Transformer DC magnetic bias monitoring and diagnosis system based on multi-source sensing

Abstract

The invention relates to a transformer direct current magnetic bias monitoring and diagnosis system based on multi-source sensing, which belongs to the technical field of intelligent monitoring and diagnosis of power equipment and comprises a current sensor, a temperature sensor, a noise sensor and a vibration sensor. The system further comprises a data processing and collecting module, a model reasoning module, a state output module and a communication protocol module. According to the invention, real-time monitoring, early warning and severity accurate classification of the direct-current magnetic bias state of the transformer are realized, and the method has stronger anti-interference capability and adaptability. The system is especially suitable for power transformer DC magnetic bias monitoring and intelligent operation and maintenance influenced by high-voltage DC power transmission, rail transit stray current and geomagnetic storm, is helpful for improving the safety and reliability of power grid operation, and meets the urgent demands of a power system for intelligent fault diagnosis and risk early warning.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and diagnosis technology of power equipment, and in particular relates to a transformer DC bias monitoring and diagnosis system based on multi-source information fusion and deep learning model. It is suitable for smart grid and transformer operation and maintenance scenarios with high requirements for power equipment status perception accuracy and fault early warning capability. Background Technology

[0002] In the field of transformer DC bias monitoring and diagnosis, accurate sensing and timely response to the degree of DC bias are crucial for ensuring the safe operation of equipment, directly affecting the transformer's thermal stability, mechanical lifespan, and power quality of the power grid. With the expansion of the power grid and the increasing complexity of the operating environment, the generation of DC bias is influenced by a variety of complex factors. Relying solely on single current monitoring is insufficient to comprehensively reflect the degree of DC bias. This not only leads to incomplete condition assessments and delayed early warnings, but also causes transformer winding overheating, increased vibration, and higher harmonic content, seriously impacting the reliable operation of the power grid.

[0003] In the diagnosis of magnetic bias, traditional methods have weak capabilities in fusing multi-source information and lack in-depth analysis of temporal evolution patterns. This makes it difficult to achieve refined diagnosis from "presence / absence judgment" to "degree classification," resulting in low accuracy and timeliness of early warnings, which cannot support the decision-making needs of intelligent operation and maintenance. Therefore, a system is needed that can fuse multi-source sensor information with deep learning models, addressing these problems through a collaborative mechanism of autonomous feature extraction and dynamic state identification. Summary of the Invention

[0004] To address the limitations of single-source monitoring methods for transformer DC bias, the inability of traditional methods to accurately assess the degree of bias, and the lack of multi-source information fusion and intelligent diagnostic capabilities in existing technologies, this invention proposes a transformer DC bias monitoring and diagnostic system based on a multi-source sensor fusion deep learning model. Through collaborative acquisition of multi-source information, intelligent extraction of temporal features, and graded diagnosis of the degree of bias, this system achieves real-time monitoring, early warning, and accurate assessment of the transformer's DC bias state. The specific technical solution is as follows:

[0005] A transformer DC bias monitoring and diagnostic system based on multi-source sensing includes a current sensor 1, a temperature sensor 2, a noise sensor 3, and a vibration sensor 4. The current sensor 1 is a zero-flux Hall current sensor, capable of measuring the DC current component under AC / DC superposition conditions under electrically isolated conditions. It is installed at the transformer neutral point grounding lead to monitor the DC current component in real time. The system also includes a data processing and acquisition module, a model inference module, a status output module, and a communication protocol module.

[0006] A method for monitoring and diagnosing DC bias magnetism in transformers based on multi-source sensing includes the following steps:

[0007] S1: By deploying multi-source sensors for current, vibration, noise, and temperature signals, the DC bias magnetic characteristic parameters of each sample transformer are measured. The characteristic parameters include: DC current component at the transformer neutral point, transformer vibration signal, noise level, and transformer temperature rise signal.

[0008] S2: Based on the measured characteristic parameters, four types of data samples are obtained, and the DC bias degree of the transformer is classified according to the DC bias risk assessment and defense guidelines of the power grid.

[0009] S3: Construct time series data samples: Cut the measured continuous time series data into segments of fixed length, each segment forming a sample with dimension (L, 4), and assign a level label to it based on the overall situation of the data within the segment.

[0010] S4: Build and train a 1D-CNN-LSTM neural network model. Take the (L, 4) time series samples from step S3 as the input to the model and input them into the trained model in sequence. Finally, output the corresponding state level label.

