A transformer fault detection method and system

Abstract

The application relates to the technical field of data processing, in particular to a transformer fault detection method and system; the method comprises the following steps: collecting current data of a transformer; the current data is divided into a plurality of time windows with a preset size, the change of the current data in the time windows is analyzed, the pulse shape kurtosis for representing pulse type faults and the periodic waveform skewness for representing periodic distortion faults are calculated; a pulse fault dominant factor is constructed based on the pulse shape kurtosis and the periodic waveform skewness, the data in the time windows is classified based on the pulse fault dominant factor, corresponding random cutting trees are constructed for different types of data, and a random cutting forest containing at least two different types of random cutting trees is formed; the real-time collected current data is input into the random cutting forest for abnormality detection, and a transformer fault detection result is obtained. The application has the effect of improving the transformer detection accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a transformer fault detection method and system. BACKGROUND

[0002] The power transformer is the core static electrical equipment for realizing voltage conversion and power transmission in the power system, and its working principle is based on the law of electromagnetic induction. The power transformer undertakes the key task of raising and lowering voltage in the power grid, and guarantees the safe use of power transmission over long distances and different levels of users. Due to the long-term operation of the transformer in a complex electromagnetic environment and variable load conditions, the transformer is prone to internal faults such as winding deformation, inter-turn short circuit, and core fault under the combined action of electric, thermal, mechanical and other stresses. Once the transformer fails, not only the equipment will be damaged, but also a large area power outage accident may occur, affecting the safe and stable operation of the power grid.

[0003] In order to improve the real-time and accuracy of transformer operation state monitoring, the prior art proposes an anomaly detection method based on robust random cut forest (RRCF). The method constructs a random cut tree set to perform real-time anomaly detection on streaming data, and is suitable for online analysis of monitoring signals such as current. However, in the actual operation process of the transformer, the fault evolution mechanism is complex, and there may be multiple mode fault features caused by different physical causes in the current signal, such as coexistence of pulse type fault and periodic distortion type fault. These two types of faults have significant differences in waveform form and numerical characteristics. The former is a transient impulse signal with sharp peaks and thick tails, and the latter is a persistent waveform distortion with destroyed periodic structure. The traditional RRCF algorithm uses a single and indiscriminate random cutting strategy when constructing a random cutting tree, and does not model different fault modes according to their characteristics. Therefore, it is easy to have a fuzzy feature boundary and unstable abnormal score in a multi-mode mixed signal, which leads to a decrease in the accuracy of the detection result and a risk of missed detection and false detection. SUMMARY

[0004] In order to improve the accuracy of transformer fault detection, the present application provides a transformer fault detection method and system.

[0005] In the first aspect, the present application provides a transformer fault detection method, which adopts the following technical scheme:

[0006] A transformer fault detection method includes collecting current data of a transformer;

[0007] The current data is divided into a plurality of time windows of a predetermined size, the change of the current data in the time window is analyzed, the pulse shape kurtosis for representing the pulse type fault and the periodic waveform skewness for representing the periodic distortion fault are calculated in the time window;

[0008] construct a pulse fault dominant factor based on the pulse shape kurtosis and the periodic waveform skewness, wherein the pulse shape kurtosis is positively correlated with the pulse fault dominant factor, and the periodic waveform skewness is negatively correlated with the pulse fault dominant factor;

[0009] Based on the pulse fault dominant factor, the data in the time window is classified, and corresponding random cut trees are constructed for different categories of data to form a random cut forest containing at least two different types of random cut trees.

[0010] The real-time collected current data is input into the random cut forest for anomaly detection to obtain the transformer fault detection result.

