Method and system for judging on-load tap-changer for transformer

Abstract

The invention discloses a method and a system for judging an on-load tap-changer for a transformer. The method comprises the following steps of: 1, synchronously acquiring multi-source signals; step 2, signal preprocessing; step 3, multi-dimensional feature extraction; 4, generating a dynamic threshold value; step 5, state judgment; 6, self-adaptive compensation is carried out; step 7, backtracking historical data; and step 8, remote communication. According to the invention, multi-source signals are synchronously acquired, vibration and current signals are acquired by using an embedded sensor array, and the comprehensiveness and reliability of signal acquisition are improved; during signal preprocessing, denoising and time-frequency matrix construction are respectively performed on different signals, so that the signal quality is improved; in multi-dimensional feature extraction, four-dimensional feature vectors are formed, dynamic threshold generation is combined with multiple algorithms, and three-level alarm is triggered through state judgment; the self-adaptive compensation link adjusts the feature weight, supports historical data backtracking, and uploads a result through a Modbus-TCP protocol to guarantee the operation safety of a power grid.

Description

Technical Field

[0001] The present invention relates to the field of power equipment monitoring, and in particular to a method and system for determining an on-load tap changer for a transformer. Background Art

[0002] The on-load tap changer (OLTC) is the core component of transformer voltage regulation. The status of its fast-acting mechanism directly affects the safe operation of the power grid.

[0003] Traditional detection methods rely on manual listening or offline testing, which suffer from poor real-time performance and high false positive rates. In recent years, single-source analysis techniques based on vibration or current signals have been gradually adopted. However, single signal sources are susceptible to environmental interference. For example, vibration signals are susceptible to mechanical coupling, and current signals have difficulty capturing transient impact characteristics.

[0004] Some existing technologies attempt to fuse multi-source signals, but most of them use fixed threshold judgment and do not establish a dynamic feature correlation model, resulting in poor adaptability to working conditions. Therefore, a method and system for judging the on-load tap changer for a transformer are proposed. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method and system for determining an on-load tap changer for a transformer, so as to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for determining the on-load tap changer for a transformer, comprising the following steps: Step 1: Synchronous acquisition of multi-source signals: The vibration signal of the switching process and the tap winding current signal are synchronously collected through an embedded sensor array, with a sampling frequency of not less than 200kHz. The vibration sensor adopts a three-axis accelerometer with a range of ±50g and a resolution of 0.1mg. The current sensor adopts a Rogowski coil with a bandwidth of 100kHz, a linearity of ±0.1%, and a signal acquisition timing error of ≤1μs. Step 2: Signal preprocessing: The vibration signal was denoised using wavelet thresholding, using the Symlet wavelet basis with a five-layer decomposition. The current signal was separated from its fundamental harmonics with a cutoff frequency of 5 kHz. A joint time-frequency matrix was constructed. The wavelet threshold was determined using the Stein unbiased risk estimate. The fundamental component was extracted using the fast Fourier transform, and the harmonic components were fitted using the Prony algorithm. The time-frequency resolution of the vibration signal was ≤0.1 ms, and the cutoff frequency for harmonic analysis of the current signal was 5 kHz. Step 3: Multidimensional feature extraction: The vibration signal's time-domain kurtosis (calculation window length is 1 / 3 of the switching process duration), frequency-domain marginal spectral entropy (band division interval is 1 kHz), and current signal transient overshoot (rising edge capture accuracy is 0.1 μs) and harmonic distortion rate (calculated up to the 21st harmonic) are extracted to form a four-dimensional feature vector. The time-domain kurtosis is calculated using a 4096-point sliding window, and the frequency-domain marginal spectral entropy is obtained by integrating the short-time Fourier transform time-frequency graph. The transient overshoot is defined as the absolute value of the difference between the current peak value and the steady-state value. Step 4: Dynamic threshold generation: A Gaussian mixture model (3-5 components) is established based on historical normal operating condition data, and the Mahalanobis distance of the feature vector is calculated in real time to generate a dynamic judgment threshold. The feature selection algorithm based on mutual information reduces the dimensionality of the four-dimensional feature vector and eliminates redundant feature dimensions; the covariance matrix of the Gaussian mixture model is regularized to prevent singularity, and feature selection retains feature combinations with mutual information greater than 0.5 and redundancy less than 0.3; Step 5: Status judgment: When the characteristic vector exceeds the dynamic threshold and lasts longer than 3ms, it is determined that the mechanism is not in place and the three-level alarm mechanism is triggered. The duration is achieved through the hardware comparator and the FPGA timer, with a time accuracy of ±0.1ms. Step 6: Adaptive compensation: Dynamically adjust the characteristic weight based on the transformer load factor. The load factor is calculated by the ratio of the effective value of the tap winding current to the rated current. The weight adjustment uses a fuzzy logic controller. The weight is linearly attenuated by 5% for every 10% increase in the load factor. The transformer load factor is calculated in real time based on the high-voltage side voltage and current, with a sampling period of 100ms. When the load factor is greater than 80%, the current characteristic weight is increased to 60%; Step 7: Historical data backtracking: Stores the feature vectors and judgment results of the last 100 switching processes (data format HDF5, compression ratio 3:1), supporting fault mode backtracking analysis; data storage uses a first-in-first-out queue structure, and each record contains a 256-dimensional feature vector, a timestamp, and a 3-bit status flag; Step 8: Remote Communication The judgment results are uploaded to the dispatching master station through the Modbus-TCP protocol. The communication cycle is adjustable (1s-5min), and the special power protection period is automatically switched to a 1s cycle; First, the mutual information between each feature and the state label in the four-dimensional feature vector is calculated, and the correlation between the feature and the state is quantified by estimating the joint probability distribution and the marginal probability distribution. Then, the redundancy between the features is calculated, and the mutual information between the features is used to measure it. Finally, according to the maximum correlation and minimum redundancy criterion, the feature combinations with mutual information greater than 0.5 and redundancy less than 0.3 are retained to ensure that the feature vector after dimensionality reduction is both effective and concise. Through a series of steps including synchronous acquisition of multi-source signals, signal preprocessing, multi-dimensional feature extraction, dynamic threshold generation, state judgment, adaptive compensation, historical data backtracking and communication management, accurate monitoring and judgment of the transformer on-load tap changer state is achieved. This method not only improves the accuracy and synchronization of signal acquisition, but also effectively improves the accuracy of fault identification through advanced signal processing and feature extraction technologies. The dynamic threshold generation and adaptive compensation mechanism enable the system to be flexibly adjusted according to actual working conditions, enhancing the robustness of judgment. The historical data backtracking function provides strong support for fault analysis, and flexible communication management ensures the timely upload and sharing of judgment results, providing a solid guarantee for the safe and stable operation of the power grid.

