Method for rotor winding inter-turn short circuit monitoring based on feature enhancement and machine learning

Abstract

The present disclosure provides a rotor winding inter-turn short circuit monitoring method based on feature enhancement and machine learning, relating to the technical field of fault monitoring. The method comprises: collecting multi-source electrical signals in the operation process of the generator rotor winding, including: a first signal for representing the reflection characteristics of the rotor winding and a second signal for representing the electromagnetic state change of the rotor; window segmentation is performed on the multi-source electrical signals to obtain a plurality of windows; for each window, the first feature representing the winding impedance state and the second feature representing the electromagnetic state change are extracted from the window, and are fused into the fusion feature of the window; for the fusion feature, a plurality of classification models with different classification mechanisms are used to perform fault judgment on the fusion feature of the window respectively, and according to the judgment results of each classification model, a fault monitoring result is generated. The present disclosure represents the inter-turn short circuit fault from multiple angles, improves the abnormal recognition ability, and can reduce the misjudgment risk through multi-model collaborative judgment.

Description

Technical Field

[0001] This disclosure relates to the field of fault monitoring technology, and in particular to a method for monitoring inter-turn short circuits in rotor windings based on feature enhancement and machine learning. Background Technology

[0002] Inter-turn short circuits in generator rotor windings are a typical hidden electrical fault in large rotating electrical machines. In the early stages of the fault, the signal amplitude is weak and easily affected by on-site electromagnetic noise, operating condition fluctuations, and baseline drift.

[0003] In related technologies, short-circuit fault monitoring typically relies on technicians conducting RSO (Resistant Circuit State) tests, which suffers from issues such as high subjectivity, insufficient quantification standards, and inadequate stability in identifying weak anomalies. Furthermore, relying solely on a single signal or model makes it difficult to fully utilize the multi-dimensional electrical state variables closely related to fault evolution, resulting in a low early fault detection rate. Summary of the Invention

[0004] To overcome the problems existing in related technologies, this disclosure provides a rotor winding inter-turn short circuit monitoring method based on feature enhancement and machine learning.

[0005] According to a first aspect of the present disclosure, a method for monitoring inter-turn short circuits in rotor windings based on feature enhancement and machine learning is provided, comprising: Collect multi-source electrical signals during the operation of the generator rotor winding; wherein, the multi-source electrical signals include: a first signal for characterizing the reflection characteristics of the rotor winding and a second signal for characterizing the changes in the electromagnetic state of the rotor; The multi-source electrical signals are windowed to obtain multiple windows; For each window, a first feature for characterizing the winding impedance state is extracted from the first signal within the window, and a second feature for characterizing the electromagnetic state change is extracted from the second signal within the window. The first feature and the second feature are then fused to obtain the fused feature of the window. For the fusion features of each window, multiple classification models with different classification mechanisms are used to determine the faults of the fusion features of the window, and the fault monitoring results are generated based on the determination results of each classification model.

[0006] In some embodiments, after acquiring the multi-source electrical signals during the operation of the generator rotor winding, the method further includes: The acquired multi-source electrical signals are preprocessed, and the preprocessing operations include at least one of the following: time alignment, outlier removal, missing value completion, filtering and noise reduction, and baseline drift correction. Based on the mean and standard deviation of the training samples of the classification model, the preprocessed multi-source electrical signals are standardized. The multi-source electrical signal is windowed to obtain multiple windows, including: The standardized multi-source electrical signals are windowed to obtain multiple windows.

[0007] In some embodiments, the preprocessing operation for the acquired multi-source electrical signals includes: Anomalies in multi-source electrical signals are identified and eliminated using the median absolute deviation criterion or the three-standard-deviation criterion; and / or Missing values ​​in multi-source electrical signals are filled in using linear interpolation or the mean of the nearest window; and / or The baseline drift term of the signal is fitted using a polynomial, and then subtracted from the original signal to perform baseline drift correction.

[0008] In some embodiments, the first signal includes: the instantaneous voltage across the slip ring and / or the RSO double-ended reflected waveform; Extracting a first feature characterizing the winding impedance state from the first signal within the window includes: Obtain a dynamic reference waveform that matches the current operating condition; The first signal at both ends of the slip ring within the window is compared with the dynamic reference waveform to obtain the first feature; The first feature includes at least one of the following: waveform overlap at both ends, average difference between both ends, propagation delay, and differential energy.

[0009] In some embodiments, acquiring a dynamic reference waveform that matches the current operating condition includes: Obtain multiple historical waveforms corresponding to the normal state of the current operating condition; The dynamic reference waveform is obtained by weighted summation of the multiple historical waveforms.

[0010] In some embodiments, the second signal includes: a rotor current signal; Extracting a second feature from the second signal within the window to characterize changes in electromagnetic state includes: The second feature is determined based on the rotor current signal within this window; The second feature includes at least one of the following: the mean, standard deviation, maximum variation amplitude, difference with adjacent windows, and current trend slope within the window of the rotor current signal.

[0011] In some embodiments, fusing the first feature and the second feature to obtain the fused feature of the window includes: The first and second features corresponding to the window are concatenated to obtain the multi-source fusion feature vector of the window. The multi-source fusion feature vector is sequentially subjected to feature filtering and dimensionality reduction operations to obtain the fusion feature corresponding to the window. The feature selection includes at least one of correlation selection, variance selection, recursive feature elimination, and model importance-based selection.

