Mutual inductor fault diagnosis method and system based on intelligent algorithm, and storage medium

Abstract

The invention relates to the technical field of fault diagnosis, and discloses a mutual inductor fault diagnosis method and system based on an intelligent algorithm, and a storage medium. The method comprises the following steps: acquiring an output signal of a secondary side of a mutual inductor at a preset sampling frequency, performing de-noising processing on the output signal, and converting the output signal into time sequence data; performing time domain feature extraction on the time sequence data to obtain a first data feature and a second data feature of the target physical quantity, extracting a zero-sequence current component and a zero-sequence voltage component from the time sequence data, and calculating a phase difference between the zero-sequence current and the zero-sequence voltage as a zero-sequence feature; whether the transformer is abnormal or not is judged, if yes, a multi-modal input vector is constructed based on the time domain features, the zero sequence features, the environmental parameters, the temperature parameters and historical fault records, the multi-modal input vector is input into a pre-trained multi-task deep learning model, and the deep learning model outputs a fault diagnosis result. According to the invention, the fault diagnosis precision of the mutual inductor is improved.

Description

Technical Field

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a method, system and storage medium for diagnosing a transformer fault based on an intelligent algorithm. Background Art

[0002] With the increasing development and complexity of power systems, the safe and stable operation of power equipment is crucial to their reliability. Instrument transformers, as crucial measuring devices in power systems, are primarily used for current and voltage measurement and protection. They are widely used in relay protection, metering equipment, and automated control systems. The operating status of instrument transformers directly impacts the normal operation of power systems, making the diagnosis and prediction of instrument transformer faults particularly important.

[0003] Existing Chinese patent applications, such as publication number CN117056725A, propose a method and device for fault diagnosis of an optical fiber current transformer. The method includes: obtaining light source monitoring data of the optical fiber current transformer, oversampling the light source monitoring data, and dividing the oversampled light source monitoring data into training samples, test samples, and verification samples, respectively extracting feature vectors for the training samples, test samples, and verification samples, constructing a training set with the feature vectors of the training samples, a test set with the feature vectors of the test samples, and a verification set with the feature vectors of the verification samples, and training a preset fault diagnosis model using the training set and the test set to obtain a trained fault diagnosis model; verifying the accuracy of the trained fault diagnosis model using the verification set, and judging whether the accuracy of the trained fault diagnosis model meets the requirements; if so, using the trained fault diagnosis model to diagnose the optical fiber current transformer. However, the fault diagnosis method of the optical fiber current transformer in the above-mentioned prior art relies on oversampling and feature extraction, which may result in excessively high computational complexity and high sensitivity to noise data.

[0004] There is also a Chinese patent application with publication number CN117171616A, which proposes a voltage transformer fault diagnosis method, device and storage medium based on data enhancement, including: obtaining historical voltage data on the secondary side of the voltage transformer; establishing a transient data model of the voltage transformer and enhancing the fault voltage data; constructing a graph convolutional neural network model and inputting the enhanced fault voltage data into the graph convolutional neural network model for training; using the trained graph convolutional neural network model to perform fault diagnosis on the current voltage transformer voltage data; however, although the voltage transformer fault diagnosis method in the above-mentioned prior art utilizes data enhancement, in the case of large data volume and complex calculations, model training may still face the problems of low efficiency and high computing requirements. Summary of the Invention

[0005] To solve the above technical problems, the present application provides a transformer fault diagnosis method, system and storage medium based on an intelligent algorithm, which are used to improve the accuracy and efficiency of transformer fault diagnosis.

[0006] In a first aspect, the present application provides a transformer fault diagnosis method based on an intelligent algorithm, the method comprising: Step S1: obtaining an output signal of the secondary side of the transformer at a preset sampling frequency, performing denoising on the output signal and converting it into time series data; Step S2: performing time domain feature extraction on the time series data to obtain a first data feature and a second data feature of a target physical quantity, extracting a zero-sequence current component and a zero-sequence voltage component from the time series data, calculating a frequency domain energy distribution and a phase offset using a fast Fourier transform, and calculating a phase difference between the zero-sequence current and the zero-sequence voltage based on the frequency domain energy distribution as a zero-sequence feature; Step S3: Determine whether the transformer is abnormal based on the time domain characteristics. If so, obtain the environmental parameters and temperature parameters of the transformer, and construct a multimodal input vector based on the time domain characteristics, zero-sequence characteristics, environmental parameters, temperature parameters and historical fault records; Step S4: Establish a multi-task deep learning model, input the multimodal input vector into the pre-trained multi-task deep learning model, and the deep learning model outputs a fault diagnosis result, including a classification result of the transformer fault type, an estimated value of the fault location, and the probability of the fault occurring.

