CN114372093A - Processing method of DGA (differential global alignment) online monitoring data of transformer - Google Patents

Abstract

The invention proposes a method for processing transformer DGA online monitoring data. According to the characteristics of the returned data, the online data is equivalent to a time series; the idea of sliding window algorithm is introduced in the first stage, and an improved sequence piecewise linearization is proposed. The algorithm divides the sequence data into several line segments characterized by slope and span, and then uses the improved K-means clustering to symbolize the online monitoring data. Finally, the APRIORI algorithm is used to mine the correlation between different indicators in the DGA, and the In the second stage, according to the filtered abnormal value sampling points, the improved support vector regression algorithm of particle swarm optimization is used to ensure the solution speed and solution diversity of the algorithm, and optimize the support vector regression algorithm. The key parameters repair these sampling points to complete the processing of transformer online DGA monitoring data.

Description

A method of processing transformer DGA online monitoring data

technical field

The invention relates to a method for processing transformer DGA online monitoring data, which belongs to the field of data cleaning of power equipment.

Background technique

Power transformers are pivotal equipment for power conversion and transmission, and their safe and stable operation is an important guarantee for the quality of power supply to users. The online data of the DGA indicators of the transformer is the real-time monitoring of the insulation performance of the equipment. Based on the analysis of the oil chromatography data, the real-time status of the transformer can be quickly obtained; at the same time, there are many indicators in the DGA data. , which helps to identify data with different abnormal patterns in online data, and can enhance the credibility of the comprehensive state evaluation results of the equipment.

Due to the operating environment of the equipment and some electromagnetic interference effects of the transformer itself, the online monitoring device is prone to randomly distributed abnormal value points in the process of data acquisition and transmission, and even data drift and transmission interruption in severe cases. For obvious data anomalies such as data drift and data interruption, the back-end system can quickly identify the problem and issue an alarm for the problem; but for those abnormal numerical points randomly distributed in normal online data, the real-time characterization of equipment status indicators is effective. It also affects the status evaluation work based on indicators, and it is easy to cause false alarms and false alarms of abnormal equipment status, resulting in a waste of equipment operation and maintenance resources.

Power transformers are important equipment to ensure the stable operation of the power transmission and distribution network. The monitoring data of the transformer's iron core grounding current is an important basis for evaluating the status of the transformer. The monitoring data for a period of time, including its overall change trend, extreme points and jump points in the change, and statistical characteristics of the data, can reflect the possible abnormal conditions inside the power transformer from many aspects.

After the long-term operation of power equipment, large-scale index data has been stored in the power database, which must contain index data of different abnormal patterns. Analyzing data of different abnormal patterns in the data based on this relationship, and effectively repairing these data, is conducive to improving the comprehensive state evaluation system of power equipment, discovering abnormal states of equipment devices early, improving equipment maintenance efficiency, and reducing equipment operation and maintenance. cost.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for processing transformer DGA online monitoring data, so as to solve the above-mentioned problems of the background technology.

The present invention is achieved through the following technical solutions, a method for processing transformer DGA online monitoring data, comprising the following steps:

S1. Sliding window processing of datasets: The idea of sliding windows is introduced, and a window of length L is used to intercept online datasets;

个数据窗口，导出所有窗口中的数据，构成待分析数据集DS_i，i∈n；S2. Traverse the online dataset by sliding the window with a certain step size: set the sliding step size to l, drag the window to slide on the overall dataset until all data is traversed; let the length of the online dataset be L ₁ , and get the result after traversing

data windows, export the data in all windows to form the data set DS _i to be analyzed, i∈n;

S3. Piecewise linearization of sequence data: A piecewise linearization algorithm of sequence data is proposed, which combines indeterminate points in online data to form multiple sets of data points; the standard for grouping data points is that all points are The error between the combined line segment and the actual data point is less than the threshold, and the slope of the line segment and the line segment span are used to characterize the fitted line segment;