[0011] The pulse shape kurtosis representing the pulse type fault and the periodic waveform skewness representing the periodic distortion fault are calculated respectively. And based on the pulse shape kurtosis and the periodic waveform skewness, a pulse fault dominant factor is constructed to reflect what factor causes the fault. Through the pulse fault dominant factor, the fault mode is pre-classified, and then a special random cut tree is constructed for different fault types to form a dual-mode random cut forest. The targeted modeling method solves the missed detection and false detection problems caused by the ambiguous feature boundary and unstable abnormal score when the traditional method deals with the composite fault of pulse and periodic distortion, thereby improving the accuracy, sensitivity and reliability of the transformer composite fault detection.

[0012] Optionally, the step of calculating the pulse shape kurtosis of the time window for representing the pulse type fault comprises:

[0013] Obtain the fourth-order statistical moment and the second-order statistical moment of the current data in the time window, and take the ratio of the fourth-order statistical moment to the second-order statistical moment as the pulse shape kurtosis corresponding to the time window.

[0014] By using the characteristics that the high-order statistical moment is more sensitive to the peak and thick tail characteristics of data distribution, the transient impact signal characteristics generated by the pulse type fault such as partial discharge are effectively captured and quantified, which provides a stable and reliable numerical basis for subsequent accurate judgment of the pulse fault dominant degree.

[0015] Optionally, the step of calculating the periodic waveform skewness of the time window for representing the periodic distortion fault comprises: for any time window, obtaining the variation period of the current data in the time window, and dividing the current data in the time window into multiple period segments based on the variation period, obtaining the SBD distance of any two period segments, taking the result after adjusting the SBD distance using an exponential function as the distance index, and taking the mean value of the multiple distance indexes as the periodic waveform skewness.

[0016] The change of the current is analyzed, the period of the current change is obtained, and whether the change of the current change occurs is determined by comparing the shapes of different periods (i.e., SBD distance), so as to reflect whether a periodic fault occurs.

[0017] Optionally, the step of constructing the pulse fault dominant factor based on the kurtosis of pulse shape and the skewness of periodic waveform includes:

[0018] The kurtosis of pulse shape and the skewness of periodic waveform are respectively subjected to sigmoid function normalization processing;

[0019] The ratio of the normalized kurtosis of pulse shape to the sum of the normalized kurtosis of pulse shape and the skewness of periodic waveform is calculated, and the ratio is taken as the pulse fault dominant factor.

[0020] The ratio of the kurtosis and the skewness after the sigmoid normalization processing is calculated. This way not only unifies the two indicators with different physical meanings and dimensions to a comparable scale, but also clearly quantifies the dominant position of the pulse type fault in the current time window through the ratio form.

[0021] Optionally, the step of constructing the corresponding random cut tree for different categories of data includes: for any time window; if the pulse fault dominant factor is greater than a preset threshold, the current data in the time window is taken as a feature vector, and a pulse type fault random cut tree is constructed;

[0022] If the pulse fault dominant factor is not greater than the preset threshold, the current data in the time window is classified as periodic distortion fault dominant data, and a feature vector is extracted based on the SBD distance between each period segment in the time window, and a periodic distortion fault random cut tree is constructed.

[0023] Optionally, the step of inputting the real-time collected current data into the random cut forest for anomaly detection includes:

[0024] The vector formed by the real-time collected current data is input into the pulse type fault random cut tree to obtain a first anomaly score;

[0025] The SBD distance matrix of the real-time current data is calculated, and a feature vector is extracted, which is input into the periodic distortion fault random cut tree to obtain a second anomaly score;

[0026] The final anomaly detection score is determined based on the first anomaly score and the second anomaly score.

[0027] The real-time data is input into two different types of random cut trees respectively, and two independent anomaly scores (i.e., the first anomaly score and the second anomaly score) are obtained. Compared with the prior art of outputting a single score by a single model, the dual-channel detection mechanism improves the detection sensitivity and the possibility of fault tracing.

[0028] Optionally, the step of calculating the SBD distance matrix of the real-time current data and extracting the feature vector comprises: obtaining the distance index between any two period segments in the time window to form the SBD distance matrix; calculating the mean, standard deviation, maximum value, and spectral norm of the SBD distance matrix to construct the feature vector.