[0007] Preferably, the embedded sensor array in step 1 needs to be calibrated before installation, and the calibration content includes sensitivity calibration of the vibration sensor and linearity calibration of the current sensor. The calibration environment temperature is controlled at 20°C ± 2°C and the humidity is controlled at 40%-60% RH; For embedded sensor array calibration, vibration sensor sensitivity calibration involves applying vibrations of known amplitude using a standard vibration table, recording the sensor output, and adjusting the calibration coefficient to align the output with the standard value. Current sensor linearity calibration involves using a standard current source to output currents of varying amplitudes, comparing the sensor output with the standard value, and adjusting the calibration curve through linear regression to ensure measurement accuracy within the specified temperature and humidity range (20°C ± 2°C, 40%-60% RH). By precisely controlling and calibrating the ambient temperature and humidity, and meticulously calibrating the sensitivity of the vibration sensor and the linearity of the current sensor, the impact of environmental factors and the sensor's inherent characteristics on the measurement results is effectively reduced. This not only ensures the authenticity and accuracy of the collected data, but also provides a solid data foundation for subsequent signal processing, feature extraction, and status judgment, thereby improving the reliability and judgment accuracy of the entire transformer on-load tap-changer status monitoring system.

[0008] Preferably, the time domain kurtosis calculation window length in step 3 is dynamically adjusted according to the duration of the switching process and updated in real time using a sliding window method; First, sensors detect the start and end signals of the switching process to determine the duration of the switching process. Then, one-third of this duration is used as the initial window length. Using the sliding window method, the window is continuously moved in the data stream, one data point at a time, and the data within the window is updated in real time to calculate the temporal kurtosis, ensuring that the window always covers the key part of the current switching process. Through synchronous acquisition of multi-source signals and sophisticated signal preprocessing, a high-quality data foundation can be obtained, providing reliable guarantees for subsequent analysis. The multidimensional feature extraction method comprehensively considers the various characteristics of vibration and current signals to form a comprehensive feature vector, which helps to more accurately capture transformer state changes. The dynamic threshold generation and state judgment mechanism can adapt to different working conditions in real time, improving the accuracy and timeliness of fault detection. The adaptive compensation strategy dynamically adjusts the feature weight according to the load rate, enhancing the adaptability and robustness of the system. The historical data backtracking function facilitates fault analysis and system optimization. The communication management unit ensures reliable data transmission, and the human-computer interaction unit provides an intuitive operation interface, which overall improves the intelligence level and practicality of the transformer on-load tapchanger monitoring system.

[0009] Preferably, the Gaussian mixture model in step 4 is iteratively trained using the EM algorithm, and the convergence threshold is set to 1×10⁻ 6 ,The training data set contains no less than 500 normal switching process data; Initialize the parameters of the Gaussian mixture model, including the mean, covariance matrix, and mixing coefficient of each component. Then, use the EM algorithm for iterative training. In each iteration, the posterior probability of each data point belonging to each component is calculated in the E step, and the model parameters are updated in the M step to maximize the likelihood function. The convergence threshold is set to 1×10⁻ 6 , when the parameter change between two consecutive iterations is less than the threshold, the iteration is stopped; the training dataset must contain data of no less than 500 normal switching processes to ensure the generalization ability of the model; The EM algorithm is used to iteratively train the Gaussian mixture model. By setting a strict convergence threshold and a training set containing a large amount of normal switching process data, this step can significantly improve the accuracy and robustness of the model. The EM algorithm effectively solves the complexity of parameter estimation of the Gaussian mixture model by iteratively optimizing parameters, and the strict convergence threshold ensures the stability of the training process and the accuracy of the final model. At the same time, the introduction of a large amount of normal switching process data enables the model to fully learn the characteristic distribution under normal operating conditions, thereby more accurately identifying abnormal conditions in subsequent judgments, providing strong guarantees for the safe operation of the transformer.