S4. Build a model describing the similarity of different line segments: build a similarity model based on the slope and span of the line segment, and use the improved K-means clustering algorithm based on the maximum and minimum distances to classify the line segments, and assign symbols to the same type of line segments, complete Symbolization of sequence data;

S5. Mining the correlation between different sequences: Based on the idea of Apriori algorithm, set the minimum confidence and support, mine the frequent itemsets existing between different sequences, and quantify the correlation between different sequences;

S6. Extract and screen out the abnormal values existing in the DGA online monitoring data: according to the strength of the correlation between the sequences, to determine the types of abnormal values existing in the data, separate data with different abnormal patterns;

S7. Improve particle swarm optimization support vector regression: define the distance between particle solution sets, calculate the density of different particles based on the distance, and define the update method of particles according to the density; use the algorithm to optimize the key parameters of support vector regression, complete Processing of DGA Online Data.

Further, the specific steps of the piecewise linearization algorithm for sequence data proposed in S3 are:

1) For online monitoring data of equipment indicators similar to DGA, it is equivalent to time series data;

2) For the time series X _K ={x ₁ ,x ₂ ,...,x _k }, use a window of length L (L<k) to intercept data points, and for the data in the intercepted window, based on the idea of sliding windows, to The data points contained in it are subjected to piecewise linear fitting;

3) Take the first data point in the window as the fitting starting point of the initial line segment, let this point be _xi , and assuming that the fitting end point of the initial line segment is _xi+m (m>1), use this m+1 data points are fitted to a line segment;

4) Then for such a line segment, use the following formula to express:

my-(X _i+m-1 -X _i )X-(m-1)X _i +X _i+m-1 =0 (2)

Take the distance from the actual data point to the fitted line segment as the fitting error; calculate the distance from all actual data points to the line segment within the step size of the fitted line segment, and use the sum as the overall fitting error ER of the line segment:

5) Set the fitting error threshold to ER _r , if ER < ER _r , it means that the line segment can continue to add fitting points, let m=m+1, and repeat the above steps; if there is ER > ER _r , then determine The line segment cannot be fitted, save the fitting end point of the current line segment as X _end =X _i+m-1 , record its data sampling time, and then return to step 3), reset the parameter m, and use the current fitting end point as the lower The fitting starting point of a line segment performs the next part of data fitting until all data points in the series are fitted.

Further, the main steps of constructing a similarity model in S4 and performing cluster analysis based on this model are:

的标准化操作；1) Form the attributes of all line segments existing in the same sequence as

standardized operations;

2) In the cluster analysis, establish a standard for measuring the similarity of line segments; extract two key parameters, the slope and the span of the line segment, use the Euclidean distance to describe the similarity between the line segments, and express the difference between the different attributes of the line segment in the form of weights. Consider the degree; the established line segment similarity model is as follows:

3) Based on the above-mentioned line segment similarity model, clustering analysis is performed on the line segment set using the improved K-means algorithm based on the maximum and minimum distances, and the similar line segments are divided into the same category.

Further, the main steps of the improved K-means algorithm based on the maximum and minimum distances in S4 are:

1) The maximum and minimum distances are also based on the Euclidean distance, which differs from the K-means algorithm in that it takes the object as far away as possible as the cluster center; for the sample set, a proportional coefficient θ (0<θ<1) is given. , any sample in the sample set _sn is taken as the initial cluster center, denoted as z ₁ ;

2) arbitrarily take the sample farthest from z ₁ in the remaining n-1 samples as the second cluster center, denoted as z ₂ ;

3) Calculate the distance between the remaining n-2 samples and z ₁ and z ₂ , and find the minimum value, namely:

D _ij =||x _i -z _j ||,j=1,2 (6)

D _i =min(D _i1 ,D _i2 ),i=1,2,...,n (7)

4) If

D _i =max{D _i }>θ×||z _i -z ₂ || (8)

Then select the corresponding sample _si as the third cluster center z ₃ ;

5) Assuming there are K cluster centers, calculate the distances from the remaining n-K samples to the cluster centers, and have:

D _r =max{min(D _i1 ,D _i2 ,...D _ik )}>θ×||z ₁ -z ₂ || (9)

Then the corresponding sample x _r is the K+1th cluster center, denoted as z _K+1 ; and this process is repeated continuously until no new cluster center appears;

6) When no new cluster centers appear, assign the samples to each category according to the principle of minimum distance.