[0029] Optionally, after the step of collecting the current data of the transformer, the method further comprises: performing mean filtering denoising processing and Z-score normalization processing on the current data.

[0030] Optionally, the mean of the first anomaly score and the second anomaly score is taken as the final anomaly detection score.

[0031] In a second aspect, the present application provides a transformer fault detection system, which adopts the following technical scheme:

[0032] A transformer fault detection system comprises a memory storing computer program instructions, which, when executed by a processor, implement the transformer fault detection method according to the above.

[0033] The transformer fault detection method is generated into a computer program and stored in the memory to be loaded and executed by the processor, so that a system is made according to the memory and the processor, and use is facilitated.

[0034] The present application has the following technical effects:

[0035] The historical current data is analyzed, the current data is classified according to the shape of the current data, then two types of random cut trees are constructed through the classified data, and two random forests are formed, the first anomaly score and the second anomaly score of the real-time data are obtained based on the two types of random forests in the subsequent real-time monitoring process, and the final anomaly score is calculated by combining the two anomaly scores to improve the accuracy of data detection. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a method flowchart of a transformer fault detection method according to an embodiment of the present application.

[0037] Figure 2 is a method flowchart of calculating the period waveform skewness of the time window for representing the periodic distortion fault in the transformer fault detection method according to an embodiment of the present application. DETAILED DESCRIPTION

[0038] The embodiment of the application discloses a transformer fault detection method, first, the current signal of the transformer is subjected to deep feature mining, and quantitative indexes capable of representing pulse type faults and periodic distortion faults, i.e., pulse morphology kurtosis and periodic waveform skewness, are respectively constructed; then, the two indexes are fused to construct a pulse fault dominant factor, so as to accurately distinguish the dominant features of the current fault mode; finally, the cutting tree construction process of the RRCF (Robust Random Cut Forest) algorithm is improved based on the dominant factor, two types of random cutting trees respectively focusing on different fault modes are formed, and thus a dual-mode random cutting forest is constituted, so that accurate and real-time detection of the transformer composite fault is realized.

[0039] Reference Figure 1 A transformer fault detection method comprises steps S1-S5.

[0040] S1: Collecting current data of the transformer.

[0041] In the embodiment, the system installs a current transformer on the high-voltage side of the transformer, and continuously collects the current signal during operation at a high sampling frequency. As a preferred scheme, the sampling frequency is set to 10 MHz, so as to ensure that transient fault features such as partial discharge and the like can be captured. The collected analog signal is converted into a digital signal sequence via a high-speed data acquisition card and transmitted to a back-end processing unit.

[0042] It can be understood that various noises will inevitably be mixed in the originally collected signal. In order to avoid the influence of noise interference on the accuracy of subsequent data analysis, the application first uses a mean filtering algorithm to denoise the current data. Meanwhile, in order to eliminate the signal amplitude variation caused by external factors such as power grid load fluctuation, the application further uses a Z-score normalization method to standardize the denoised data, so that the mean value is adjusted to 0 and the standard deviation is 1, thereby enhancing the stability and comparability of subsequent feature calculation. The mean filtering and Z-score normalization are both well-known techniques in the art, and the specific implementation process is not described here. After preprocessing, the current data for subsequent analysis is obtained.

[0043] S2: Dividing the current data into a plurality of time windows of a preset size, analyzing the change of the current data in the time window, and calculating the pulse morphology kurtosis of the time window for representing the pulse type fault and the periodic waveform skewness for representing the periodic distortion fault.

[0044] In transformer faults, partial discharge and winding breakdown are typical pulse-type faults, which are physically caused by the instantaneous energy release of the dielectric within a short period of time (nanoseconds to microseconds). This manifests in the current signal as a non-Gaussian distributed impulse pulse with sharp peaks and heavy tails. To effectively and stably identify these characteristics, this application constructs a pulse kurtosis index.