[0010] Preferably, the three-level alarm mechanism in step 5 includes: a first-level sound and light alarm (delay 500ms), a second-level switching prohibition (delay 1s) and a third-level trip protection (delay 2s). The first-level alarm triggers a local buzzer and LED indicator light, the second-level alarm locks the tap changer operation circuit through hard wiring, and the third-level alarm directly drives the circuit breaker trip coil. An independent alarm control module is set up in the system. This module receives the alarm signal from the status judgment unit. When it detects that the characteristic vector exceeds the dynamic threshold and the duration is greater than 3ms, the control module initiates corresponding actions according to the alarm level: the first-level alarm is achieved by driving the local buzzer to sound and the LED indicator to flash; the second-level alarm is directly cut off the power supply of the tap changer operating circuit through hard wiring, prohibiting switching operations; the third-level alarm is achieved by driving the circuit breaker tripping coil to achieve trip protection and ensure system safety. The three-level alarm mechanism effectively improves the system's ability to respond to abnormal situations through graded response. The first-level sound and light alarm quickly notifies on-site personnel, facilitating timely inspection; the second-level switching prohibition prevents continued operation under fault conditions, avoiding further damage; the third-level trip protection serves as the last line of defense, quickly cutting off power in the event of a serious fault to protect equipment and personnel. This hierarchical alarm strategy not only ensures timely response, but also avoids unnecessary system downtime, improving system reliability and safety.

[0011] Preferably, the load factor calculation in step 6 adopts a sliding window method, with a window length of 10 minutes and a data update period of 100ms. The load factor is calculated as the ratio of the average value of the effective current value in the window to the rated current, and a recursive averaging algorithm is used to achieve real-time updates; For the specific implementation of load factor calculation, a sliding window method is used, with a window length of 10 minutes and a data update period of 100ms. In each update period, the average value of the effective current in the window is calculated and the ratio is calculated with the rated current to obtain the load factor. To update the load factor in real time, a recursive averaging algorithm is used. That is, each time new data arrives, only the average value in the current window is updated, avoiding recalculation of the entire window data and improving calculation efficiency. The sliding window method combined with the recursive averaging algorithm to calculate the load factor has significant advantages. This method can reflect the load status of the transformer in real time. By dynamically adjusting the data in the window, the timeliness and accuracy of the load factor calculation are ensured. The application of the recursive averaging algorithm reduces the amount of calculation and improves the response speed of the system. In addition, this implementation method can smooth out the fluctuations of the load factor and avoid misjudgment caused by instantaneous current changes. It provides a reliable basis for judging the status of the transformer's on-load tap changer, which helps to improve the stability and safety of the power system.

[0012] Preferably, the data storage in step seven adopts a ring buffer structure, with a coverage period of ≥72 hours and a power-off protection time of ≥72 hours. The ring buffer adopts a circular queue structure, and each storage unit contains a feature vector, a timestamp, and a judgment result. The power-off protection maintains the real-time clock and SRAM data persistence through a supercapacitor; In the ring buffer structure, a circular queue is used to manage data storage units. Each unit is designed to be fixed in size and contains a 256-dimensional feature vector, a timestamp accurate to milliseconds, and a 3-bit status flag. When new data arrives, the oldest data is overwritten in sequence, ensuring an overwriting period of at least 72 hours. Power-off protection is achieved through an integrated supercapacitor, which maintains the real-time clock operation and persistent storage of key data in SRAM during power outages, ensuring that data is not lost for at least 72 hours after a power outage. The use of a ring buffer structure for data storage, combined with circular queue management, effectively ensures data continuity and timeliness, with a coverage period of up to 72 hours, providing sufficient historical data support for fault backtracing. The power-off protection mechanism maintains the real-time clock and SRAM data persistence through supercapacitors, ensuring the security and integrity of critical data even in the event of a sudden power outage, thereby enhancing the reliability and stability of the system. This design not only optimizes the utilization of storage resources, but also improves the system's ability to respond to abnormal situations, providing a solid data foundation for condition monitoring and fault diagnosis of transformer on-load tap-changers.

[0013] Preferably, the communication cycle in step eight is automatically adjusted according to the grid dispatch level, and the adjustment strategy is stored in an EEPROM non-volatile memory; The system automatically selects the appropriate communication cycle based on the current grid dispatch level by reading the preset grid dispatch level and communication cycle correspondence table in the EEPROM. At the same time, the system supports modifying the correspondence table in the EEPROM through the human-computer interaction unit or remote commands to adapt to the needs of different grid environments. This method effectively improves the accuracy of fault diagnosis through synchronous acquisition and precise preprocessing of multi-source signals. Secondly, multi-dimensional feature extraction and dynamic threshold generation mechanism enable the system to adapt to different working conditions in real time, enhancing the flexibility and reliability of fault diagnosis. Furthermore, adaptive compensation and historical data backtracking functions not only optimize system performance, but also provide strong support for fault analysis. Finally, the judgment results are flexibly uploaded to the dispatching master station through the Modbus-TCP protocol, and the communication cycle is automatically adjusted according to the grid dispatching level, ensuring close coordination between the system and grid dispatching, and improving overall operating efficiency and safety.