[0012] In some embodiments, the multi-source electrical signals further include: auxiliary operating quantities, which include at least one of unit vibration, winding temperature, winding humidity, excitation current, unit load, and speed; The step of concatenating the first and second features corresponding to the window to obtain the multi-source fusion feature vector of the window includes: The first feature, the second feature, and the statistical features of the auxiliary operation quantity corresponding to the window are concatenated to obtain the multi-source fusion feature vector of the window.

[0013] In some embodiments, for the fusion features of each window, multiple classification models with different classification mechanisms are used to determine the faults of the fusion features of that window, and fault monitoring results are generated based on the determination results of each classification model, including: For the fusion features of each window, a first classification model based on maximum margin classification, a second classification model based on nearest neighbor sample matching, and a third classification model based on multi-layer nonlinear mapping are used to determine the fault of the fusion features of the window, and the determination results output by each model are obtained. Based on the pre-configured model weights, the determination results output by each model are weighted and fused to generate a weighted fusion result. Based on the weighted fusion results of each window, fault monitoring results are generated; The model weights are determined based on at least one of the failure recall, false positive rate, and accuracy of each model on the validation set.

[0014] In some embodiments, the weighted fusion result includes a fault risk score; the fault risk score is obtained by weighted summation of the confidence scores of each fault category output by each model. The generation of fault monitoring results based on the weighted fusion results of each window includes: The fault risk score of each window is compared with the multi-level warning threshold to determine the status level corresponding to the window. The status levels include: normal status, suspected abnormal status and inter-turn short circuit fault warning status. If multiple consecutive windows correspond to the same status level, that status level is determined as the fault monitoring result.

[0015] The solution provided in this disclosure can transform the traditional experience-based judgment process that relies on manual observation of waveforms into a computable feature extraction and model discrimination process. By simultaneously characterizing inter-turn short-circuit faults from two perspectives—waveform propagation and reflection characteristics and rotor electromagnetic state changes—it improves the ability to identify early weak anomalies. Furthermore, by using multi-model collaborative discrimination, it reduces the risk of misjudgment by a single model, thereby achieving continuous monitoring and hierarchical early warning. Attached Figure Description

[0016] Figure 1 This diagram illustrates a flowchart of a rotor winding inter-turn short-circuit monitoring method based on feature enhancement and machine learning, according to an embodiment of this disclosure. Figure 2 This diagram illustrates the signal acquisition process in an embodiment of the present disclosure. Figure 3 This diagram illustrates data preprocessing and feature enhancement in an embodiment of this disclosure. Figure 4 This diagram illustrates a multi-model classification fusion and online early warning method in an embodiment of this disclosure. Detailed Implementation

[0017] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0018] The following will describe the exemplary implementation method in detail with reference to the accompanying drawings and embodiments.

[0019] This disclosure provides a method for monitoring inter-turn short circuits in rotor windings based on feature enhancement and machine learning. This method can be executed by any electronic device, such as a server, edge computing unit, or industrial control computer deployed in a generator condition monitoring system or digital operation and maintenance platform. It should be noted that the "monitoring" involved in this disclosure does not refer only to manual visual judgment of a single offline RSO (Recurrent Surge Oscillograph) test waveform, but rather to the continuous windowing processing, automatic feature calculation, model discrimination, and risk classification output of multi-source sampled data during testing or operation and maintenance, thereby forming an intelligent diagnostic process that can be deployed in a generator condition monitoring system or digital operation and maintenance platform.

[0020] Figure 1 This diagram illustrates a flowchart of a rotor winding inter-turn short-circuit monitoring method based on feature enhancement and machine learning, as shown in an embodiment of this disclosure. Figure 1As shown, the method includes the following steps S101 to S104.

[0021] S101, collects multi-source electrical signals during the operation of the generator rotor winding; wherein, the multi-source electrical signals include: a first signal for characterizing the reflection characteristics of the rotor winding and a second signal for characterizing the changes in the electromagnetic state of the rotor.

[0022] In some embodiments, the first signal can be the instantaneous voltage waveform at both ends of the slip ring, acquired by a high-speed acquisition unit after a fast step voltage signal is injected into the generator rotor winding via an RSO test device. In actual engineering scenarios, the RSO test injects a fast step or pulse signal into the rotor winding. When there is an inter-turn short circuit, local insulation degradation, or impedance abrupt change within the winding, the propagating signal will be reflected, distorted, or delayed at abnormal locations, resulting in phenomena such as decreased overlap of the two-end waveforms, local peak-valley shift, and increased differential energy. Therefore, the first signal can reflect the impedance state of the rotor winding from the perspective of waveform propagation and reflection characteristics.

[0023] The second signal can be the rotor current signal, which is acquired by a current sensor (such as a Hall current sensor or a shunt). During the development of an inter-turn short circuit, the rotor current may experience fluctuations in amplitude, short-term abrupt changes, or steady-state shifts, which can supplement the reflection of winding anomalies from an electromagnetic state perspective. Therefore, the second signal can characterize the operating state of the rotor winding from the perspective of electromagnetic state changes.