[0011] Furthermore, in step S1, current sensor 1, model LEM ITN 1000-S, is installed at the neutral point grounding lead of the transformer. It is a DC current sensor based on the Hall effect and zero flux principle, used to monitor the DC current component at the neutral point. Temperature sensor 2, model Omega PR-21 PT100 RTD, is installed in the thermometer holder at the top of the transformer tank, used to monitor the temperature rise of the windings and transformer oil. Noise sensor 3, model GRAS 46AE, is placed outside the transformer tank at a horizontal distance of 1.5-3 meters and a height of 1.2-1.5 meters, used to collect audible noise levels. Vibration sensor 4, model Siemens SITRANS VS300, is symmetrically installed in the upper part of the transformer tank wall, corresponding to the projection area of ​​the core and windings, used to collect mechanical vibration signals caused by DC bias. All sensors are designed to resist electromagnetic interference and are suitable for the strong electromagnetic environment of substations.

[0012] The Hall effect formula for DC current measurement:

[0013]

[0014] in, The voltage signal output by the sensor. The sensitivity of Hall elements is sparse. The control current for the Hall element, The magnetic flux density is perpendicular to the Hall element;

[0015] Ampere's circuital law is the foundation of the zero flux principle. A sensor generates a magnetic field that is equal in magnitude and opposite in direction to the magnetic field generated by the measured current by applying a reverse current.

[0016]

[0017] in, Magnetic flux density For path infinitesimal vectors, The permeability of free space, Let C be the total current flowing through the area enclosed by the closed path C.

[0018] Furthermore, the specific steps for establishing and training the 1D-CNN-LSTM neural network model described in step S4 are as follows:

[0019] S41: Determine the neural network structure: The established 1D-CNN-LSTM neural network model includes an input layer, a 1D convolutional layer, a pooling layer, an LSTM layer, a fully connected layer, and an output layer;

[0020] S42: Determine the input feature vector: The input is time series data composed of multi-source feature parameters;

[0021] S43: Determine the 1D-CNN structure: 1D convolutional layers are used to extract local features, and pooling layers are used for feature dimensionality reduction;

[0022] The output of one-dimensional convolution for extracting features from time series data is:

[0023]

[0024] In the formula, Features extracted; for Activation function; This is the weight matrix; It is a time series; It is the bias vector;

[0025] S44: Determine the LSTM layer structure: LSTM layers are used to capture long-term dependencies in time series;

[0026] The forget gate, input gate, and output gate of an LSTM network all employ activation functions. First, the activation function controls the flow of data. Then, the long-term memory mode updates the data state relatively stably, while the short-term memory mode adapts to changes at each time step. Addition operations are performed between the old and new memories to prevent gradient vanishing. Finally, the current state is output. The specific calculation formula is as follows:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] In the formula, , , , This is the weight matrix; , , , This is the corresponding bias vector; It is the hyperbolic tangent function; Dot product; This is the output from the previous time step; Values ​​representing the degree of retention; The memory state of the previous moment; Add a degree value to the current state; This is an intermediate state; This is the current state; To output the degree value; This is the output at the current moment; This is the output at the current moment.

[0034] Furthermore, the software portion of this invention runs in an embedded processing module. It receives multi-source sensor data, performs time-series data preprocessing, calls a pre-trained 1D-CNN-LSTM model for state reasoning, and outputs early warning information and maintenance suggestions based on the diagnostic results. Together with the hardware module, it constitutes a complete intelligent monitoring and diagnostic system. The system's software program achieves intelligent classification of DC bias through closed-loop diagnostic logic. It adopts a deep learning-based multi-source time-series pattern recognition algorithm, classifies different degree levels according to the "Guidelines for DC Bias Risk Assessment and Defense of Power Grids," and performs targeted state output and early warning. Specifically, it pre-sets five levels of bias and their corresponding feature threshold ranges, collects multi-source time-series data in real time as model input, outputs classification results after judgment by the 1D-CNN-LSTM model, and triggers corresponding early warning and control strategies for different levels.

[0035] Beneficial effects:

[0036] This invention integrates multi-source sensing with deep learning to construct an intelligent monitoring and diagnostic system for DC bias in transformers. Its core advantages are as follows:

[0037] 1. The effective fusion of multi-source information enables comprehensive perception from a single-dimensional to a multi-dimensional state. By collaboratively acquiring and processing neutral point DC current, vibration, noise, and temperature rise signals, the system can deeply capture the comprehensive performance of the transformer under different degrees of magnetic bias. This fundamentally overcomes the limitations and singularity of traditional monitoring methods, significantly improving the comprehensiveness and accuracy of condition assessment, and providing a solid basis for the safe operation and precise maintenance of transformers.