[0007] In combination with the first aspect, in a first implementation of the first aspect of the present application, performing denoising processing on the output signal includes: Performing preliminary filtering on the collected output signal based on a bandpass filter, wherein the output signal is current or voltage waveform data, setting a frequency range of the bandpass filter based on the operating frequency of the power system, identifying output signals that are not within the frequency range, determining the presence of noise components, and removing the noise components; The output signal after preliminary processing is further denoised based on an adaptive filter to remove external noise, environmental noise and interference noise within the power system. The denoised signal is converted from the time domain to the frequency domain using wavelet transform, and the key frequency components in the signal are extracted. The noise generated outside the system and the environment is removed through frequency domain analysis.

[0008] In combination with the first aspect, in a second implementation of the first aspect of the present application, obtaining the first data feature and the second data feature of the target physical quantity includes: A latest preset number of time series data are extracted from the time series data, wherein the first data feature is the average value and the first standard deviation of the target physical quantity in the latest time series data, and the second data feature includes the time change rate and the second standard deviation of the target physical quantity in the latest time series data. A regression analysis is performed on the latest preset number of time series data, a linear regression equation is established based on the acquisition time point, the predicted value of the target physical quantity at each acquisition time point is calculated, the optimal slope and the optimal intercept are calculated based on the error function, and the optimal slope is defined as the time change rate of the target physical quantity.

[0009] In combination with the first aspect, in a third implementation of the first aspect of the present application, performing regression analysis on the latest preset amount of time series data includes: An optimal linear regression equation is determined based on the optimal slope and the optimal intercept, an optimal predicted value of a target physical quantity is determined based on the optimal linear regression equation, the sum of squared residuals between actual measured values ​​and optimal predicted values ​​is obtained, the cumulative sum of squares of the differences between each acquisition time point and the average value of the time series is calculated, and the second standard deviation of the time change rate is defined by dividing the sum of squared residuals by the degree of dispersion.

[0010] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, the frequency domain energy distribution and phase offset are calculated using a fast Fourier transform, including: A fast Fourier transform is performed on the zero-sequence current and the zero-sequence voltage to extract the frequency domain signal, the energy of each frequency component is calculated in the frequency domain, the phase of each frequency component is extracted based on the complex result obtained by the fast Fourier transform, and the phase difference between the zero-sequence current and the zero-sequence voltage is calculated as the phase offset.

[0011] In combination with the first aspect, in a fifth implementation of the first aspect of the present application, determining whether an abnormality occurs in the mutual inductor based on the time domain characteristics includes: Set reference values ​​for the first data feature and the second data feature. If the average value and time rate of change of the target physical quantity in the time series data are greater than a preset reference value, and the first standard deviation and the second standard deviation are less than or equal to the preset reference value, it is determined that an abnormality has occurred in the mutual inductor.

[0012] In combination with the first aspect, in a sixth implementation of the first aspect of this application, a multi-task deep learning model is established, including: The multi-task deep learning model includes at least one shared layer and multiple task-specific layers. The multimodal input vector is input into the pre-trained multi-task deep learning model. The multi-task deep learning model extracts common features through the shared layer, uses historical fault records as training data, and combines time domain features, zero-sequence features, environmental parameters and temperature parameters for training. The model parameters are optimized through back propagation, and the performance of the model is tuned through cross-validation. Fault diagnosis is performed separately through multiple task-specific layers, and the fault diagnosis results are output.

[0013] In a second aspect, the present application provides a transformer fault diagnosis system based on an intelligent algorithm, the system comprising: A preprocessing module, which obtains the output signal of the secondary side of the mutual inductor at a preset sampling frequency, performs denoising on the output signal and converts it into time series data; an extraction module, performing time domain feature extraction on the time series data to obtain a first data feature and a second data feature of a target physical quantity, and extracting a zero-sequence current component and a zero-sequence voltage component from the time series data, calculating a frequency domain energy distribution and a phase offset using a fast Fourier transform, and calculating a phase difference between the zero-sequence current and the zero-sequence voltage based on the frequency domain energy distribution as a zero-sequence feature; The fusion module determines whether the transformer is abnormal based on the time domain characteristics. If so, it obtains the environmental parameters and temperature parameters of the transformer and constructs a multimodal input vector based on the time domain characteristics, zero-sequence characteristics, environmental parameters, temperature parameters and historical fault records. The diagnosis module establishes a multi-task deep learning model, inputs the multimodal input vector into the pre-trained multi-task deep learning model, and the deep learning model outputs a fault diagnosis result, including a classification result of the transformer fault type, an estimated value of the fault location, and the probability of the fault occurring.

[0014] A third aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned transformer fault diagnosis method based on the intelligent algorithm.