[0024] Understandably, by simultaneously acquiring the first and second signals, this disclosure can comprehensively characterize inter-turn short-circuit faults from two dimensions: waveform propagation and reflection characteristics and rotor electromagnetic state changes. Compared with the judgment method that relies on only a single signal, it has stronger information integrity and fault identification capabilities.

[0025] S102 performs window segmentation on multi-source electrical signals to obtain multiple windows.

[0026] In some embodiments, to accommodate continuous monitoring, the system can slide-divide continuously sampled data according to a fixed window length W, with a window sliding step size of Δ. Each window serves as a sample input into the subsequent feature calculation process. Through windowing processing, the system can transform one-time waveform observations into continuous sample sequences, providing a foundation for trend monitoring and continuous early warning.

[0027] For example, the window length W can be set according to the actual sampling frequency and signal characteristics. For instance, in an RSO test scenario, the window length can cover a complete pulse injection and reflection process (e.g., a few microseconds to tens of microseconds); in a continuous operation monitoring scenario, the window length can cover several power frequency cycles (e.g., integer multiples of 20ms) to ensure that the window contains sufficient signal information. The sliding step size Δ can be less than or equal to the window length W. When Δ is less than W, there is overlap between adjacent windows, which can improve the sensitivity to capturing short-term anomalies.

[0028] S103, for each window, extract a first feature from the first signal within the window to characterize the winding impedance state, and extract a second feature from the second signal within the window to characterize the electromagnetic state change. Merge the first feature and the second feature to obtain the fused feature of the window.

[0029] In some embodiments, for each window, the system extracts features from a first signal and a second signal, respectively. The first feature characterizes the winding impedance state and may include waveform overlap at both ends, average difference between the two ends, propagation delay, and differential energy, etc.; the second feature characterizes electromagnetic state changes and may include the mean, standard deviation, and maximum amplitude of rotor current, etc. By fusing the first and second features, a fused feature for the window can be obtained, which can simultaneously reflect waveform propagation and reflection differences and rotor electromagnetic state changes.

[0030] S104. For the fusion features of each window, multiple classification models with different classification mechanisms are used to determine the faults of the fusion features of the window respectively, and the fault monitoring results are generated based on the determination results of each classification model.

[0031] In some embodiments, the system can employ multiple classification models with different classification mechanisms to determine faults based on the fused features. Since different classification models have different classification mechanisms and varying sensitivities to fault features, they can assess the fault state from different perspectives. By integrating the determination results of each classification model, the risk of misjudgment by a single model can be reduced, improving the stability and reliability of fault monitoring results.

[0032] For example, fault monitoring results may include normal status, suspected abnormal status, and inter-turn short-circuit fault warning status. The system can perform trend analysis based on the monitoring results of continuous windows, providing auxiliary decision-making basis for generator planned maintenance, insulation condition assessment, and digital operation and maintenance.

[0033] This disclosure transforms the traditional experience-based judgment process, which relies on manual waveform observation, into a computationally achievable feature extraction and model discrimination process. This reduces subjective human bias and improves the consistency and reproducibility of fault identification results. Simultaneously, using the first and second signals as primary inputs, it can characterize inter-turn short-circuit faults from both waveform propagation and reflection characteristics and rotor electromagnetic state changes, improving the ability to identify early weak anomalies. Furthermore, it employs multiple classification models with different classification mechanisms for collaborative discrimination, fully utilizing the complementary advantages of different models and reducing the risk of misjudgment by a single model through fusion.

[0034] The above embodiments have illustrated the overall technical concept of this disclosure. For ease of understanding, the following will combine... Figure 2 The signal acquisition process described in this disclosure is explained.

[0035] In some embodiments, a fast step voltage signal is injected into the generator rotor winding using an RSO test device, and the instantaneous voltage waveforms at both ends of the slip ring are acquired by a high-speed acquisition unit. and Simultaneously, the rotor current is collected via a current sensor. If the monitoring system has the necessary capabilities, it can also simultaneously collect data on unit vibration. Winding temperature ,humidity Auxiliary operating parameters such as excitation current, unit load, and speed.

[0036] The multi-source state vector can be represented as: .in, express The multi-source state vector at any given time. If only the instantaneous slip ring voltage and rotor current are collected on-site, then... , , Set to empty or exclude from subsequent calculations.

[0037] To accommodate continuous monitoring, the system divides continuously sampled data into windows of fixed lengths. Perform sliding segmentation, with a window sliding step size of... . No. A monitoring window can be represented as: Through windowing processing, the system can transform one-time waveform observations into continuous sample sequences, providing a foundation for trend monitoring and continuous early warning.

[0038] Understandably, in actual engineering sites, the multi-source electrical signals collected often suffer from problems such as inconsistent timestamps, different sampling frequencies, high-frequency noise, power frequency interference, baseline drift, outliers, and short-term missing data. These problems directly affect the accuracy of subsequent feature calculations and model discrimination. Therefore, for ease of understanding, the preprocessing and standardization methods for multi-source electrical signals will be explained in detail below.