[0014] Preferably, the feature selection algorithm in step 4 adopts the maximum correlation and minimum redundancy criterion, and retains feature combinations with inter-feature correlation less than 0.8; When using the maximum correlation and minimum redundancy criterion for feature selection, the correlation between each feature and the target variable (such as institutional status) is first calculated. The redundancy between features is also evaluated. By traversing all feature combinations, those with high correlation with the target variable (correlation coefficient greater than a certain threshold, such as 0.5, but here it is explicitly required to keep the correlation less than 0.8 to avoid overfitting) and low redundancy between features are selected. In specific implementation, mutual information can be used as the correlation measure, and conditional entropy can be used to calculate redundancy. An iterative optimization algorithm (such as a greedy algorithm) can be used to find a feature subset that meets the conditions. Using the maximum correlation and minimum redundancy criterion for feature selection can accurately identify and retain the features that are most critical for judging the status of the transformer's on-load tap changer, while eliminating those features that do not contribute much to the judgment result or may cause information redundancy. The advantage of this method is that it effectively improves the accuracy and efficiency of subsequent status judgments, reduces unnecessary computational burden, and enables the system to respond to tap changer status changes more quickly and accurately. In addition, by retaining feature combinations with inter-feature correlations less than 0.8, excessive coupling between features is avoided, further enhancing the robustness and generalization ability of the system.

[0015] A system for determining whether a transformer on-load tap changer is used, using the above-mentioned method for determining whether a transformer on-load tap changer is used, comprises: A signal acquisition unit, an edge computing unit, a communication management unit, and a human-computer interaction unit, wherein the output end of the signal acquisition unit is electrically connected to the edge computing unit, the output end of the edge computing unit is electrically connected to the communication management unit, and the output end of the communication management unit is electrically connected to the human-computer interaction unit; Signal acquisition unit: integrated with a three-axis accelerometer and Rogowski coil, supporting IEEE 1588 synchronization and timing; Edge computing unit: runs feature extraction and judgment algorithms. The core algorithm is implemented through a hardware acceleration engine. Communication management unit: equipped with dual Ethernet interfaces, supports IEC 61850 protocol conversion, and automatically switches to SMS backup channel when communication is interrupted; Human-computer interaction unit: equipped with a touch screen to display real-time waveforms and judgment results, and supports historical data curve playback function; When a communication failure is detected on the primary communication link (e.g., dual Ethernet interfaces), the link monitoring module built into the communication management unit immediately triggers the switching logic. A text message containing the key judgment result is automatically sent to a preset receiving number via the built-in GSM / GPRS module to ensure that the information is not lost. During this process, a hardware watchdog timer monitors the communication status to ensure that the switching action is completed within milliseconds, thus ensuring system reliability. The signal acquisition unit's high-precision synchronization and timing capability ensures the accuracy and real-time nature of the data; the edge computing unit utilizes a hardware acceleration engine to efficiently execute feature extraction and judgment algorithms, significantly improving processing speed and accuracy; the communication management unit's dual Ethernet interfaces and IEC 61850 protocol conversion function enhance the system's compatibility and flexibility, while the SMS backup channel provides reliable data transmission in the event of communication interruptions; the human-computer interaction unit's touch screen design not only intuitively displays real-time waveforms and judgment results, but also supports playback of historical data curves, greatly facilitating the monitoring and troubleshooting work of operation and maintenance personnel.

[0016] In summary, compared with the prior art, the present invention provides a method and system for determining the on-load tap changer for a transformer, which has the following beneficial effects: This invention uses multi-source signal synchronous acquisition and an embedded sensor array to synchronously acquire the vibration signal of the switching process and the tap winding current signal. It comprehensively utilizes the two types of signal information, avoids the problem that a single signal source is susceptible to environmental interference, and improves the comprehensiveness and reliability of signal acquisition. In the signal preprocessing stage, wavelet threshold denoising is performed on the vibration signal, and fundamental harmonic separation is performed on the current signal to construct a joint time-frequency matrix, which effectively improves the signal quality and provides a good foundation for subsequent analysis. In the multidimensional feature extraction step, the multidimensional features of the vibration signal and current signal are extracted to form a four-dimensional feature vector, which comprehensively characterizes the signal characteristics. Dynamic threshold generation establishes a Gaussian mixture model based on historical normal operating condition data and calculates the Mahalanobis distance in real time. The feature selection algorithm combined with mutual information is used to reduce the dimensionality of the feature vector, so that the judgment threshold can change dynamically with the operating conditions, improving the accuracy and adaptability of the judgment. When judging the state, the feature vector, dynamic threshold and duration are comprehensively considered to trigger the three-level alarm mechanism, ensuring the timeliness of fault judgment. The adaptive compensation link dynamically adjusts the feature weight according to the transformer load rate and adopts a fuzzy logic controller to make the judgment results more in line with the actual working conditions. The historical data backtracking function supports fault mode analysis and facilitates problem tracing. The judgment results are uploaded to the dispatching master station via the Modbus-TCP protocol. The communication cycle is adjustable, and the special power protection period automatically switches to a 1s cycle, ensuring the timeliness and flexibility of information transmission and effectively protecting the safe operation of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a step diagram of the transformer on-load tap changer judgment method of the invention.

[0018] Figure 2 The present invention is a flow chart of the transformer on-load tap changer judgment method.