[0015] Compared with the prior art, the beneficial effects of the present invention are at least as follows: The present invention realizes accurate diagnosis of transformer faults through the application of multimodal data input and multi-task deep learning model. The method realizes the combination of denoising processing, time domain feature extraction, zero-sequence feature calculation and multi-task learning of the transformer output signal through steps S1 to S4, and can effectively extract key features from current and voltage signals and perform real-time analysis. First, a variety of noises, including external noise, environmental noise and internal interference of the power system, are removed by bandpass filters and adaptive filters, thereby improving signal quality; then, frequency domain analysis is performed by wavelet transform, frequency domain features are accurately extracted, and noise components are further eliminated. Secondly, the time rate of change of the target physical quantity is calculated by regression analysis method, and the dynamic changes of the signal are accurately captured by combining time domain and zero-sequence features, so as to determine whether there is an abnormality. In addition, a multimodal input vector is constructed by combining environmental parameters, temperature parameters and historical fault records, and multiple fault diagnosis tasks are jointly trained by deep learning models to output fault type classification, fault location estimation and the probability of fault occurrence. This method can achieve early warning and precise positioning of transformer faults, significantly improve the accuracy and reliability of fault detection, reduce misdiagnosis and missed diagnosis, optimize the operation and maintenance management of the power system, and ensure the safe and stable operation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 Schematic diagram of an embodiment of a transformer fault diagnosis method based on an intelligent algorithm in an embodiment of the present application; Figure 2 This is a schematic diagram of an embodiment of a transformer fault diagnosis system based on an intelligent algorithm in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The embodiments of the present application provide a method, system and storage medium for diagnosing a transformer fault based on an intelligent algorithm. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way are interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or are inherent to these processes, methods, products or devices.

[0019] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of a transformer fault diagnosis method based on an intelligent algorithm includes: Step S1: Obtain the output signal of the secondary side of the current or voltage transformer at a preset sampling frequency, perform denoising on the output signal and convert it into time series data.

[0020] Power systems face many complex challenges when diagnosing instrument transformer faults, primarily due to factors such as unstable signal quality, external interference, and aging equipment. As core measurement devices in power systems, instrument transformers are responsible for monitoring changes in current or voltage and providing essential data for protection, control, and monitoring. However, due to their complex operating environments and prolonged exposure to high voltages, harsh climates, and electromagnetic interference, their output signals are susceptible to noise, resulting in inaccurate or distorted measurement data. Failure to promptly identify and address these issues can lead to serious consequences, such as misdiagnosis or missed diagnosis of power system faults, impacting the safety and stability of the entire power system.

[0021] Therefore, one of the key issues in transformer fault diagnosis is how to extract effective information from noise-affected current or voltage signals to ensure accurate and timely fault detection. Because the signals themselves contain a large amount of environmental noise and interference, traditional diagnostic methods are prone to failure in complex environments and cannot provide accurate fault diagnosis results.

[0022] Specifically, an output signal is obtained from the secondary side of the current or voltage transformer at a preset sampling frequency. For example, assuming that the operating frequency of the voltage signal monitored by the transformer is 50 Hz (i.e., the operating frequency of the power system is 50 Hz). In order to capture the signal details within this frequency range, the sampling frequency needs to be at least 100 Hz, so 100 data points will be collected per second. The output signal is the waveform data of the secondary side current or voltage. The collected signal is often interfered with by external noise and inherent noise of the equipment. Therefore, the present invention performs denoising on the signal through bandpass filters, adaptive filters, and wavelet transforms. The specific denoising process will be described later. The denoised signal is converted into time series data. The format of the time series data can accurately capture the trend and fluctuation of the signal over time. By obtaining the signal at a preset sampling frequency and performing denoising, the accuracy and reliability of the signal data are ensured, and the errors caused by noise are eliminated, thereby improving the accuracy of transformer fault diagnosis.

[0023] Step S2: Perform time domain feature extraction on the time series data to obtain the first data feature and the second data feature of the target physical quantity, and extract the zero-sequence current component and the zero-sequence voltage component from the time series data at the same time, use fast Fourier transform to calculate the frequency domain energy distribution and phase offset, and calculate the phase difference between the zero-sequence current and the zero-sequence voltage as the zero-sequence feature based on the frequency domain energy distribution.

[0024] Specifically, the latest predetermined number of time series data is determined from the time series data, and time domain features are extracted therefrom. The target physical quantity includes the secondary side current or the secondary side voltage. The first data feature includes the average value and the first standard deviation of the target physical quantity. The second data feature includes the time change rate of the target feature and the second standard deviation of the time change rate.