[0039] Specifically, please refer to Figure 3 After acquiring multi-source electrical signals during the operation of the generator rotor winding, preprocessing operations can be performed on the acquired multi-source electrical signals. These preprocessing operations include at least one of the following: time alignment, outlier removal, missing value completion, filtering and noise reduction, and baseline drift correction. Subsequently, based on the mean and standard deviation of the corresponding training samples in the classification model, the preprocessed multi-source electrical signals are standardized. Step S102 above includes: windowing the standardized multi-source electrical signals to obtain multiple windows.

[0040] For example, regarding time alignment, since multi-source electrical signals may come from different acquisition channels (e.g., RSO acquisition card and current acquisition card), the sampling timestamps of each channel may not be completely consistent. Therefore, it is necessary to perform unified timestamp alignment and equal-interval resampling on various types of signals to ensure that each signal strictly corresponds in the time dimension during subsequent feature calculations.

[0041] For example, the Median Absolute Deviation (MAD) criterion can be used for outlier removal. This criterion first calculates the median of the signal sequence. Then calculate each data point The median of the absolute values ​​of the deviations from the median, i.e. If the deviation of a data point exceeds a preset multiple (e.g., 3 or 5 times) of the MAD value, then the condition is satisfied. (in If a preset threshold is used, the point is considered an outlier and removed. Alternatively, the three-standard-deviation (3σ) criterion can be used, which calculates the mean of the signal sequence. and standard deviation If a data point deviates from the mean by more than three standard deviations, it satisfies the condition. If the signal is abnormal, it is considered an outlier. Both methods can effectively identify and eliminate outliers caused by transient interference from sensors or sampling anomalies.

[0042] For example, linear interpolation can be used for missing value completion. Assume the missing value is located at... and Between, then the missing point The missing values ​​can be filled by linear interpolation of the effective sampled values ​​before and after. Alternatively, the mean of the nearest window can be used, which replaces the missing values ​​with the mean of the corresponding positions in the adjacent window. Linear interpolation is suitable for cases with a small number of missing points and can maintain the local continuity of the signal; the mean of the nearest window is suitable for cases with short-term missing values ​​and slow signal changes and can maintain the overall trend of the signal.

[0043] For example, a polynomial fitting method can be used for baseline drift correction. Specifically, for the original sampled signal... ,use Baseline drift term obtained by polynomial fitting of order 1 ,Right now Then, the baseline drift term is subtracted from the original signal to obtain the corrected signal. This processing can highlight local distortions, peak-to-valley shifts, and propagation delay differences in the reflected waveform. Baseline drift correction is particularly important for RSO waveforms because low-frequency drift directly affects the accuracy of calculating the differences between the two ends of the waveform.

[0044] For filtering and noise reduction, different methods can be used depending on the data type, such as low-pass filtering, band-pass filtering, wavelet denoising, or adaptive filtering. For example, for rotor current signals, low-pass filtering can be used to remove high-frequency noise components; for RSO waveform signals, band-pass filtering can be used to retain effective frequency band information, or wavelet threshold denoising can be used to reduce noise while preserving waveform characteristics.

[0045] After preprocessing, a standardization operation is performed. Based on the mean and standard deviation of the training samples in the classification model, the preprocessed multi-source electrical signals are standardized. Specifically, the standardization formula is as follows: ,in This represents the standardized signal. and represents the mean and standard deviation of the corresponding variable in the training samples, respectively.

[0046] It should be noted that the standardized parameters obtained in the offline training phase are consistently used during the online monitoring phase (i.e., and This ensures consistency between training and deployment. In other words, the mean and standard deviation calculated during model training are saved and directly used during subsequent online monitoring without recalculation. This ensures that the data distribution during online monitoring remains consistent with the data distribution during training, avoiding performance degradation due to data distribution shifts.

[0047] After the standardization process is completed, when executing S102 above, the standardized multi-source electrical signals can be windowed to obtain multiple windows.

[0048] This disclosure improves data quality under conditions of field noise, operating condition fluctuations, and unstable sampling by performing preprocessing operations on multi-source electrical signals, including time alignment, outlier removal, missing value completion, filtering and noise reduction, and baseline drift correction. This makes subsequent feature calculations and model discrimination more stable. Furthermore, standardization based on the mean and standard deviation of training samples eliminates dimensional differences between different physical quantities, ensuring consistency between model training and online deployment.

[0049] The above embodiments have provided a detailed description of the preprocessing and standardization processes. Next, the extraction method of the first feature and the construction method of the dynamic reference waveform will be described in detail.

[0050] Please continue to refer to Figure 3 In some embodiments, the first signal includes: the instantaneous voltage across the slip ring and / or the RSO double-ended reflected waveform. It is understood that the instantaneous voltage across the slip ring is the instantaneous voltage reflected waveform collected at both ends of the slip ring during the RSO test, denoted as... and When the rotor winding is in normal condition, the waveforms at both ends usually have high consistency due to the symmetrical winding structure. However, when there is a local inter-turn short circuit or impedance change, the propagating signal is reflected and distorted at abnormal locations, resulting in a decrease in the consistency of the waveforms at both ends.