[0038] 2. The combination of deep learning models and traditional diagnostic methods enables automatic feature extraction and intelligent mining of long-term temporal patterns. The system employs a 1D-CNN-LSTM hybrid model, utilizing 1D-CNN to automatically extract local features and spatial correlations from monitoring data, and using an LSTM network to learn the dynamic evolution of DC bias. The two work together to fully mine the spatiotemporal information in the data, giving the diagnostic model stronger generalization ability and higher classification accuracy, enabling real-time monitoring and early warning of the transformer's DC bias state.

[0039] 3. The system possesses real-time response and intelligent decision-making capabilities, automatically generating differentiated operation and maintenance recommendations based on diagnostic results. For different degrees of magnetic bias, the system can quickly provide tiered response strategies, ranging from enhanced monitoring and the installation of current-limiting devices to adjustments in operating modes, significantly improving the timeliness and targeted nature of fault handling. Simultaneously, the system supports online updates and adaptive optimization, continuously adapting to different operating environments and equipment states, effectively enhancing the intelligence level and safety margin of power grid operation. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the overall framework of the present invention.

[0041] Figure 2 This is a front view of the current sensor installed next to the transformer.

[0042] Figure 3 This is a top view of the temperature sensor installed on top of the transformer tank.

[0043] Figure 4 This is a K-direction view showing the noise sensor installed next to the transformer.

[0044] Figure 5 This is a side view of the transformer structure where the vibration sensor is installed in the upper part of the transformer tank wall. Detailed Implementation

[0045] Example 1

[0046] In this monitoring system, multi-source sensors serve as the core of state perception and are deployed at key locations on the transformer to achieve synchronous acquisition of DC bias characteristic parameters. Among them, the DC current measuring device at the transformer neutral point uses a high-precision, high-reliability DC current measuring device based on the Hall effect and zero flux principle. It can measure the DC current component under AC / DC superposition conditions under electrically isolated conditions and is installed at the transformer neutral point grounding lead to monitor the DC current component in real time.

[0047] Temperature sensor 2 is installed in the thermometer holder at the top of the transformer tank to monitor changes in winding and oil temperature; noise sensor 3 is arranged at a horizontal distance of 1.5-3 meters from the outside of the transformer tank to collect audible noise level; vibration sensor 4 is symmetrically installed in the upper middle part of the transformer tank wall, corresponding to the projection area of ​​the transformer core and winding, to capture mechanical vibration signals caused by DC bias magnetism; all sensors are designed to resist electromagnetic interference and adapt to the strong electromagnetic environment of the substation.

[0048] The analog signals output from each sensor are filtered, amplified, and converted from analog to digital by the signal conditioning circuit before being synchronously acquired by the data acquisition unit to form multi-channel time series data. The acquisition system operates continuously at a fixed sampling rate, and the data is transmitted via network to edge computing nodes or cloud servers for subsequent processing and model analysis.

[0049] Among them, the Hall effect formula for DC current measurement is:

[0050]

[0051] The voltage signal output by the sensor. The sensitivity of Hall elements is sparse. The control current for the Hall element, The magnetic flux density is perpendicular to the Hall element.

[0052] Ampere's circuital law is the basis of the zero flux principle. The sensor applies a reverse current so that the magnetic field it generates is equal in magnitude and opposite in direction to the magnetic field generated by the current being measured.

[0053]

[0054] Magnetic flux density For path infinitesimal vectors, The permeability of free space, Let C be the total current flowing through the area enclosed by the closed path C.

[0055] Example 2

[0056] The system of this invention uses an embedded processing unit as its core, which is connected to a multi-channel signal acquisition module, a communication module, and a power management module. The processing unit continuously receives digital signals from various sensors, performs preprocessing and feature extraction, and then inputs them into a pre-trained 1D-CNN-LSTM model for state determination.

[0057] The signal acquisition module employs multi-channel input, connecting to the analog outputs of current, vibration, noise, and temperature sensors. Vibration and noise sensor signals undergo additional bandpass filtering and integral amplification to extract characteristic amplitudes within the effective frequency band; the current sensor output undergoes direct scaling and DC component extraction; and the temperature signal is linearized and normalized. All processed data are time-aligned to form (L,4)-dimensional time-series samples, which are then fed into the neural network model.