[0025] However, in power systems, faults such as ground faults, short circuit faults or load imbalance usually lead to imbalance in current and voltage signals; failures of the transformer itself are usually caused by equipment aging, damage or other reasons, which lead to distortion of its measurement signal. Therefore, in the case of transformer failure and system failure, both current and voltage output waveforms may be abnormal, but the sources and natures of these two abnormalities are different. Therefore, it may be difficult to accurately distinguish between system failures and transformer failures by relying solely on the time domain features in the time series data. Therefore, it is necessary to extract zero-sequence current components and zero-sequence voltage components from the time series data.

[0026] In a three-phase power system, the zero-sequence component reflects the unbalanced part of the system's current and voltage. Under normal circumstances, the three-phase current and voltage are balanced, so the zero-sequence current and zero-sequence voltage should be very close to zero. A significant increase in zero-sequence current and zero-sequence voltage usually means that the system is unbalanced. This application uses fast Fourier transform (FFT) to convert the time domain signal into frequency domain data, and extracts the energy distribution and phase offset of each frequency component. The specific extraction method will be explained later. Relying solely on time domain features may lead to misjudgment of signal fluctuations, because time domain features cannot capture changes in frequency components in the signal, and these changes are often obvious characteristics when a fault occurs. The introduction of zero-sequence current and zero-sequence voltage analysis can effectively distinguish between anomalies caused by transformer faults and power system faults, thereby reducing the probability of misdiagnosis and missed diagnosis.

[0027] Step S3: Determine whether the transformer is abnormal based on the time domain characteristics. If so, obtain the environmental parameters and temperature parameters of the transformer, and construct a multimodal input vector based on the time domain characteristics, zero-sequence characteristics, environmental parameters, temperature parameters and historical fault records.

[0028] Specifically, when a transformer fails or has an abnormality, the time domain characteristics will change significantly. Therefore, the time domain characteristics are analyzed to determine whether the transformer has an abnormality. However, a single time domain feature may not be sufficient to accurately determine the type and source of the fault. Therefore, it is necessary to construct a multimodal input vector, which not only includes the time domain characteristics of the signal, but also takes into account the zero-sequence characteristics, environmental parameters, temperature parameters and historical fault records of the power system.

[0029] The introduction of zero-sequence features helps distinguish signal anomalies caused by transformer failures from those caused by system faults, reducing the possibility of misjudgment. Environmental parameters include factors in the system operating environment, such as temperature, humidity, air pressure, and electromagnetic interference. Environmental factors significantly impact transformer performance. For example, high temperatures can cause the transformer's coil insulation to age, or electromagnetic interference can lead to measurement signal errors. Extreme environmental fluctuations (such as excessively high or low temperatures) can cause equipment to malfunction. Rising temperatures can also cause transformer performance degradation or failure. Therefore, monitoring temperature changes helps promptly identify equipment health and prevent failure due to overheating in high-temperature environments. Historical fault records contain past fault types, occurrence times, frequency, and maintenance records of the transformer. These records can help the system identify long-term fault trends. In complex environments, a single feature may not be sufficient to address a wide range of fault types. By integrating multiple features, the robustness of the model can be enhanced, making it more adaptable to various operating environments and fault types, especially when dealing with complex or uncertain factors.

[0030] Step S4: Establish a multi-task deep learning model, input the multimodal input vector into the pre-trained multi-task deep learning model, and the deep learning model outputs the fault diagnosis results, including the classification results of the transformer fault type, the estimated value of the fault location, and the probability of the fault occurring.

[0031] Specifically, the multimodal input vectors are first fed into a pre-trained multi-task deep learning model for comprehensive processing. The multi-task deep learning model extracts common features from this data through a shared layer, while simultaneously learning its respective tasks through multiple dedicated layers. Outputs include diagnostic results such as fault type classification, fault location estimation, and fault probability. By integrating multiple features, the deep learning model can more comprehensively understand the operating status of the transformer, thereby improving the accuracy of transformer fault diagnosis. The fusion of data from different features effectively avoids misjudgments or missed diagnoses caused by a single feature.

[0032] In a specific embodiment, performing denoising on the output signal specifically includes the following steps: Perform preliminary filtering on the collected output signal based on the bandpass filter. The output signal is current or voltage waveform data. The frequency range of the bandpass filter is set based on the operating frequency of the power system. The output signal that is not within the frequency range is identified, and the presence of noise components is determined, and the noise components are removed. The output signal after preliminary processing is further denoised based on an adaptive filter to remove external noise, environmental noise and interference noise within the power system. The denoised signal is converted from the time domain to the frequency domain using wavelet transform, and the key frequency components in the signal are extracted. The noise generated by the outside of the system and the environment is removed through frequency domain analysis.

[0033] Specifically, in actual work, there are noises from various sources in the power system, such as electromagnetic interference, environmental noise, sensor noise, etc. These noises will interfere with the signals measured by the transformer, causing signal distortion or inaccuracy, thereby affecting the reliability of fault diagnosis and system monitoring. Therefore, it is necessary to denoise the output signal.