[0019] Figure 3 This is a system diagram for determining the on-load tap changer for the transformer of the invention. DETAILED DESCRIPTION

[0020] The present invention provides a technical solution, a method for determining the load tap changer for a transformer, see Figure 1 、 Figure 2 and Figure 3 , including the following steps: Step 1: Synchronous acquisition of multi-source signals: The vibration signal of the switching process and the tap winding current signal are synchronously collected through an embedded sensor array with a sampling frequency of no less than 200kHz. The vibration sensor uses a three-axis accelerometer with a range of ±50g and a resolution of 0.1mg. The current sensor uses a Rogowski coil with a bandwidth of 100kHz, a linearity of ±0.1%, and a signal acquisition timing error of ≤1μs. Step 2: Signal preprocessing: The vibration signal was denoised using wavelet thresholding, using the Symlet wavelet basis with a five-layer decomposition. The current signal was separated from its fundamental harmonics with a cutoff frequency of 5 kHz. A joint time-frequency matrix was constructed. The wavelet threshold was determined using the Stein unbiased risk estimate. The fundamental component was extracted using the fast Fourier transform, and the harmonic components were fitted using the Prony algorithm. The time-frequency resolution of the vibration signal was ≤0.1 ms, and the cutoff frequency for harmonic analysis of the current signal was 5 kHz. Step 3: Multidimensional feature extraction: The vibration signal's time-domain kurtosis (calculation window length is 1 / 3 of the switching process duration), frequency-domain marginal spectral entropy (band division interval is 1 kHz), and current signal transient overshoot (rising edge capture accuracy is 0.1 μs) and harmonic distortion rate (calculated up to the 21st harmonic) are extracted to form a four-dimensional feature vector. The time-domain kurtosis is calculated using a 4096-point sliding window, and the frequency-domain marginal spectral entropy is obtained by integrating the short-time Fourier transform time-frequency graph. The transient overshoot is defined as the absolute value of the difference between the current peak value and the steady-state value. Step 4: Dynamic threshold generation: A Gaussian mixture model (3-5 components) is established based on historical normal operating condition data, and the Mahalanobis distance of the feature vector is calculated in real time to generate a dynamic judgment threshold. The feature selection algorithm based on mutual information reduces the dimensionality of the four-dimensional feature vector and eliminates redundant feature dimensions; the covariance matrix of the Gaussian mixture model is regularized to prevent singularity, and feature selection retains feature combinations with mutual information greater than 0.5 and redundancy less than 0.3; Step 5: Status judgment: When the characteristic vector exceeds the dynamic threshold and lasts longer than 3ms, it is determined that the mechanism is not in place and the three-level alarm mechanism is triggered. The duration is achieved through the hardware comparator and the FPGA timer, with a time accuracy of ±0.1ms. Step 6: Adaptive compensation: Dynamically adjust the characteristic weight based on the transformer load factor. The load factor is calculated by the ratio of the effective value of the tap winding current to the rated current. The weight adjustment uses a fuzzy logic controller. The weight is linearly attenuated by 5% for every 10% increase in the load factor. The transformer load factor is calculated in real time based on the high-voltage side voltage and current, with a sampling period of 100ms. When the load factor is greater than 80%, the current characteristic weight is increased to 60%; Step 7: Historical data backtracking: Stores the feature vectors and judgment results of the last 100 switching processes (data format HDF5, compression ratio 3:1), supporting fault mode backtracking analysis; data storage uses a first-in-first-out queue structure, and each record contains a 256-dimensional feature vector, a timestamp, and a 3-bit status flag; Step 8: Remote Communication The judgment results are uploaded to the dispatching master station through the Modbus-TCP protocol. The communication cycle is adjustable (1s-5min), and the special power protection period is automatically switched to a 1s cycle; First, the mutual information between each feature and the state label in the four-dimensional feature vector is calculated, and the correlation between the feature and the state is quantified by estimating the joint probability distribution and the marginal probability distribution. Then, the redundancy between the features is calculated, and the mutual information between the features is used to measure it. Finally, according to the maximum correlation and minimum redundancy criterion, the feature combinations with mutual information greater than 0.5 and redundancy less than 0.3 are retained to ensure that the feature vector after dimensionality reduction is both effective and concise. Through a series of steps including synchronous acquisition of multi-source signals, signal preprocessing, multi-dimensional feature extraction, dynamic threshold generation, state judgment, adaptive compensation, historical data backtracking and communication management, accurate monitoring and judgment of the transformer on-load tap changer state is achieved. This method not only improves the accuracy and synchronization of signal acquisition, but also effectively improves the accuracy of fault identification through advanced signal processing and feature extraction technologies. The dynamic threshold generation and adaptive compensation mechanism enable the system to be flexibly adjusted according to actual working conditions, enhancing the robustness of judgment. The historical data backtracking function provides strong support for fault analysis, and flexible communication management ensures the timely upload and sharing of judgment results, providing a solid guarantee for the safe and stable operation of the power grid.

[0021] See also Figure 1 、 Figure 2 and Figure 3,The embedded sensor array in step 1 needs to be calibrated before installation. ,The calibration content includes sensitivity calibration of vibration sensor and ,linearity calibration of current sensor. ,The calibration environment temperature is controlled at 20℃±2℃ and ,humidity is controlled at 40%-60%RH; For embedded sensor array calibration, vibration sensor sensitivity calibration involves applying vibrations of known amplitude using a standard vibration table, recording the sensor output, and adjusting the calibration coefficient to align the output with the standard value. Current sensor linearity calibration involves using a standard current source to output currents of varying amplitudes, comparing the sensor output with the standard value, and adjusting the calibration curve through linear regression to ensure measurement accuracy within the specified temperature and humidity range (20°C ± 2°C, 40%-60% RH). By precisely controlling and calibrating the ambient temperature and humidity, and meticulously calibrating the sensitivity of the vibration sensor and the linearity of the current sensor, the impact of environmental factors and the sensor's inherent characteristics on the measurement results is effectively reduced. This not only ensures the authenticity and accuracy of the collected data, but also provides a solid data foundation for subsequent signal processing, feature extraction, and status judgment, thereby improving the reliability and judgment accuracy of the entire transformer on-load tap-changer status monitoring system.