[0058] Example 3

[0059] The system software adopts a modular design, including data acquisition, preprocessing, model inference, status output, and communication modules. After power-on, the main program initializes all peripherals and loads the neural network model parameters. A time-series sample of length L=100 is generated every 10 seconds, corresponding to 4 feature channels. The preprocessed sample is input into a 1D-CNN-LSTM model, with the following structure: Input layer: receives (100, 4) samples; 1D-CNN layer: extracts local spatiotemporal features; LSTM layer: learns long-term dependencies, receiving the feature sequence processed by the CNN and learning the rules and dependencies of feature evolution throughout the time period. Output layer: outputs results with five classification labels. Based on the category with the highest output probability, a diagnostic result is generated and output using One-Hot encoding.

[0060] The output of one-dimensional convolution for extracting features from time series data is:

[0061]

[0062] (The formula edited using MathType can be converted to Word's built-in formula insert function. It has been checked and is working correctly on my end; it displays W.) X)

[0063] In the formula, Features extracted; for Activation function; This is the weight matrix; It is a time series; This is the bias vector.

[0064] LSTM networks are an improvement on recurrent neural networks (RNNs). By incorporating gate structures, they effectively address the vanishing and exploding gradient problems inherent in RNNs. The inclusion of cellular memory units in LSTM networks gives them excellent memory capabilities, making them widely used in time series forecasting. The specific calculation formula is as follows:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071] In the formula, , , , This is the weight matrix; , , , This is the corresponding bias vector; It is the hyperbolic tangent function; Dot product; This is the output from the previous time step; Values ​​representing the degree of retention; The memory state of the previous moment; Add a degree value to the current state; This is an intermediate state; This is the current state; To output the degree value; This is the output at the current moment; This is the output at the current moment.

[0072] At the same time, the system triggers corresponding warnings based on the diagnostic level:

[0073] Normal [1,0,0,0,0]: Logs are recorded, no alerts are issued;

[0074] General [0,1,0,0,0]: Increase monitoring frequency;

[0075] Mild [0,0,1,0,0]: Upload alarm, closely monitor the changing trends of various parameters;

[0076] Moderate [0,0,0,1,0]: It is recommended to install relevant current limiting and DC blocking devices;

[0077] Severe [0,0,0,0,1]: Emergency alarm, it is recommended to adjust the location or number of compensation devices;

[0078] A 220kV transformer with a directly grounded neutral point was selected as the test object. Current, vibration, noise, and temperature sensors were deployed according to the requirements of Example 1. The sampling frequency was set to 1kHz, and a time-series sample was formed every 10 seconds. Partial data of system monitoring and diagnosis results were obtained, as shown in the table below.

[0079] Time series samples Neutral point current (A) <![CDATA[Vibration amplitude (mm / s 2 )]]> Noise (dB) Temperature rise (°C) System diagnostic level 1 1.6 0.05 62 2.1 normal 2 4.75 0.12 68 4.3 generally 3 8.1 0.28 75 6.8 Mild 4 15.3 0.51 82 9.9 moderate 5 20.6 0.89 90 12.2 Severe

Claims

1. A transformer DC bias monitoring and diagnostic system based on multi-source sensing, comprising a current sensor (1), a temperature sensor (2), a noise sensor (3), and a vibration sensor (4), characterized in that, The current sensor (1) is a zero-flux Hall current sensor, which can measure the DC current component under AC and DC superposition under electrical isolation conditions. It is installed at the neutral point grounding lead of the transformer to monitor the DC current component in real time. The system also includes: a data processing and acquisition module, a model inference module, a status output module, and a communication protocol module.

2. A method for monitoring and diagnosing DC bias magnetism in transformers based on multi-source sensing, comprising the following steps: S1: By deploying multi-source sensors for current, vibration, noise, and temperature signals, the DC bias magnetic characteristic parameters of each sample transformer are measured. These characteristic parameters include: The DC current component at the transformer neutral point, the transformer vibration signal, the noise level, and the transformer temperature rise signal; S2: Based on the measured characteristic parameters, four types of data samples are obtained, and the DC bias degree of the transformer is classified according to the DC bias risk assessment and defense guidelines of the power grid. S3: Construct time series data samples: Cut the measured continuous time series data into segments of fixed length, each segment forming a sample with dimension (L, 4), and assign a level label to it based on the overall situation of the data within the segment. S4: Build and train a 1D-CNN-LSTM neural network model. Take the (L, 4) time series samples from step S3 as the input to the model and input them into the trained model in sequence. Finally, output the corresponding state level label.