[0034] This application first performs preliminary filtering on the collected current or voltage signal using a bandpass filter. Power systems typically operate within a certain frequency range, with the signal frequency typically concentrated near the operating frequency (e.g., 50Hz or 60Hz). Any signal outside this frequency range may be noise or interference. Therefore, the role of the bandpass filter is to remove high-frequency noise and low-frequency interference by setting an appropriate frequency range (e.g., 50Hz ± 5Hz), while retaining the operating frequency components of the power system.

[0035] Secondly, this application uses an adaptive filter to further denoise the signal after preliminary filtering. By continuously adjusting its filtering characteristics, the adaptive filter can dynamically adapt to different types of noise, especially in an environment with constantly changing noise characteristics. It has a strong denoising ability. Unlike traditional fixed filters, the adaptive filter can self-adjust according to changes in the input signal, removing external noise, environmental noise, and interference noise within the power system.

[0036] Finally, the denoised signal is converted from the time domain to the frequency domain using a wavelet transform. Frequency-domain analysis further removes noise generated external to the system and by the environment. The wavelet transform has excellent time-frequency localization properties, allowing it to decompose the signal at different scales, thereby extracting its key frequency components. This series of denoising processes resolves the unstable signal quality of the transformer. Multi-stage filtering effectively removes different types of noise, resulting in a cleaner signal and significant denoising effect.

[0037] In a specific embodiment, obtaining the first data feature and the second data feature of the target physical quantity specifically includes the following steps: A latest preset number of time series data are extracted from the time series data, wherein the first data feature is the average value and the first standard deviation of the target physical quantity in the latest time series data, and the second data feature includes the time change rate and the second standard deviation of the target physical quantity in the latest time series data. A regression analysis is performed on the latest preset number of time series data, a linear regression equation is established based on the acquisition time point, the predicted value of the target physical quantity at each acquisition time point is calculated, the optimal slope and the optimal intercept are calculated based on the error function, and the optimal slope is defined as the time change rate of the target physical quantity.

[0038] Specifically, first collect the data of the target physical quantity from the secondary side of the transformer, usually the current or voltage signal. For example, if the target physical quantity is the secondary side current signal, then the data extracted from the time series are the current signals at different time points. The secondary side current data collected at the Similarly, if the target physical quantity is a voltage signal, the collected data is the voltage value. This application takes the secondary side voltage as an example, and the average value calculation formula is: , the first standard deviation of the mean is calculated as: , assuming that there are 5 current signals in the time series data, namely: 3.5, 3.8, 3.6, 3.7, 3.9, then =3.7, =0.2.

[0039] Traditional fault diagnosis methods usually rely on simple statistical analysis or feature extraction of signals (such as mean and standard deviation), but these methods often ignore the dynamic changes of signals in the time dimension, especially when the signal changes drastically, and may fail to capture subtle fault signals. Calculating the time rate of change through a linear regression equation is a method that can accurately capture the trend of signal changes, and is particularly suitable for analyzing continuous time series data. Linear regression analysis is performed on the latest n number of time series data to establish the relationship between the target physical quantity and the acquisition time point. The linear regression equation is: y=ax+b, where y represents the predicted value of the target physical quantity at time point x, a is the slope of the regression equation, and b is the intercept. The slope of the regression equation represents the time rate of change of the target physical quantity. Compared with more complex nonlinear models, linear regression models are relatively simple and less prone to overfitting problems, and are particularly suitable for real-time data analysis. Set the error function to: , collect N latest time series data points ,in is the i-th time point, is the actual measured value of the target physical quantity corresponding to the time point, By substituting the error function and taking the partial derivatives with respect to a and b, setting them to zero, we find the point that minimizes the error function, thereby obtaining the optimal slope and intercept. In a linear regression model, the optimal slope represents the rate of change of the target physical quantity over time, while the optimal intercept provides a prediction of the initial state of the device. The optimal slope and intercept are obtained by minimizing the error function, ensuring that the regression model better fits the actual data and reduces prediction error. For dynamic changes in the target physical quantity, the optimized regression model can more accurately reflect the changing trend of the signal, helping to improve the accuracy of fault diagnosis.

[0040] In a specific embodiment, performing regression analysis on the latest preset amount of time series data specifically includes the following steps: The optimal linear regression equation is determined based on the optimal slope and the optimal intercept. The optimal predicted value of the target physical quantity is determined based on the optimal linear regression equation. The residual sum of squares between the actual measured value and the optimal predicted value is obtained. The cumulative sum of the squares of the difference between each acquisition time point and the time series average is calculated, and the residual sum of squares is divided by the degree of dispersion to define the second standard deviation of the time change rate.