[0045] First, the preprocessed high-frequency current data sequence is divided into several non-overlapping time windows, each lasting 10 milliseconds. For any given time window, the current data within it can be represented as a vector. ;in This indicates the number of current data points within the time window.

[0046] Based on the above analysis, higher-order statistics are used to sensitively capture the kurtosis and tail characteristics of the data distribution. Fourth-order statistical moments are more sensitive to extreme values ​​(i.e., spikes) in the data sequence than second-order statistical moments (variance). Therefore, in this embodiment, the fourth-order and second-order statistical moments of the current data within a time window are obtained, and the ratio of the fourth-order to the second-order statistical moments is used as the kurtosis of the pulse shape corresponding to that time window.

[0047] Specifically, for any given time window, the formula for calculating the pulse morphological kurtosis can be expressed as: ;in, Indicates time window Pulse morphology kurtosis; This indicates the number of current data points within the time window; Indicates time window The Middle Current data; Indicates time window The mean of the medium current data.

[0048] When sharp pulses caused by partial discharge occur within the time window, the probability density function of the data sequence exhibits a peak-tailed shape. Because the fourth-order central moment has a power-law amplification effect on extreme data points far from the mean, the molecule... The growth rate will be much faster than the square of the variance in the denominator, that is... This leads to the calculated pulse morphological kurtosis. The value increases significantly. Therefore, An increase in the value directly reflects the severity of the pulse-type fault, providing a reliable quantitative basis for subsequent identification.

[0049] For non-pulse faults, such as inter-turn short circuits and winding mechanical resonance, the characteristic is that a high-frequency oscillation signal is superimposed on the power frequency current, resulting in a continuous and periodic distortion of the current waveform. To quantify this distortion characteristic, this application constructs a periodic waveform skewness index.

[0050] Reference Figure 2 The steps for calculating the periodic waveform skewness of the time window used to characterize the periodic distortion fault include: steps S21-S22.

[0051] S21: For any given time window, obtain the period of change of the current data within the time window, and divide the current data within the time window into multiple period segments based on the period of change.

[0052] For any given time window, calculate the ACF autocorrelation coefficient for different delay step sizes, where the range of the delay step size is... The time window is divided into integers between 1 and 2, with the delay step size at which the ACF autocorrelation coefficient reaches its maximum value taken as the variation period T of the high-frequency current data. There are several periodic segments, among which , Represents the floor function; This indicates the number of periodic segments. The calculation process for the ACF autocorrelation coefficient is a well-known technique, and the specific process will not be elaborated further.

[0053] S22: Obtain the SBD distance between any two period segments, use the result of adjusting the SBD distance using an exponential function as the distance exponent, and use the average of multiple distance exponents as the skewness of the period waveform.

[0054] Periodic waveform skewness is primarily used to measure the similarity between waveforms of different periodic segments within a time window, reflecting whether the periodic structure has been disrupted. When a transformer is operating normally, the waveforms of each periodic segment are highly similar; however, when faults such as inter-turn short circuits occur, the superimposed oscillation signals cause significant differences in the waveforms of each periodic segment. Therefore, this application uses Shape-Based Distance (SBD) to measure the similarity between any two periodic segment waveforms and calculates the periodic waveform skewness based on this.

[0055] Specifically, the formula for calculating the skewness of a periodic waveform can be expressed as:

[0056] ;in, Indicates the first The periodic waveform skewness of each time window Indicates time window The number of inner period segments, when When the value is ≤2, it indicates that the time window lacks periodicity or has a small number of periods, making periodic comparisons impossible. In this case, it can be... Defined as a preset value, for example , to indicate that there is no assessable periodic distortion. This represents an exponential function with the natural constant as its base. represents the SBD distance of the current data in the first period segment and the second period segment in the time window, which can effectively measure the shape similarity of the current data in the two period segments. The calculation method of the SBD distance is a known technology, and the specific process will not be described again.