Further, the main process of sequence correlation mining in S5 is:

1) The setting of the minimum support and minimum confidence parameters; the confidence and support thresholds are the basis for determining the sequence association and frequent itemsets, and the minimum support thresholds for frequent-1 and frequent-2 itemsets are minsup ₁ and minsup ₂ , the minimum confidence threshold in sequence association mining is mincon;

其中

两序列对应的所有符号类别为：{A₁,A₂,…,A_CA}和{B₁,B₂,…,B_CB}，基于Apriori算法的基本思想，通过对事务集的两阶段扫描，得到序列的频繁项集；根据式(10)计算序列中每个符号的置信度：2) Generation of frequent itemsets; use the two symbolized sequences after summarization as transaction sets, denoted as

in

All symbol categories corresponding to the two sequences are: {A ₁ ,A ₂ ,…,A _CA } and {B ₁ ,B ₂ ,…,B _CB }, based on the basic idea of the Apriori algorithm, through two-stage scanning of the transaction set , get the frequent itemsets of the sequence; calculate the confidence of each symbol in the sequence according to formula (10):

In the formula, N ^t represents the number of transaction sets, that is, the number of elements in the sequence, and the support degree represents the proportion of items in the transaction set. When excavating frequent-1 itemsets, the items whose support degree is greater than minsup ₁ are selected. Divide into a set of frequent-1 itemsets;

Denote the sets of frequent-1 itemsets of two sequences in association mining as P _A and P _B respectively. According to the index parameters, the items in the sets are paired in pairs to form 2-itemsets of the form (P _Ai , P _Bi ), Calculate the support degree of each item in the 2-item set, and divide the items whose support degree is greater than minsup ₂ into frequent-2 itemsets, denoted as {P _A , P _B } _freq ;

3) Mining of sequence associations; all sequences are combined in pairs, and the support of items in the frequent-2 item set and the confidence between the corresponding association mining sequences are counted respectively;

According to formula (11), the support degrees of all frequent-2 itemsets between the two index parameters are accumulated, and the support degrees of these two parameter sequences in all multivariate sequences are counted;

σ(X ^A )=sum(σ(P _A )) (12)

σ(X ^B )=sum(σ(P _B )) (13)

Where m=CA+CB, is the total number of line segment categories divided after the cluster analysis of the two sequences; at the same time, the minimum support threshold at the index sequence level is minsup ₃ , if the support at the parameter index level is greater than the set threshold, Then calculate the confidence degree con(X ^A →X ^B ) of the combination of symbolic itemsets in the two sequences, as shown in formula (14):

When the confidence is greater than the set minimum confidence threshold, the association rule X ^A → X ^B is retained, the confidence is used to describe the strength of the association between the two indicators, and it is determined that there is a strong correlation between the two indicators.

Further, the improved particle swarm optimization support vector regression in S7 is mainly: for the vacant numerical points caused by the deletion of outliers, use the improved particle swarm optimization support vector regression algorithm to repair; the main steps are as follows:

其中元素表达为

1) Specify the number of variables m, and generate N m-dimensional particles in the space of feasible solutions. S ^t is the t-th generation particle in the iteration, where the elements are

where the elements are expressed as

2) Determine the inertia weight, and its specific expression is:

Among them, w _a and w _z represent the maximum and minimum values of inertia weights, f, f _z , and f _pj respectively represent the fitness value of the particle, the minimum fitness value of all particles, and the average fitness value of all particles.

3) Divide the type of particle population, and express the distance between each particle by Euclidean distance:

Define a standard distance:

In the formula, r is the dividing radius, and the density c _i of the i particle is calculated:

n _i is the number of particles in the i particle community, and N is the number of particles in the generated solution set.