[0051] Extracting a first feature from the first signal within the window to characterize the winding impedance state includes: acquiring a dynamic reference waveform that matches the current operating condition; comparing the first signals at both ends of the slip ring within the window with the dynamic reference waveform to obtain the first feature; wherein the first feature includes at least one of the following: double-ended waveform overlap, double-ended average difference, propagation delay, and differential energy.

[0052] Specifically, obtaining a dynamic reference waveform that matches the current operating condition includes: obtaining multiple historical waveforms under normal conditions corresponding to the current operating condition; and performing a weighted summation of the multiple historical waveforms to obtain the dynamic reference waveform.

[0053] Dynamic reference waveform It can be calculated using the following formula: ,in Indicates the first one that is similar to the current working condition. A historical normal waveform, This represents the total number of historical normal waveforms. Represents the corresponding weight, and satisfies .

[0054] It should be noted that the purpose of constructing the dynamic reference waveform is to reduce the impact of operating conditions such as load, temperature, and rotor current variations on the waveform morphology. Under different operating conditions, even if the rotor winding is in normal condition, the morphology of the RSO double-ended waveform may differ. By selecting historical normal waveforms similar to the current operating condition and performing weighted summation, a reference benchmark that reflects the normal state under the current operating condition can be obtained, thereby improving the accuracy of subsequent difference characteristic calculations.

[0055] For example, the judgment of similar operating conditions can be based on parameters such as load level, rotor current magnitude, and winding temperature. Weighting The weighting can be set based on the similarity between historical waveforms and the current operating condition; waveforms with higher similarity are assigned greater weights. For example, the weights can be determined by calculating the Euclidean distance between the current operating condition parameter vector and the historical operating condition parameter vectors; the smaller the distance, the greater the weight.

[0056] After acquiring the dynamic reference waveform, the first signals at both ends of the slip ring within the window are compared with the dynamic reference waveform to obtain the first feature.

[0057] Specifically, the overlap of the waveforms at both ends Used for measurement and The consistency can be calculated using the following formula:

[0058] in, Indicates the window length (i.e., the total number of sampling points within the window). and These represent the mean values ​​of the waveforms at both ends within the window, i.e. , . The closer to 1, the more consistent the waveforms at both ends; A decrease usually indicates a difference in reflection or an increase in local distortion.

[0059] Mean difference between two ends The average difference intensity between the two waveforms can be calculated using the following formula:

[0060] An increase in the value indicates an increase in the difference between the two-terminal waveforms, which can be used as a sensitive indicator of inter-turn short-circuit faults.

[0061] Propagation delay The change in propagation delay between the two waveforms can be determined through cross-correlation analysis:

[0062] in Represents the cross-correlation function. This represents the time offset. When there is a sudden change in impedance within the winding, the signal propagation path and speed will change, resulting in a propagation delay. An anomaly has occurred.

[0063] Differential energy The overall energy of a two-ended differential waveform can be calculated using the following formula: ,in

[0064] Differential energy An increase in the difference between the two-terminal waveforms indicates an overall increase in energy, which can be used as another sensitive indicator of inter-turn short-circuit faults.

[0065] This embodiment of the disclosure effectively reduces the interference of operating condition fluctuations on waveform difference calculation by constructing a dynamic reference waveform, making the extracted first feature more stable and reliable. Simultaneously, by extracting multiple features such as waveform overlap at both ends, average difference between both ends, propagation delay, and differential energy, the differences in waveform propagation and reflection can be characterized from different perspectives, avoiding the problem that a single feature cannot cover complex fault modes.

[0066] The above embodiments have provided a detailed explanation of the extraction method for the first feature. Next, the extraction method for the second feature will be introduced.

[0067] Please continue to refer to Figure 3 In some embodiments, the second signal includes a rotor current signal; extracting a second feature from the second signal within the window to characterize changes in electromagnetic state includes determining the second feature based on the rotor current signal within the window; wherein the second feature includes at least one of the mean, standard deviation, maximum change amplitude, difference with adjacent windows, and current trend slope within the window of the rotor current signal within the window.

[0068] Specifically, for the first There are several windows, and the rotor current signal sequence within each window is given by... , The following second feature can be extracted: Average rotor current This reflects the steady-state level of the rotor current within this window, and the calculation formula is:

[0069] When an inter-turn short circuit reduces the effective number of turns in the winding, the steady-state value of the rotor current may shift under the same excitation voltage.

[0070] Rotor current standard deviation This reflects the degree of fluctuation in rotor current within the window, and the calculation formula is as follows:

[0071] When an inter-turn short circuit causes an imbalance in electromagnetic force or localized heating, the rotor current may fluctuate irregularly, leading to an increase in the standard deviation.

[0072] Maximum variation amplitude of rotor current This reflects the short-time abrupt change in rotor current within the window, and the calculation formula is as follows:

[0073] When an inter-turn short circuit develops suddenly or occurs intermittently, the rotor current may experience a short-term abrupt change, and the maximum change amplitude can capture such transient anomalies.

[0074] Difference with adjacent windows This reflects the change in the average rotor current between adjacent windows, and the calculation formula is:

[0075] in Indicates the first The average rotor current of each window. This feature can detect step changes in rotor current.