[0057] When the transformer is in normal operation, the waveform heights of different period segments are similar, the calculation result tends to 1. Therefore, the value tends to a smaller positive number, so that the final period waveform skewness is close to 0. On the contrary, when a turn-to-turn short circuit fault occurs, the current waveform appears periodic distortion, and the waveform similarity between different period segments decreases, the calculation result will decrease, thereby increasing the value of , and finally leading to a significant increase. Therefore, the increase of the value reflects the severity of the periodic distortion fault of the current.

[0058] S3: wherein the kurtosis of pulse form is positively correlated with the pulse fault dominant factor, and the period waveform skewness is negatively correlated with the pulse fault dominant factor.

[0059] The transformer may have a composite fault, i.e. pulse fault and periodic distortion fault exist at the same time, but the intensity and dominance of the two are different. Therefore, in this embodiment, the pulse fault dominant factor is constructed based on the kurtosis of pulse form and the period waveform skewness.

[0060] Specifically, the calculation formula of the pulse fault dominant factor can be represented as: ;

[0061] In the formula, represents the pulse fault dominant factor of the time window , represents the function for normalizing the data, and respectively represent the kurtosis of pulse form and the period waveform skewness of the time window .

[0062] S4: based on the pulse fault dominant factor, the data in the time window is classified, and the corresponding random cut tree is constructed for the data of different categories, forming a random cut forest containing at least two different types of random cut trees.

[0063] In the composite fault of the transformer, if The proportion is large, indicating that the pulse type fault occupies the dominant degree of the current composite fault, that is, the pulse type fault occupies the dominant degree of the fault characteristics, so the calculated pulse fault dominant factor is large; if The proportion is large, indicating that the periodic distortion fault occupies the dominant degree of the current composite fault, that is, the periodic distortion fault occupies the dominant degree of the fault characteristics, so the calculated pulse fault dominant factor is small.

[0064] Therefore, for any time window, if the pulse fault dominant factor is greater than the preset threshold, the current data in the time window is taken as a feature vector to construct a pulse type fault random cut tree.

[0065] A time window includes multiple current data, and the multiple current data are taken as a multi-dimensional feature vector of a data point to construct a pulse type fault random cut tree.

[0066] If the pulse fault dominant factor is not greater than the preset threshold, the current data of the time window is classified as periodic distortion fault dominant data, and a feature vector is extracted based on the SBD distance between each cycle segment in the time window to construct a periodic distortion fault random cut tree.

[0067] The step of extracting a feature vector based on the SBD distance between each cycle segment in the time window includes: obtaining the SBD distance between any two cycle segments (the distance has been calculated in the above step, so the distance can be directly extracted), and further calculating the negative index between each two cycle segments, that is, The calculation result is taken as a distance index, and a plurality of distance indexes constitute an SBD distance matrix, which can be represented as: In the formula, The distance index between the first cycle segment and the The distance index between the first cycle segment and the

[0068] The mean, standard deviation, maximum value and spectral norm of all elements in the SBD distance matrix are calculated, and the vector formed by the mean, standard deviation, maximum value and spectral norm is taken as the feature vector of the time window. In the process of constructing the periodic distortion fault random cut tree, the feature vector is a multi-dimensional vector corresponding to a data point. Based on the feature vectors corresponding to each time window, a periodic distortion fault random cut tree is constructed.

[0069] The preset threshold in this embodiment is 0.5, which can be adjusted by the staff based on actual production experience in other embodiments.

[0070] S5: inputting the real-time collected current data into the random cut forest for anomaly detection.

[0071] The vector formed by the real-time collected current data is input into the pulse type fault random cut tree to obtain a first anomaly score;

[0072] calculating an SBD distance matrix of the real-time current data and extracting a feature vector, inputting the feature vector into a periodic distortion fault random cut tree to obtain a second anomaly score;

[0073] determining a final anomaly detection score based on the first anomaly score and the second anomaly score.

[0074] in response to the final anomaly detection score being greater than a preset anomaly threshold, issuing an alarm.