[0022] See also Figure 1 、 Figure 2 and Figure 3 ,The time domain kurtosis calculation window length in step 3 is dynamically adjusted according to the ,duration of the switching process, and is updated in real time using the ,sliding window method; First, sensors detect the start and end signals of the switching process to determine the duration of the switching process. Then, one-third of this duration is used as the initial window length. Using the sliding window method, the window is continuously moved in the data stream, one data point at a time, and the data within the window is updated in real time to calculate the temporal kurtosis, ensuring that the window always covers the key part of the current switching process. Through synchronous acquisition of multi-source signals and sophisticated signal preprocessing, a high-quality data foundation can be obtained, providing reliable guarantees for subsequent analysis. The multidimensional feature extraction method comprehensively considers the various characteristics of vibration and current signals to form a comprehensive feature vector, which helps to more accurately capture transformer state changes. The dynamic threshold generation and state judgment mechanism can adapt to different working conditions in real time, improving the accuracy and timeliness of fault detection. The adaptive compensation strategy dynamically adjusts the feature weight according to the load rate, enhancing the adaptability and robustness of the system. The historical data backtracking function facilitates fault analysis and system optimization. The communication management unit ensures reliable data transmission, and the human-computer interaction unit provides an intuitive operation interface, which overall improves the intelligence level and practicality of the transformer on-load tapchanger monitoring system.

[0023] See also Figure 1 、 Figure 2 and Figure 3 The Gaussian mixture model in step 4 is iteratively trained using the EM algorithm, and the convergence threshold is set to 1×10⁻ 6 ,The training data set contains no less than 500 normal switching process data; Initialize the parameters of the Gaussian mixture model, including the mean, covariance matrix, and mixing coefficient of each component. Then, use the EM algorithm for iterative training. In each iteration, the posterior probability of each data point belonging to each component is calculated in the E step, and the model parameters are updated in the M step to maximize the likelihood function. The convergence threshold is set to 1×10⁻ 6 , when the parameter change between two consecutive iterations is less than the threshold, the iteration is stopped; the training dataset must contain data of no less than 500 normal switching processes to ensure the generalization ability of the model; The EM algorithm is used to iteratively train the Gaussian mixture model. By setting a strict convergence threshold and a training set containing a large amount of normal switching process data, this step can significantly improve the accuracy and robustness of the model. The EM algorithm effectively solves the complexity of parameter estimation of the Gaussian mixture model by iteratively optimizing parameters, and the strict convergence threshold ensures the stability of the training process and the accuracy of the final model. At the same time, the introduction of a large amount of normal switching process data enables the model to fully learn the characteristic distribution under normal operating conditions, thereby more accurately identifying abnormal conditions in subsequent judgments, providing strong guarantees for the safe operation of the transformer.

[0024] See also Figure 1 、 Figure 2 and Figure 3 The three-level alarm mechanism in step five includes: first-level sound and light alarm (delay 500ms), second-level switching prohibition (delay 1s) and third-level trip protection (delay 2s). The first-level alarm triggers the local buzzer and LED indicator, the second-level alarm locks the tap changer operation circuit through hard wiring, and the third-level alarm directly drives the circuit breaker trip coil; An independent alarm control module is set up in the system. This module receives the alarm signal from the status judgment unit. When it detects that the characteristic vector exceeds the dynamic threshold and the duration is greater than 3ms, the control module initiates corresponding actions according to the alarm level: the first-level alarm is achieved by driving the local buzzer to sound and the LED indicator to flash; the second-level alarm is directly cut off the power supply of the tap changer operating circuit through hard wiring, prohibiting switching operations; the third-level alarm is achieved by driving the circuit breaker tripping coil to achieve trip protection and ensure system safety. The three-level alarm mechanism effectively improves the system's ability to respond to abnormal situations through graded response. The first-level sound and light alarm quickly notifies on-site personnel, facilitating timely inspection; the second-level switching prohibition prevents continued operation under fault conditions, avoiding further damage; the third-level trip protection serves as the last line of defense, quickly cutting off power in the event of a serious fault to protect equipment and personnel. This hierarchical alarm strategy not only ensures timely response, but also avoids unnecessary system downtime, improving system reliability and safety.

[0025] See also Figure 1 、 Figure 2 and Figure 3 The load factor calculation in step six adopts the sliding window method, with a window length of 10 minutes and a data update period of 100ms. The load factor is calculated as the ratio of the average value of the effective current in the window to the rated current, and a recursive averaging algorithm is used to achieve real-time updates; For the specific implementation of load factor calculation, a sliding window method is used, with a window length of 10 minutes and a data update period of 100ms. In each update period, the average value of the effective current in the window is calculated and the ratio is calculated with the rated current to obtain the load factor. To update the load factor in real time, a recursive averaging algorithm is used. That is, each time new data arrives, only the average value in the current window is updated, avoiding recalculation of the entire window data and improving calculation efficiency. The sliding window method combined with the recursive averaging algorithm to calculate the load factor has significant advantages. This method can reflect the load status of the transformer in real time. By dynamically adjusting the data in the window, the timeliness and accuracy of the load factor calculation are ensured. The application of the recursive averaging algorithm reduces the amount of calculation and improves the response speed of the system. In addition, this implementation method can smooth out the fluctuations of the load factor and avoid misjudgment caused by instantaneous current changes. It provides a reliable basis for judging the status of the transformer's on-load tap changer, which helps to improve the stability and safety of the power system.