3. The method for monitoring and diagnosing DC bias magnetism of a transformer based on multi-source sensing according to claim 2, characterized in that, In step S1, the current sensor (1) is a LEM ITN 1000-S, installed at the neutral point grounding lead of the transformer. It is a DC current sensor based on the Hall effect and zero flux principle, used to monitor the DC current component at the neutral point; the temperature sensor (2) is an Omega PR-21 PT100 RTD, installed in the thermometer seat at the top of the transformer tank, used to monitor the temperature rise of the winding and transformer oil; the noise sensor (3) is a GRAS 46AE, placed at a horizontal distance of 1.5-3 meters and a height of 1.2-1.5 meters outside the transformer tank, used to collect the audible noise level; the vibration sensor (4) is a Siemens SITRANS VS300, symmetrically installed in the upper part of the transformer tank wall, corresponding to the projection area of ​​the iron core and winding, used to collect the mechanical vibration signal caused by DC bias; all sensors are designed to resist electromagnetic interference and are suitable for the strong electromagnetic environment of substations. The Hall effect formula for DC current measurement: ; in, The voltage signal output by the sensor. The sensitivity of Hall elements is sparse. The control current for the Hall element, The magnetic flux density is perpendicular to the Hall element; Ampere's circuital law is the foundation of the zero flux principle. A sensor applies a reverse current, causing its generated magnetic field to be equal in magnitude and opposite in direction to the magnetic field generated by the measured current. ; in, Magnetic flux density For path infinitesimal vectors, The permeability of free space, Let C be the total current flowing through the area enclosed by the closed path C.

4. The method for monitoring and diagnosing DC bias magnetism of a transformer based on multi-source sensing according to claim 2, characterized in that, The specific steps for establishing and training the 1D-CNN-LSTM neural network model described in step S4 are as follows: S41: Determine the neural network structure: The established 1D-CNN-LSTM neural network model includes an input layer, a 1D convolutional layer, a pooling layer, an LSTM layer, a fully connected layer, and an output layer; S42: Determine the input feature vector: The input is time series data composed of multi-source feature parameters; S43: Determine the 1D-CNN structure: 1D convolutional layers are used to extract local features, and pooling layers are used for feature dimensionality reduction; The output of one-dimensional convolution for extracting features from time series data is: ; In the formula, Features extracted; for Activation function; This is the weight matrix; It is a time series; It is the bias vector; S44: Determine the LSTM layer structure: LSTM layers are used to capture long-term dependencies in time series; The forget gate, input gate, and output gate of an LSTM network all employ activation functions. First, the activation function controls the flow of data. Then, the long-term memory mode updates the data state relatively stably, while the short-term memory mode adapts to changes at each time step. Addition operations are performed between the old and new memories to prevent gradient vanishing. Finally, the current state is output. The specific calculation formula is as follows: ; ; ; ; ; ; In the formula, , , , This is the weight matrix; , , , This is the corresponding bias vector; It is the hyperbolic tangent function; Dot product; This is the output from the previous time step; Values ​​representing the degree of retention; The memory state of the previous moment; Add a degree value to the current state; This is an intermediate state; This is the current state; To output the degree value; This is the output at the current moment; This is the output at the current moment.

5. The method for monitoring and diagnosing DC bias magnetism of a transformer based on multi-source sensing according to claim 2, characterized in that, The system's software runs within an embedded processing module. It receives multi-source sensor data, performs time-series data preprocessing, calls a pre-trained 1D-CNN-LSTM model for state reasoning, and outputs early warning information and maintenance suggestions based on the diagnostic results. Together with the hardware module, it constitutes a complete intelligent monitoring and diagnostic system. The system's software program achieves intelligent classification of DC bias through closed-loop diagnostic logic. It adopts a deep learning-based multi-source time-series pattern recognition algorithm, classifies different degree levels according to the "Guidelines for DC Bias Risk Assessment and Defense of Power Grids," and provides targeted state output and early warning. Specifically, it pre-sets five levels of bias and their corresponding feature threshold ranges, collects multi-source time-series data in real time as model input, outputs classification results after judgment by the 1D-CNN-LSTM model, and triggers corresponding early warning and control strategies for different levels.

Images