[0076] Current trend slope within the window This reflects the changing trend of the rotor current within the window. The trend slope can be obtained by linearly fitting the rotor current signal within the window and taking the slope of the fitted line. Let the fitted line be... Then the slope It can be obtained using the least squares method:

[0077] in This is the mean value indexed by time. When an inter-turn short circuit develops slowly, the rotor current may show a gradual increasing or decreasing trend, and the slope of the trend can reflect this slow change.

[0078] Understandably, if an inter-turn short circuit causes a change in the electromagnetic state, the aforementioned rotor current characteristics, together with the RSO waveform difference characteristics, can enhance the fault characterization capability. This embodiment of the present disclosure, by extracting various statistical features from the rotor current signal, can supplement the reflection of winding anomalies from the perspective of electromagnetic state changes, thus complementing the first feature.

[0079] Next, we will introduce feature fusion, feature selection, and dimensionality reduction methods in detail.

[0080] Please continue to refer to Figure 3In some embodiments, the first feature and the second feature are fused to obtain the fused feature of the window, including: concatenating the first feature and the second feature corresponding to the window to obtain the multi-source fused feature vector of the window; performing feature filtering and dimensionality reduction operations on the multi-source fused feature vector in sequence to obtain the fused feature corresponding to the window; wherein, feature filtering includes at least one of correlation filtering, variance filtering, recursive feature elimination and model importance-based filtering.

[0081] Specifically, the RSO waveform characteristics (i.e., the first characteristic) and the rotor current characteristics (i.e., the second characteristic) can be spliced ​​together to form the second characteristic. Multi-source fusion feature vector of each window .

[0082] In some embodiments, the multi-source electrical signals further include: auxiliary operating quantities, which include at least one of unit vibration, winding temperature, winding humidity, excitation current, unit load, and speed; and concatenating the first feature and the second feature corresponding to the window to obtain the multi-source fusion feature vector of the window, including: concatenating the first feature, the second feature, and the statistical features of the auxiliary operating quantities corresponding to the window to obtain the multi-source fusion feature vector of the window.

[0083] For example, if the monitoring system has the necessary capabilities, it can simultaneously collect data on unit vibration. Winding temperature ,humidity Auxiliary operating parameters, etc. For auxiliary operating quantities, their statistical characteristics (such as root mean square of vibration) can be extracted. Average temperature Average humidity wait).

[0084] In some embodiments, wavelet energy entropy can be further extracted for feature extraction of RSO waveforms. As a supplementary feature, specifically, the RSO waveform within the window is subjected to multi-level wavelet decomposition, and the energy of the detail coefficients and approximation coefficients at each level is calculated. Then, based on the energy distribution, the information entropy is calculated to obtain the wavelet energy entropy. ,in Wavelet energy entropy can reflect the energy distribution of a signal in different frequency bands and capture frequency domain feature changes caused by faults.

[0085] In some embodiments, multi-source fusion feature vectors It can be represented as:

[0086] in, Indicates the first The overlap of the waveforms at both ends of each window. Indicates the first The average difference between the two ends of each window, Indicates the first The propagation delay of each window, Indicates the first The differential energy of each window, Indicates the first Wavelet energy entropy or piecewise energy entropy for a window Indicates the first The average rotor current of each window, Indicates the first The standard deviation of rotor current in each window, Indicates the first The rotor current difference between each window and its adjacent windows. Indicates the first Optional root-mean-square vibration characteristics for each window Indicates the first Winding temperature characteristics for each window, Indicates the first The winding humidity characteristics of each window. This feature vector can simultaneously reflect the differences in waveform propagation and reflection as well as changes in rotor electromagnetic state, providing stronger information completeness compared to manual judgment based on a single waveform.

[0087] By sequentially performing feature filtering and dimensionality reduction operations on the multi-source fusion feature vector, the fusion feature corresponding to the window can be obtained.

[0088] In some embodiments, for correlation screening, the Pearson correlation coefficient between each feature and the fault label can be calculated, and features with an absolute value of the correlation coefficient lower than a preset threshold can be removed.

[0089] For variance screening, the variance of each feature can be calculated, and features with excessively low variance can be removed.

[0090] For recursive feature elimination (RFE), the model can be trained repeatedly and the least important features can be removed to gradually select the optimal feature subset.

[0091] For model-based selection, a random forest model can be trained first, and the importance of features can be evaluated based on the reduction in Gini impurity or information gain of the model output, and features with lower importance rankings can be eliminated.

[0092] After feature selection, dimensionality reduction methods such as Principal Component Analysis (PCA) can be used to map high-dimensional fused features into low-dimensional sensitive feature vectors. Specifically, ,in Indicates the first Feature vectors after dimensionality reduction within a window Indicates from the previous The principal component directions (i.e., the front of the covariance matrix) The projection matrix consisting of the eigenvectors corresponding to the largest eigenvalues, This represents the mean of the features of the training samples. The dimensionality-reduced feature vector retains the main fault information while reducing the complexity of model training.

[0093] Next, we will combine Figure 4 This paper provides a detailed introduction to multi-model classification fusion and online early warning methods.

[0094] Figure 4 The flowchart of the multi-model classification fusion and early warning method in the embodiments of this disclosure is shown.