[0041] Specifically, the optimal linear regression equation is determined based on the optimal slope and the optimal intercept, and the predicted value of the target physical quantity is calculated for each acquisition time point. The residual sum of squares between the actual measured value and the optimal predicted value is obtained. In order to calculate the degree of dispersion, the difference between the target physical quantity at each acquisition time point and the average value of the time series is calculated, and the cumulative sum of the squares of the difference between each acquisition time point and the average value of the time series is obtained. The residual sum of squares is divided by the degree of dispersion to obtain the second standard deviation of the time change rate. The calculation formula of the second standard deviation is as follows: The second standard deviation of the time rate of change measures the reliability of the slope estimate. A large residual sum of squares or an abnormally large second standard deviation may indicate a transformer fault (such as a ground fault or overload). This method, through regression analysis and combined with the actual signal trend, can provide more accurate fault detection, help identify equipment anomalies, mitigate the impact of equipment failures on the power system, and improve the accuracy and response speed of fault diagnosis.

[0042] In a specific embodiment, the calculation of frequency domain energy distribution and phase offset using fast Fourier transform specifically includes the following steps: Perform fast Fourier transform on the zero-sequence current and zero-sequence voltage to extract the frequency domain signal, calculate the energy of each frequency component in the frequency domain, extract the phase of each frequency component based on the complex number result obtained by fast Fourier transform, and calculate the phase difference between the zero-sequence current and zero-sequence voltage as the phase offset.

[0043] Specifically, the acquired time-domain signal is first converted into a frequency-domain signal using a Fast Fourier Transform (FFT). The FFT decomposes the time-domain signal into its frequency components, revealing the signal's spectral characteristics. Frequency-domain analysis calculates the energy of each frequency component. The magnitude of this energy reflects the importance of each frequency component in the signal, helping to identify the primary components and noise within the signal.

[0044] Based on the complex results obtained by FFT, the phase information of each frequency component is further extracted. This phase information is used to reveal the phase difference between the current and voltage signals. In fault diagnosis, the phase difference between the zero-sequence current and zero-sequence voltage is a key characteristic, reflecting asymmetrical loads or fault conditions in the power system. By calculating the phase difference between the zero-sequence current and zero-sequence voltage, the phase offset is obtained, providing an important basis for fault diagnosis. This technology can effectively identify and locate fault types, especially phase imbalances between current and voltage, ensuring the accuracy and timeliness of diagnostic results, and solving the technical problem that traditional diagnostic methods have difficulty in accurately identifying complex fault problems.

[0045] In a specific embodiment, determining whether a mutual inductor is abnormal based on the time domain characteristics specifically includes the following steps: Set reference values ​​for the first data feature and the second data feature. If the average value and time change rate of the target physical quantity in the time series data are greater than the preset reference value, and the first standard deviation and the second standard deviation are less than or equal to the preset reference value, it is determined that the mutual inductor is abnormal.

[0046] Specifically, in instrument transformer fault diagnosis, the process of determining whether an instrument transformer is abnormal based on time-domain characteristics first requires setting reference values. These reference values ​​are typically based on historical data or statistical data from normal operation. First, by analyzing time series data, the average value and time rate of change of the target physical quantity are calculated. If these values ​​exceed the preset reference values, the change in the target physical quantity has exceeded the normal fluctuation range, which may indicate a device abnormality. Furthermore, the standard deviation and second standard deviation are calculated. These indicators reflect the volatility of the data. If these deviations are less than or equal to the preset reference values, the signal volatility meets the normal operating conditions of the device and there is no abnormal fluctuation. For example, if the average value of the current signal of a certain instrument transformer is 3.8A and the time rate of change is 0.02A / s, and these values ​​are greater than the preset reference values, and the standard deviation is 0.1A and the second standard deviation is 0.05A, both less than or equal to the preset reference values, the instrument transformer is determined to be operating normally. However, if the average value and time rate of change of the current exceed the reference values ​​and the deviation is large, it indicates a device abnormality, such as a ground fault or load overload.

[0047] This approach effectively determines whether a transformer fault has occurred, ensuring stable power system operation and avoiding downtime or accidents caused by equipment failure, thereby improving the accuracy and timeliness of fault diagnosis. By properly setting reference values, this technology can accurately identify potential faults, addressing the inability of traditional diagnostic methods to handle complex signals and unstable data.

[0048] In a specific embodiment, establishing a multi-task deep learning model specifically includes the following steps: The multi-task deep learning model includes at least one shared layer and multiple task-specific layers. The multimodal input vector is input into the pre-trained multi-task deep learning model. The multi-task deep learning model extracts common features through the shared layer, uses historical fault records as training data, and combines time domain features, zero-sequence features, environmental parameters and temperature parameters for training. The model parameters are optimized through back propagation, and the performance of the model is tuned through cross-validation. Fault diagnosis is performed separately through multiple task-specific layers, and the fault diagnosis results are output.