4) The particles initialize the two learning factors μ ₁ and μ ₂ of the algorithm according to the category of the population they belong to; when the particle density c _i is greater than a certain threshold, the update method is:

When the particle density c _i is less than a certain threshold, the update method is:

The beneficial effects of the present invention are:

Through sequence segmentation and correlation analysis algorithm, the correlation between different indicators of transformer oil chromatography is mined, the abnormal points in the transformer DGA online monitoring data are identified, and these abnormal points are repaired according to the regression algorithm, which effectively improves the transformer DGA online detection data processing. speed.

Description of drawings

Fig. 1 is the flow chart of the method of the present invention;

Fig. 2 is the sequence segmentation algorithm flow chart;

Fig. 3 is the solution flow chart of the improved particle swarm algorithm;

Figure 4 is a hydrogen index fitting comparison diagram;

Fig. 5 is the methane index fitting comparison chart;

Figure 6 is the abnormal point detected by the hydrogen and methane sequence;

Fig. 7 is a data repair result diagram;

Detailed ways

A method for processing transformer DGA on-line monitoring data of the present invention will be described in detail below with reference to the embodiments and the accompanying drawings.

A method for processing transformer DGA online monitoring data, as shown in Figure 1, includes the following steps:

S1. Import of DGA online data and basic parameter setting of sliding window algorithm: The significance of online monitoring data lies in the real-time reflection of equipment indicators. After long-term operation of equipment, the scale of online data sets is generally relatively large. The analysis is complex and unfeasible, and the online data is time-sensitive, that is, when analyzing a sampling point, the closer the sampling point is to the changed point, the greater the significance of the analysis, and vice versa. The invention introduces the idea of sliding window, uses a window of length L to intercept the online data set, and analyzes the data in the window to reduce the complexity of the process.

个数据窗口，导出所有窗口中的数据，构成待分析数据集DS_i，i∈n。S2. Traverse the online dataset by sliding the window with a certain step size: set the sliding step size to l, drag the window to slide on the overall dataset until all data is traversed; let the length of the online dataset be L ₁ , and get the result after traversing

There are data windows, and the data in all windows are exported to form the data set DS _i to be analyzed, i∈n.

S3. Piecewise linearization of sequence data: since online data is usually a numerical variable, it is not suitable for correlation mining of sequence data; the present invention proposes a piecewise linearization algorithm of sequence data, which will Quantitative points are combined to form multiple sets of data points; the standard of grouping of data points is that the error between the line segment fitted by all points and the actual data point is less than the threshold, and the slope of the line segment and the line segment span are used to characterize The fitted line segment.

S4. Build a model describing the similarity of different line segments: build a similarity model based on the slope and span of the line segment, and use the improved K-means clustering algorithm based on the maximum and minimum distances to classify the line segments, and assign symbols to the same type of line segments, complete Symbolization of sequence data.

S5. Mining the correlation between different sequences: Based on the idea of Apriori algorithm, set the minimum confidence and support, mine the frequent itemsets existing between different sequences, and quantify the correlation between different sequences.

S6. Extract and screen out the abnormal values existing in the DGA online monitoring data: According to the strength of the correlation between the sequences, the types of abnormal values existing in the judgment data are separated, and the data of different abnormal patterns are separated.

S7. Improve particle swarm optimization support vector regression: define the distance between particle solution sets, calculate the density of different particles based on the distance, and define the update method of particles according to the density, so as to improve the solution speed of the algorithm and the diversity of solutions ; Use algorithms to optimize key parameters of support vector regression, improve data regression accuracy, and complete DGA online data processing.

The object studied by the method of the present invention is the DGA online monitoring data of a certain main transformer equipment.

As shown in Figure 2, the specific steps of the piecewise linearization algorithm for sequence data proposed in S3 are:

1) For the online monitoring data of equipment indicators similar to DGA, its essence can be regarded as the state indicator values collected one by one according to a certain time interval sequence. It can be seen that the data has a strong time attribute and can be equivalent to time series data.