[0095] In some embodiments, S104 includes: for the fusion features of each window, using a first classification model based on maximum margin classification, a second classification model based on nearest neighbor sample matching, and a third classification model based on multi-layer nonlinear mapping to determine the faults of the fusion features of the window, respectively, to obtain the determination results output by each model; and performing weighted fusion of the determination results output by each model based on pre-configured model weights to generate a weighted fusion result. The model weights are determined based on at least one of the fault recall rate, false positive rate, and accuracy of each model on the validation set.

[0096] Based on the weighted fusion results of each window, the fault monitoring results can be generated.

[0097] Specifically, the first classification model is the Support Vector Machine (SVM) model. SVM separates samples of different classes by finding the optimal hyperplane, and its output can be represented as... ,in Represents the features after kernel mapping. and These are the weight vector and bias of the classification hyperplane, respectively.

[0098] The second classification model is the K-Nearest Neighbors (KNN) model. KNN calculates the distance between the current sample and the training samples, and selects the K nearest training samples for voting. Its output can be represented as... ,in This represents the set of the K closest historical samples to the current sample. These are the category labels for the corresponding samples. This indicates taking the mode.

[0099] The third classification model is the Multilayer Perceptron (MLP) model. MLPs learn complex feature combinations through multilayer nonlinear mapping. Its output can be represented as... ,in This represents a non-linear activation function (such as ReLU). , , , These are the weight matrix and bias vector for the hidden layer and output layer, respectively.

[0100] Based on pre-configured model weights, the judgment results output by each model are weighted and fused to generate a weighted fusion result. The fusion method can employ majority voting, and the final output category can be determined by the following formula:

[0101] in, Indicates the first The final output category of each window, Indicates the candidate category, This represents the total number of models (3 in this case). Indicates the first The voting weights of each model As an indicator function, when the model Output Category Set weights to 1 if the condition is met, otherwise set them to 0. Model weights The accuracy is determined based on at least one of the following: recall, false positive rate, and precision, on the validation set. For example, ,in For the first The F1 score of each model.

[0102] In some embodiments, during the model training process described above, when historical data labels are incomplete, clustering algorithms (such as K-Means or DBSCAN) can be used to initially screen samples for anomalies, classifying them into normal, suspected anomaly, and fault candidate categories. Then, professionals verify these categories using RSO test records, maintenance records, and operation logs to generate training labels for supervised learning. This step does not directly replace diagnostic conclusions with clustering results, but rather addresses the problem of insufficient on-site historical data labels.

[0103] In some embodiments, the weighted fusion result includes a fault risk score, which is obtained by weighted summation of the confidence scores of each fault category output by each model.

[0104] The method for generating fault monitoring results based on the weighted fusion results of each window is as follows: the fault risk score of each window is compared with the multi-level early warning threshold to determine the status level corresponding to the window. The status levels include: normal status, suspected abnormal status and inter-turn short circuit fault early warning status. If the status levels corresponding to multiple consecutive windows are the same, the status level is determined as the fault monitoring result.

[0105] Specifically, fault risk score It can be calculated using the following formula:

[0106] in, Indicates category The corresponding risk weights (e.g., 0 for normal, 0.5 for suspected abnormality, and 1 for failure). Indicates the first Each model output category The confidence level (for SVM and KNN, it can be converted into probability confidence level through Plattscaling or the inverse of distance).

[0107] The failure risk score for each window With multi-level early warning thresholds (such as the first threshold) Second threshold Compare them. When When, the status level is normal; when When, the status level is suspected abnormal; when When the status level is the inter-turn short-circuit fault warning status, if the status level is the same for multiple consecutive windows (e.g., three consecutive windows), the status level will be determined as the fault monitoring result to avoid false alarms caused by occasional noise.

[0108] In some embodiments, the system may also output the following information during online monitoring to assist maintenance personnel in verification: the difference curve of the two-end waveform, the trend of key feature changes, the model confidence level, and the suggested maintenance direction.

[0109] For example, in the online monitoring output: at the first The first window shows a normal identification result, with the diagnostic description indicating high consistency of the RSO two-terminal waveform, small rotor current fluctuation, and a risk score below the concern threshold. In the second window... to The first window identified a suspected abnormal state, with the diagnosis indicating a continuous decrease in the overlap of the dual-ended waveforms, an increase in differential energy, and slight fluctuations in the rotor current. The system output a warning. In the second window... to The system displays a window indicating an inter-turn short circuit fault warning. The diagnostic description is that the multi-model voting results show a continuous tendency towards the fault category, and the risk score exceeds the warning threshold. The system prompts the user to review the RSO test results and check the rotor winding insulation status.

[0110] During long-term operation, the system can re-inject new samples, after manual verification or maintenance, into the sample library for periodic updates of standardized parameters, dynamic reference waveforms, feature selection rules, dimensionality reduction matrices, and classification models. Model updates should retain version numbers and verification records to prevent unverified models from directly affecting on-site alarm results.

[0111] This disclosed embodiment extends offline test analysis into a deployable continuous monitoring and tiered early warning process, outputting statuses such as normal, alert, warning, and fault alarm, providing auxiliary decision-making basis for generator planned maintenance, insulation condition assessment, and digital operation and maintenance. Simultaneously, the system supports sample refeeding and model update mechanisms, enabling continuous optimization of model performance as operation and maintenance data accumulates, improving the accuracy and stability of fault monitoring.