[0026] See also Figure 1 、 Figure 2 and Figure 3 ,The data storage in step seven adopts a ring buffer structure, with a coverage period ≥ 72 hours and a power-off protection time ≥ 72 hours. The ring buffer adopts a circular queue structure. Each storage unit contains a feature vector, a timestamp and a judgment result. The power-off protection maintains the real-time clock and SRAM data persistence through a super capacitor; In the ring buffer structure, a circular queue is used to manage data storage units. Each unit is designed to be fixed in size and contains a 256-dimensional feature vector, a timestamp accurate to milliseconds, and a 3-bit status flag. When new data arrives, the oldest data is overwritten in sequence, ensuring an overwriting period of at least 72 hours. Power-off protection is achieved through an integrated supercapacitor, which maintains the real-time clock operation and persistent storage of key data in SRAM during power outages, ensuring that data is not lost for at least 72 hours after a power outage. The use of a ring buffer structure for data storage, combined with circular queue management, effectively ensures data continuity and timeliness, with a coverage period of up to 72 hours, providing sufficient historical data support for fault backtracing. The power-off protection mechanism maintains the real-time clock and SRAM data persistence through supercapacitors, ensuring the security and integrity of critical data even in the event of a sudden power outage, thereby enhancing the reliability and stability of the system. This design not only optimizes the utilization of storage resources, but also improves the system's ability to respond to abnormal situations, providing a solid data foundation for condition monitoring and fault diagnosis of transformer on-load tap-changers.

[0027] See also Figure 1 、 Figure 2 and Figure 3 ,The communication cycle in step eight is automatically adjusted according to the grid dispatch level, and the adjustment strategy is stored in the EEPROM non-volatile memory; The system automatically selects the appropriate communication cycle based on the current grid dispatch level by reading the preset grid dispatch level and communication cycle correspondence table in the EEPROM. At the same time, the system supports modifying the correspondence table in the EEPROM through the human-computer interaction unit or remote commands to adapt to the needs of different grid environments. This method effectively improves the accuracy of fault diagnosis through synchronous acquisition and precise preprocessing of multi-source signals. Secondly, multi-dimensional feature extraction and dynamic threshold generation mechanism enable the system to adapt to different working conditions in real time, enhancing the flexibility and reliability of fault diagnosis. Furthermore, adaptive compensation and historical data backtracking functions not only optimize system performance, but also provide strong support for fault analysis. Finally, the judgment results are flexibly uploaded to the dispatching master station through the Modbus-TCP protocol, and the communication cycle is automatically adjusted according to the grid dispatching level, ensuring close coordination between the system and grid dispatching, and improving overall operating efficiency and safety.

[0028] See also Figure 1 、 Figure 2 and Figure 3 ,The feature selection algorithm in step 4 adopts the maximum correlation and minimum redundancy criterion, and retains the feature combinations with inter-feature correlation < 0.8; When using the maximum correlation and minimum redundancy criterion for feature selection, the correlation between each feature and the target variable (such as institutional status) is first calculated. The redundancy between features is also evaluated. By traversing all feature combinations, those with high correlation with the target variable (correlation coefficient greater than a certain threshold, such as 0.5, but here it is explicitly required to keep the correlation less than 0.8 to avoid overfitting) and low redundancy between features are selected. In specific implementation, mutual information can be used as the correlation measure, and conditional entropy can be used to calculate redundancy. An iterative optimization algorithm (such as a greedy algorithm) can be used to find a feature subset that meets the conditions. Using the maximum correlation and minimum redundancy criterion for feature selection can accurately identify and retain the features that are most critical for judging the status of the transformer's on-load tap changer, while eliminating those features that do not contribute much to the judgment result or may cause information redundancy. The advantage of this method is that it effectively improves the accuracy and efficiency of subsequent status judgments, reduces unnecessary computational burden, and enables the system to respond to tap changer status changes more quickly and accurately. In addition, by retaining feature combinations with inter-feature correlations less than 0.8, excessive coupling between features is avoided, further enhancing the robustness and generalization ability of the system.

[0029] A system for determining the use of on-load tap changers for transformers, using the above-mentioned method for determining the use of on-load tap changers for transformers, see Figure 3 ,include: A signal acquisition unit, an edge computing unit, a communication management unit, and a human-computer interaction unit, wherein the output end of the signal acquisition unit is electrically connected to the edge computing unit, the output end of the edge computing unit is electrically connected to the communication management unit, and the output end of the communication management unit is electrically connected to the human-computer interaction unit; Signal acquisition unit: integrated with a three-axis accelerometer and Rogowski coil, supporting IEEE 1588 synchronization and timing; Edge computing unit: runs feature extraction and judgment algorithms. The core algorithm is implemented through a hardware acceleration engine. Communication management unit: equipped with dual Ethernet interfaces, supports IEC 61850 protocol conversion, and automatically switches to SMS backup channel when communication is interrupted; Human-computer interaction unit: equipped with a touch screen to display real-time waveforms and judgment results, and supports historical data curve playback function; When a communication failure is detected on the primary communication link (e.g., dual Ethernet interfaces), the link monitoring module built into the communication management unit immediately triggers the switching logic. A text message containing the key judgment result is automatically sent to a preset receiving number via the built-in GSM / GPRS module to ensure that the information is not lost. During this process, a hardware watchdog timer monitors the communication status to ensure that the switching action is completed within milliseconds, thus ensuring system reliability. The signal acquisition unit's high-precision synchronization and timing capability ensures the accuracy and real-time nature of the data; the edge computing unit utilizes a hardware acceleration engine to efficiently execute feature extraction and judgment algorithms, significantly improving processing speed and accuracy; the communication management unit's dual Ethernet interfaces and IEC 61850 protocol conversion function enhance the system's compatibility and flexibility, while the SMS backup channel provides reliable data transmission in the event of communication interruptions; the human-computer interaction unit's touch screen design not only intuitively displays real-time waveforms and judgment results, but also supports playback of historical data curves, greatly facilitating the monitoring and troubleshooting work of operation and maintenance personnel.