2) For the time series X _K ={x ₁ ,x ₂ ,...,x _k }, use a window of length L (L<k) to intercept data points, and for the data in the intercepted window, based on the idea of sliding windows, to The data points contained therein are fitted with a piecewise linear fit.

3) Take the first data point in the window as the fitting starting point of the initial line segment, let this point be _xi , and assuming that the fitting end point of the initial line segment is _xi+m (m>1), use this m+1 data Points fit as a line segment.

4) Then for such a line segment, it can be expressed by the following formula:

my-(X _i+m-1 -X _i )X-(m-1)X _i +X _i+m-1 =0 (2)

Use the distance from the actual data point to the fitting line segment as the fitting error to improve the fitting accuracy of the fitting line segment to the actual value point; The fitted overall error ER for this line segment:

5) Set the fitting error threshold to ER _r , if ER < ER _r , it means that the line segment can continue to add fitting points, let m=m+1, and repeat the above steps; if there is ER > ER _r , then determine The line segment cannot be fitted, save the fitting end point of the current line segment as X _end =X _i+m-1 , record its data sampling time, and then return to step 3), reset the parameter m, and use the current fitting end point as the lower The fitting starting point of a line segment performs the next part of data fitting until all data points in the series are fitted.

The main steps to build a similarity model in S4 and perform cluster analysis based on this model are:

的标准化操作。1) Since there are certain order of magnitude differences between different indicators in DGA online monitoring, it is first necessary to form the attributes of all line segments existing in the same sequence as

standardized operation.

2) In cluster analysis, it is necessary to establish a standard for measuring the similarity of line segments; DGA online data reflects the real-time indicators of equipment, and the change trend and shape of parameters can best reflect the change of equipment operating status. Therefore, when establishing a measurement line segment In the similarity model, different attributes of line segments need to be considered differently. The present invention extracts two key parameters, the slope and the span of the line segment, and uses the Euclidean distance to describe the similarity between the line segments, in which the different attributes of the line segments are expressed in the form of weights. The degree of consideration; the established line segment similarity model is as follows:

3) Based on the above-mentioned line segment similarity model, clustering analysis is performed on the line segment set using the improved K-means algorithm based on the maximum and minimum distances, and the similar line segments are divided into the same category.

The main steps of the improved K-means algorithm based on the maximum and minimum distance in S4 are:

1) The maximum and minimum distances are also based on the Euclidean distance, which differs from the K-means algorithm in that it takes the object as far away as possible as the cluster center; for the sample set, a proportional coefficient θ (0<θ<1) is given. , any sample in the sample set _sn is taken as the initial cluster center, denoted as z ₁ ;

2) arbitrarily take the sample farthest from z ₁ in the remaining n-1 samples as the second cluster center, denoted as z ₂ ;

3) Calculate the distance between the remaining n-2 samples and z ₁ and z ₂ , and find the minimum value, namely:

D _ij =||x _i -z _j ||,j=1,2 (6)

D _i =min(D _i1 ,D _i2 ),i=1,2,...,n (7)

4) If

D _i =max{D _i }>θ×||z _i -z ₂ || (8)

Then select the corresponding sample _si as the third cluster center z ₃ ;

5) Assuming there are K cluster centers, calculate the distances from the remaining n-K samples to the cluster centers, and have:

D _r =max{min(D _i1 ,D _i2 ,...D _ik )}>θ×||z ₁ -z ₂ || (9)

Then the corresponding sample x _r is the K+1th cluster center, denoted as z _K+1 ; and this process is repeated continuously until no new cluster center appears;

6) When no new cluster centers appear, assign the samples to each category according to the principle of minimum distance. The advantage of the improved K-means clustering algorithm based on the maximum and minimum distance is that it ensures that the cluster centers are consistent in each clustering analysis, removes the randomness of the traditional K-means algorithm to select the clustering centers, and can effectively improve the efficiency of clustering analysis. Accuracy and Speed.