[0075] The embodiment of the present application also discloses a transformer fault detection system, comprising a processor and a memory, the memory storing computer program instructions, when the computer program instructions are executed by the processor, realizing a transformer fault detection method according to the present application.

[0076] The above system also comprises other components such as a communication bus and a communication interface which are well known to those skilled in the art, and the setting and functions thereof are known in the art, thus not described here.

[0077] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application, therefore: any equivalent changes made on the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A transformer fault detection method, characterized by, The method comprises the following steps: Collecting current data of a transformer; Dividing the current data into a plurality of time windows of a preset size, analyzing the changes of the current data in the time windows, calculating the kurtosis of the pulse shape of the time windows for representing pulse-type faults, and the skewness of the periodic waveform for representing periodic distortion faults, comprising: obtaining the fourth-order statistical moment and the second-order statistical moment of the current data in the time window, and taking the ratio of the fourth-order statistical moment to the second-order statistical moment as the kurtosis of the pulse shape corresponding to the time window; for any time window, obtaining the change period of the current data in the time window, and dividing the current data in the time window into a plurality of period segments based on the change period, obtaining the SBD distance of any two period segments, and taking the result of adjusting the SBD distance using an exponential function as the distance index, and taking the average of a plurality of distance indexes as the skewness of the periodic waveform; Constructing a pulse fault dominant factor based on the kurtosis of the pulse shape and the skewness of the periodic waveform, wherein the kurtosis of the pulse shape is positively correlated with the pulse fault dominant factor, and the skewness of the periodic waveform is negatively correlated with the pulse fault dominant factor; Based on the pulse fault dominant factor, the data in the time window is classified, and a corresponding random cutting tree is constructed for different categories of data to form a random cutting forest containing at least two different types of random cutting trees; the steps of constructing a corresponding random cutting tree for different categories of data include: for any time window; if the pulse fault dominant factor is greater than a preset threshold, the current data in the time window is taken as a feature vector, and a pulse-type fault random cutting tree is constructed; If the pulse fault dominant factor is not greater than the preset threshold, the current data of the time window is classified as periodic distortion fault dominant data, and a feature vector is extracted based on the SBD distance between each period segment in the time window to construct a periodic distortion fault random cutting tree; Inputting the real-time collected current data into the random cutting forest for anomaly detection to obtain a transformer fault detection result.

2. The method of claim 1, wherein, The steps of constructing a pulse fault dominant factor based on the kurtosis of the pulse shape and the skewness of the periodic waveform include: Respectively performing sigmoid function normalization processing on the kurtosis of the pulse shape and the skewness of the periodic waveform; Calculate the ratio of the normalized kurtosis of the pulse shape to the sum of the normalized kurtosis of the pulse shape and the skewness of the periodic waveform, and take the ratio as the pulse fault dominant factor.

3. The method of claim 1, wherein, The steps of inputting the real-time collected current data into the random cutting forest for anomaly detection include: Inputting the vector formed by the real-time collected current data into the pulse-type fault random cutting tree to obtain a first anomaly score; Calculate the SBD distance matrix of the real-time current data and extract the feature vector, input the feature vector into the periodic distortion fault random cutting tree to obtain a second anomaly score; Determine the final anomaly detection score based on the first anomaly score and the second anomaly score.

4. The method of claim 3, wherein, The steps of calculating the SBD distance matrix of the real-time current data and extracting the feature vector include: obtaining the distance index between any two period segments in the time window to form the SBD distance matrix; calculating the mean, standard deviation, maximum value and spectral norm of the SBD distance matrix to construct the feature vector.

5. The method of claim 1, wherein, The step of collecting current data of the transformer further comprises: performing mean filtering denoising processing and Z-score normalization processing on the current data.

6. The method of claim 3, wherein, The mean value of the first abnormal score and the second abnormal score is taken as a final abnormality detection score.

7. A transformer fault detection system characterized by, The method comprises the steps of: The processor and the memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, realize a kind of transformer fault detection method according to any one of claims 1-6.

Images