[0030] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0031] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for determining the on-load tap changer for a transformer, characterized in that: The steps include: Step 1: Synchronous acquisition of multi-source signals: The vibration signal of the switching process and the current signal of the tap winding are synchronously collected through an embedded sensor array. The vibration sensor adopts a three-axis accelerometer, and the current sensor adopts a Rogowski coil. Step 2: Signal preprocessing: The vibration signal is denoised using wavelet thresholding and the Symlet wavelet basis is used. The fundamental and harmonic components of the current signal are separated and a joint time-frequency matrix is ​​constructed. The wavelet threshold is determined using the Stein unbiased risk estimation, and the fundamental component is extracted by fast Fourier transform. Step 3: Multidimensional feature extraction: The time-domain kurtosis and frequency-domain marginal spectral entropy of the vibration signal, as well as the transient overshoot and harmonic distortion rate of the current signal, are extracted to form a four-dimensional feature vector. The time-domain kurtosis is calculated using a 4096-point sliding window, and the frequency-domain marginal spectral entropy is obtained by integrating the short-time Fourier transform time-frequency graph. Step 4: Dynamic threshold generation: A Gaussian mixture model is established based on historical normal operating condition data, and the Mahalanobis distance of the feature vector is calculated in real time to generate a dynamic judgment threshold. The feature selection algorithm based on mutual information is used to reduce the dimensionality of the four-dimensional feature vector; the covariance matrix of the Gaussian mixture model is regularized to prevent singularity; Step 5: Status judgment: When the characteristic vector exceeds the dynamic threshold and lasts longer than 3ms, it is determined that the mechanism is not in place and the three-level alarm mechanism is triggered. The duration is achieved by the hardware comparator and the FPGA timer. Step 6: Adaptive compensation: Dynamically adjust the characteristic weights according to the transformer load factor. The load factor is calculated by the ratio of the tap winding current RMS to the rated current. The weight adjustment adopts a fuzzy logic controller. Step 7: Historical data backtracking: Store the feature vectors and judgment results of the latest 100 switching processes; Step 8: Remote Communication The judgment results are uploaded to the dispatching master station through the Modbus-TCP protocol. The communication cycle is adjustable, and the special power protection period is automatically switched to a 1s cycle.

2. The method for determining the on-load tap changer for a transformer according to claim 1, characterized in that: The embedded sensor array in step 1 needs to be calibrated before installation, and the calibration content includes sensitivity calibration of the vibration sensor and linearity calibration of the current sensor.

3. The method for determining the on-load tap changer for a transformer according to claim 1, characterized in that: The time domain kurtosis calculation window length in step three is dynamically adjusted according to the duration of the switching process and is updated in real time using a sliding window method.

4. The method for determining the on-load tap changer for a transformer according to claim 1, characterized in that: The Gaussian mixture model in step 4 is iteratively trained using the EM algorithm, and the convergence threshold is set to 1×10⁻ 6 ,The training dataset contains no less than 500 normal switching process data.

5. The method for determining the on-load tap changer for a transformer according to claim 1, characterized in that: The three-level alarm mechanism in step five includes: a first-level sound and light alarm, a second-level switching prohibition, and a third-level trip protection. The first-level alarm triggers a local buzzer and LED indicator light, the second-level alarm locks the tap changer operating circuit through hard wiring, and the third-level alarm directly drives the circuit breaker tripping coil.

6. The method for determining the on-load tap changer for a transformer according to claim 1, characterized in that: The load factor calculation in step six adopts the sliding window method. The load factor is calculated as the ratio of the average value of the effective current in the window to the rated current, and a recursive averaging algorithm is used to achieve real-time update.

7. The method for determining the on-load tap changer for a transformer according to claim 1, characterized in that: The data storage in step seven adopts a ring buffer structure with a coverage period of ≥72 hours and a power-off protection time of ≥72 hours. The ring buffer adopts a circular queue structure, and each storage unit contains a feature vector, a timestamp, and a judgment result. The power-off protection maintains the real-time clock and SRAM data persistence through a super capacitor.

8. The method for determining the on-load tap changer of a transformer according to claim 1, characterized in that: The communication cycle in step eight is automatically adjusted according to the grid dispatch level, and the adjustment strategy is stored in the EEPROM non-volatile memory.

9. The method for determining the on-load tap changer for a transformer according to claim 1, characterized in that: The feature selection algorithm in step 4 adopts the maximum correlation and minimum redundancy criterion, and retains feature combinations with inter-feature correlation less than 0.

8.

10. A system for determining the status of an on-load tap changer for a transformer, using the method for determining the status of an on-load tap changer for a transformer according to any one of claims 1 to 9, characterized in that: include: A signal acquisition unit, an edge computing unit, a communication management unit and a human-computer interaction unit, wherein the output end of the signal acquisition unit is electrically connected to the edge computing unit, the output end of the edge computing unit is electrically connected to the communication management unit, and the output end of the communication management unit is electrically connected to the human-computer interaction unit.

Images