The main process of sequence correlation mining in S5 is as follows:

1) The setting of the minimum support and minimum confidence parameters; the confidence and support thresholds are the basis for judging the sequence association and frequent itemsets. Appropriate threshold parameters are beneficial to enhance the credibility of the association relationship, and record frequent -1 and frequent The minimum support thresholds of -2 itemsets are minsup ₁ and minsup ₂ , and the minimum confidence threshold in sequence association mining is mincon.

其中

两序列对应的所有符号类别为：{A₁,A₂,…,A_CA}和{B₁,B₂,…,B_CB}，基于Apriori算法的基本思想，本发明通过对事务集的两阶段扫描，得到序列的频繁项集；根据式(10)计算序列中每个符号的置信度：2) Generation of frequent itemsets; use the two symbolized sequences after summarization as transaction sets, denoted as

in

All symbol categories corresponding to the two sequences are: {A ₁ , A ₂ ,...,A _CA } and {B ₁ ,B ₂ ,...,B _CB }. Based on the basic idea of the Apriori algorithm, the present invention uses two Step scan to get the frequent itemsets of the sequence; calculate the confidence of each symbol in the sequence according to formula (10):

In the formula, N ^t represents the number of transaction sets, that is, the number of elements in the sequence, and the support degree represents the proportion of items in the transaction set. When excavating frequent-1 itemsets, the items whose support degree is greater than minsup ₁ are selected. Divide into sets of frequent-1 itemsets.

Denote the sets of frequent-1 itemsets of two sequences in association mining as P _A and P _B respectively. According to the index parameters, the items in the sets are paired in pairs to form 2-itemsets of the form (P _Ai , P _Bi ), Calculate the support degree of each item in the 2-item set, and divide the items whose support degree is greater than minsup ₂ into frequent-2 itemsets, denoted as {P _A , P _B } _freq ;

3) Mining of sequence associations; all sequences are combined in pairs, and the support of items in the frequent-2 item set and the confidence between the corresponding association mining sequences are counted respectively;

According to formula (11), the support degrees of all frequent-2 itemsets between the two index parameters are accumulated, and the support degrees of these two parameter sequences in all multivariate sequences are counted;

σ(X ^A )=sum(σ(P _A )) (12)

σ(X ^B )=sum(σ(P _B )) (13)

Where m=CA+CB, is the total number of line segment categories divided after the cluster analysis of the two sequences; at the same time, the minimum support threshold at the index sequence level is minsup ₃ , if the support at the parameter index level is greater than the set threshold, Then calculate the confidence degree con(X ^A →X ^B ) of the combination of symbolic itemsets in the two sequences, as shown in formula (14):

When the confidence is greater than the set minimum confidence threshold, the association rule X ^A → X ^B is retained, the confidence is used to describe the strength of the association between the two indicators, and it is determined that there is a strong correlation between the two indicators.

The improved particle swarm optimization support vector regression in S7 mainly includes: the present invention proposes an improved particle swarm optimization support vector regression algorithm to repair the vacant numerical points caused by the deletion of outliers. As shown in Figure 3, the main steps are as follows:

其中元素表达为

1) Specify the number of variables m, and generate N m-dimensional particles in the space of feasible solutions. S ^t is the t-th generation particle in the iteration, where the elements are

where the elements are expressed as

2) Determine the inertia weight, and its specific expression is:

Among them, w _a and w _z represent the maximum and minimum values of inertia weights, f, f _z , and f _pj respectively represent the fitness value of the particle, the minimum fitness value of all particles, and the average fitness value of all particles.

3) Divide the type of particle population, and express the distance between each particle by Euclidean distance:

Define a standard distance:

In the formula, r is the dividing radius, and the density c _i of the i particle is calculated:

n _i is the number of particles in the i particle community, and N is the number of particles in the generated solution set.