[0049] Specifically, the implementation process of building a multi-task deep learning model for transformer fault diagnosis involves several key steps. First, the multi-task deep learning model consists of at least one shared layer and multiple task-specific layers. The shared layer extracts common features from all input data, including current, voltage, zero-sequence components, and environmental data. The task-specific layers handle different fault diagnosis tasks, such as fault type classification (transformer coil short circuit, aging, insulation damage, etc.), fault location estimation, and fault probability prediction. During the pre-training phase of the multi-task deep learning model, the model is first trained using historical data. This training data includes historical fault records of the equipment, time-domain features, zero-sequence features, environmental parameters, and temperature parameters. Through pre-training, the model learns the basic feature representations for each task from the data. During model training, the model uses historical fault records as label data. These records include information such as the time, type, and duration of the transformer fault, as well as whether it has been repaired. These records help the model identify the equipment's fault mode and perform appropriate training. During model training, a backpropagation algorithm is used to optimize the model parameters. The backpropagation algorithm calculates the error for each task and updates the weights of the shared and task-specific layers based on the error. This process is iteratively optimized until the model achieves satisfactory performance on each task.

[0050] During the training process, these multimodal input data are first fed into a pre-trained multi-task deep learning model. The model extracts common features from the data through shared layers. Then, incorporating historical fault records and other environmental data, the model performs backpropagation, continuously adjusting parameters to optimize the network. Cross-validation is used to evaluate the model's performance and fine-tune model parameters to ensure accuracy across different datasets. Through multiple task-specific layers, the model can handle different fault diagnosis tasks separately, ultimately outputting accurate fault diagnosis results, such as identifying the transformer fault type, estimating the fault location, and predicting the probability of fault occurrence. By combining multimodal input data, the multi-task deep learning model is able to jointly learn across multiple tasks, capable of classifying the fault type, estimating the fault location, and predicting the probability of fault occurrence. This multi-task learning approach helps improve the system's fault identification and prediction capabilities and is capable of comprehensively addressing transformer fault diagnosis.

[0051] The above describes a transformer fault diagnosis method based on an intelligent algorithm in an embodiment of the present application. The following describes a transformer fault diagnosis system based on an intelligent algorithm in an embodiment of the present application. Figure 2 In one embodiment of the present application, a transformer fault diagnosis system based on an intelligent algorithm includes: The preprocessing module obtains the output signal of the secondary side of the transformer at a preset sampling frequency, denoises the output signal and converts it into time series data; An extraction module performs time domain feature extraction on the time series data to obtain the first data feature and the second data feature of the target physical quantity. The module also extracts the zero-sequence current component and the zero-sequence voltage component from the time series data, calculates the frequency domain energy distribution and phase offset using fast Fourier transform, and calculates the phase difference between the zero-sequence current and the zero-sequence voltage as the zero-sequence feature based on the frequency domain energy distribution. The fusion module determines whether the transformer is abnormal based on the time domain characteristics. If so, it obtains the environmental parameters and temperature parameters of the transformer and constructs a multimodal input vector based on the time domain characteristics, zero-sequence characteristics, environmental parameters, temperature parameters and historical fault records. The diagnosis module establishes a multi-task deep learning model, inputs the multimodal input vector into the pre-trained multi-task deep learning model, and the deep learning model outputs the fault diagnosis results, including the classification results of the transformer fault type, the estimated value of the fault location, and the probability of the fault occurring.

[0052] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of a transformer fault diagnosis method based on an intelligent algorithm.

[0053] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0054] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0055] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A transformer fault diagnosis method based on intelligent algorithm, characterized in that: The method comprises: Step S1: obtaining an output signal of the secondary side of the transformer at a preset sampling frequency, performing denoising on the output signal and converting it into time series data; Step S2: performing time domain feature extraction on the time series data to obtain a first data feature and a second data feature of a target physical quantity, extracting a zero-sequence current component and a zero-sequence voltage component from the time series data, calculating a frequency domain energy distribution and a phase offset using a fast Fourier transform, and calculating a phase difference between the zero-sequence current and the zero-sequence voltage based on the frequency domain energy distribution as a zero-sequence feature; Step S3: Determine whether the transformer is abnormal based on the time domain characteristics. If so, obtain the environmental parameters and temperature parameters of the transformer, and construct a multimodal input vector based on the time domain characteristics, zero-sequence characteristics, environmental parameters, temperature parameters and historical fault records; Step S4: Establish a multi-task deep learning model, input the multimodal input vector into the pre-trained multi-task deep learning model, and the deep learning model outputs a fault diagnosis result, including a classification result of the transformer fault type, an estimated value of the fault location, and the probability of the fault occurring.