[0112] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A method for monitoring inter-turn short circuits in rotor windings based on feature enhancement and machine learning, characterized in that, include: Collect multi-source electrical signals during the operation of the generator rotor winding; wherein, the multi-source electrical signals include: a first signal for characterizing the reflection characteristics of the rotor winding and a second signal for characterizing the changes in the electromagnetic state of the rotor; The multi-source electrical signals are windowed to obtain multiple windows; For each window, a first feature for characterizing the winding impedance state is extracted from the first signal within the window, and a second feature for characterizing the electromagnetic state change is extracted from the second signal within the window. The first feature and the second feature are then fused to obtain the fused feature of the window. For the fusion features of each window, multiple classification models with different classification mechanisms are used to determine the faults of the fusion features of the window, and the fault monitoring results are generated based on the determination results of each classification model.

2. The method according to claim 1, characterized in that, After acquiring the multi-source electrical signals during the operation of the generator rotor winding, the method further includes: The acquired multi-source electrical signals are preprocessed, and the preprocessing operations include at least one of the following: time alignment, outlier removal, missing value completion, filtering and noise reduction, and baseline drift correction. Based on the mean and standard deviation of the training samples of the classification model, the preprocessed multi-source electrical signals are standardized. The multi-source electrical signal is windowed to obtain multiple windows, including: The standardized multi-source electrical signals are windowed to obtain multiple windows.

3. The method according to claim 2, characterized in that, The preprocessing operation for the acquired multi-source electrical signals includes: Anomalies in multi-source electrical signals are identified and eliminated using the median absolute deviation criterion or the three-standard-deviation criterion; and / or Missing values ​​in multi-source electrical signals are filled in using linear interpolation or the mean of the nearest window; and / or The baseline drift term of the signal is fitted using a polynomial, and then subtracted from the original signal to perform baseline drift correction.

4. The method according to claim 1, characterized in that, The first signal includes: the instantaneous voltage across the slip ring and / or the RSO double-ended reflected waveform; Extracting a first feature characterizing the winding impedance state from the first signal within the window includes: Obtain a dynamic reference waveform that matches the current operating condition; The first signal at both ends of the slip ring within the window is compared with the dynamic reference waveform to obtain the first feature; The first feature includes at least one of the following: waveform overlap at both ends, average difference between both ends, propagation delay, and differential energy.

5. The method according to claim 4, characterized in that, The acquisition of the dynamic reference waveform matching the current operating condition includes: Obtain multiple historical waveforms corresponding to the normal state of the current operating condition; The dynamic reference waveform is obtained by weighted summation of the multiple historical waveforms.

6. The method according to claim 1, characterized in that, The second signal includes: rotor current signal; Extracting a second feature from the second signal within the window to characterize changes in electromagnetic state includes: The second feature is determined based on the rotor current signal within this window; The second feature includes at least one of the following: the mean, standard deviation, maximum variation amplitude, difference with adjacent windows, and current trend slope within the window of the rotor current signal.

7. The method according to claim 1, characterized in that, The step of fusing the first feature and the second feature to obtain the fused feature of the window includes: The first and second features corresponding to the window are concatenated to obtain the multi-source fusion feature vector of the window. The multi-source fusion feature vector is sequentially subjected to feature filtering and dimensionality reduction operations to obtain the fusion feature corresponding to the window. The feature selection includes at least one of correlation selection, variance selection, recursive feature elimination, and model importance-based selection.

8. The method according to claim 7, characterized in that, The multi-source electrical signals also include: auxiliary operating quantities, which include at least one of unit vibration, winding temperature, winding humidity, excitation current, unit load, and speed; The step of concatenating the first and second features corresponding to the window to obtain the multi-source fusion feature vector of the window includes: The first feature, the second feature, and the statistical features of the auxiliary operation quantity corresponding to the window are concatenated to obtain the multi-source fusion feature vector of the window.

9. The method according to claim 1, characterized in that, For the fusion features of each window, multiple classification models with different classification mechanisms are used to determine the faults of the fusion features of that window. Based on the determination results of each classification model, fault monitoring results are generated, including: For the fusion features of each window, a first classification model based on maximum margin classification, a second classification model based on nearest neighbor sample matching, and a third classification model based on multi-layer nonlinear mapping are used to determine the fault of the fusion features of the window, and the determination results output by each model are obtained. Based on the pre-configured model weights, the determination results output by each model are weighted and fused to generate a weighted fusion result. Based on the weighted fusion results of each window, fault monitoring results are generated; The model weights are determined based on at least one of the failure recall, false positive rate, and accuracy of each model on the validation set.

10. The method according to claim 9, characterized in that, The weighted fusion result includes a fault risk score; the fault risk score is obtained by weighted summation of the confidence scores of each fault category output by each model. The generation of fault monitoring results based on the weighted fusion results of each window includes: The fault risk score of each window is compared with the multi-level warning threshold to determine the status level corresponding to the window. The status levels include: normal status, suspected abnormal status and inter-turn short circuit fault warning status. If multiple consecutive windows correspond to the same status level, that status level is determined as the fault monitoring result.

Images