training stage, the time series of gas signals that selectively detect hydrogen and acetylene output by the gas sensor and the standard concentration time series of the corresponding characteristic gases are used as dual inputs to optimize the negative cosine similarity loss function. In the fine-tuning stage, only some network parameters are adjusted in the field environment to achieve low-dimensional linear transformation of the parameter space. S400: Time series normalization and feature extraction of multiple parameters such as oil level, oil temperature, characteristic gas concentration and gas pressure; S500: The Maximum Information Coefficient (MIC) method is used to calculate the correlation between each feature and the fault type, and feature weights are generated. S600: Input weighted features into the machine learning model and output the sudden failure state matrix GZ to achieve the classification and severity assessment of sudden failures.

Images