2. The method according to claim 1, characterized in that Performing denoising processing on the output signal, comprising: Performing preliminary filtering on the collected output signal based on a bandpass filter, wherein the output signal is current or voltage waveform data, setting a frequency range of the bandpass filter based on the operating frequency of the power system, identifying output signals that are not within the frequency range, determining the presence of noise components, and removing the noise components; The output signal after preliminary processing is further denoised based on an adaptive filter to remove external noise, environmental noise and interference noise within the power system. The denoised signal is converted from the time domain to the frequency domain using wavelet transform, and the key frequency components in the signal are extracted. The noise generated outside the system and the environment is removed through frequency domain analysis.

3. The method according to claim 1, characterized in that Obtaining a first data feature and a second data feature of a target physical quantity, comprising: A latest preset number of time series data are extracted from the time series data, wherein the first data feature is the average value and the first standard deviation of the target physical quantity in the latest time series data, and the second data feature includes the time change rate and the second standard deviation of the target physical quantity in the latest time series data. A regression analysis is performed on the latest preset number of time series data, a linear regression equation is established based on the acquisition time point, the predicted value of the target physical quantity at each acquisition time point is calculated, the optimal slope and the optimal intercept are calculated based on the error function, and the optimal slope is defined as the time change rate of the target physical quantity.

4. The method according to claim 3, characterized in that Perform regression analysis on the latest preset amount of time series data, including: An optimal linear regression equation is determined based on the optimal slope and the optimal intercept, an optimal predicted value of a target physical quantity is determined based on the optimal linear regression equation, the sum of squared residuals between actual measured values ​​and optimal predicted values ​​is obtained, the cumulative sum of squares of the differences between each acquisition time point and the average value of the time series is calculated, and the second standard deviation of the time change rate is defined by dividing the sum of squared residuals by the degree of dispersion.

5. The method according to claim 1, wherein Fast Fourier transform is used to calculate the frequency domain energy distribution and phase offset, including: A fast Fourier transform is performed on the zero-sequence current and the zero-sequence voltage to extract the frequency domain signal, the energy of each frequency component is calculated in the frequency domain, the phase of each frequency component is extracted based on the complex result obtained by the fast Fourier transform, and the phase difference between the zero-sequence current and the zero-sequence voltage is calculated as the phase offset.

6. The method according to claim 3, characterized in that Determine whether the transformer is abnormal based on time domain characteristics, including: Set reference values ​​for the first data feature and the second data feature. If the average value and time rate of change of the target physical quantity in the time series data are greater than a preset reference value, and the first standard deviation and the second standard deviation are less than or equal to the preset reference value, it is determined that an abnormality has occurred in the mutual inductor.

7. The method according to claim 1, characterized in that Build multi-task deep learning models, including: The multi-task deep learning model includes at least one shared layer and multiple task-specific layers. The multimodal input vector is input into the pre-trained multi-task deep learning model. The multi-task deep learning model extracts common features through the shared layer, uses historical fault records as training data, and combines time domain features, zero-sequence features, environmental parameters and temperature parameters for training. The model parameters are optimized through back propagation, and the performance of the model is tuned through cross-validation. Fault diagnosis is performed separately through multiple task-specific layers, and the fault diagnosis results are output.

8. A transformer fault diagnosis system based on an intelligent algorithm, used to implement a transformer fault diagnosis method based on an intelligent algorithm as claimed in any one of claims 1 to 7, characterized in that: The system comprises: A preprocessing module, which obtains the output signal of the secondary side of the mutual inductor at a preset sampling frequency, performs denoising on the output signal and converts it into time series data; an extraction module, performing time domain feature extraction on the time series data to obtain a first data feature and a second data feature of a target physical quantity, and extracting a zero-sequence current component and a zero-sequence voltage component from the time series data, calculating a frequency domain energy distribution and a phase offset using a fast Fourier transform, and calculating a phase difference between the zero-sequence current and the zero-sequence voltage based on the frequency domain energy distribution as a zero-sequence feature; The fusion module determines whether the transformer is abnormal based on the time domain characteristics. If so, it obtains the environmental parameters and temperature parameters of the transformer and constructs a multimodal input vector based on the time domain characteristics, zero-sequence characteristics, environmental parameters, temperature parameters and historical fault records. The diagnosis module establishes a multi-task deep learning model, inputs the multimodal input vector into the pre-trained multi-task deep learning model, and the deep learning model outputs a fault diagnosis result, including a classification result of the transformer fault type, an estimated value of the fault location, and the probability of the fault occurring.

9. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the transformer fault diagnosis method based on the intelligent algorithm as described in any one of claims 1 to 7 is implemented.

Images