4) Two learning factors μ ₁ and μ ₂ of the initialization algorithm; when the particle density c _i is greater than a certain threshold, the update method is:

When the particle density c _i is less than a certain threshold, the update method is:

Specific examples are given below:

A method for processing transformer DGA online monitoring data, the steps are as follows:

S1. Sliding window processing of data sets: After years of operation of power transformers, the DGA online monitoring data of power transformers usually have a large scale, and processing the entire data set at the same time usually increases the complexity of the algorithm and the operating pressure of the server, which is feasible. It has low performance; a DGA online data processing method based on the idea of sliding window is proposed, a data window of length L is established, and the data is intercepted in the data set using this window.

个数据窗口，将所得的窗口数据导出，得到待处理的数据窗口集合{DS_i},i∈n，数据处理将以数据窗口作为分析的基本单位。S2. Intercept the data set according to a certain step size: with a step size of l, drag the data window to slide in the online monitoring data set of length L ₁ , and the intercepted data can be obtained.

A data window is obtained, and the obtained window data is derived to obtain a set of data windows to be processed {DS _i }, i ∈ n. The data processing will take the data window as the basic unit of analysis.

S3. Piecewise linearization of the sequence data in the window: For the intercepted data window W _i , extract the corresponding sequence data according to the DGA monitoring indicators. This example mainly studies two types of gases, H ₂ and CH ₄ in the DGA. Therefore, two corresponding sequences can be obtained in the data window Wi, and the sequence is _piecewise linearized.

S4. Cluster analysis of line segment sets: For the line segment set expressed in the form of an array, this example uses the Euclidean distance method to establish a model ds _ij describing the similarity of the line segments based on the relevant parameters in it, and according to this similarity model, use the maximum The minimum distance improved K-means clustering algorithm performs cluster analysis on the line segment set, merges the line segments with high similarity into one category, and assigns a symbol to each category line segment to complete the symbolization of the sequence data.

S5. Correlation mining between sequences: For the two sequences that complete the summarization operation, based on the idea of the Apriori algorithm, the minimum support threshold minsup _i at different levels and the minimum confidence threshold mincon at the index level are set. The frequent itemsets that exist between the sequences are continuously mined, and the relationship between the indicators is finally determined.

S6. Extraction and screening of abnormal data based on association relationship: According to the association relationship between sequences, the invalid abnormal data existing in the sequence is filtered and extracted.

S7. Data repair: use the improved particle swarm optimization support vector regression to repair the DGA online monitoring data, and complete the processing of the DGA online data.

Taking the hydrogen and methane gas indicators in the DGA historical online monitoring data of a main transformer as the research object, the method proposed in the present invention is used to perform piecewise linearization fitting on the above window sequence data. It should be noted here: due to different indicators The order of magnitude of the data is different, so when using the method proposed by the present invention to perform piecewise linearization fitting, an appropriate fitting error threshold should be selected for different index data. The specific fitting results of each index data are shown in Figures 4 and 4. shown in Figure 5.

The fitting result proves the feasibility of the online data piecewise linearization algorithm proposed by the present invention, the fitting error of each line segment is less than the set fitting error threshold, and the fitted line segment can better reflect the fitting interval The changing trend of online data points, the validity of the algorithm is verified.

Sequence correlation mining: After obtaining the corresponding frequent itemsets, the method proposed in the present invention is used to analyze the correlation between the two indicators, and the support degree is convenient for the confidence degree to express the strength of the correlation relationship, which is supported by H ₂ →CH ₄ The degree and confidence are 0.5050 and 0.6804 respectively, which are both greater than the set minimum relevant threshold, indicating that the rule is a strong association rule, indicating that there is a strong correlation between hydrogen and methane indicators. The detection results are shown in Figure 6. The result of repairing DGA online data is shown in Figure 7.

It can be seen that after the screened out data points are repaired by the method of the present invention relying on several other characteristic gases, all the values return to the normal level, and the online data is effectively cleaned.

Claims (6)

1. A processing method of transformer DGA online monitoring data is characterized in that: the method comprises the following steps:

s1, sliding window processing of the data set: introducing a sliding window idea, and intercepting an online data set by using a window with the length of L;

s2, traversing the online data set by sliding a window with a certain step size: setting the sliding step length as l, dragging the window to slide on the whole data set until all data are traversed; let the length of the online data set be L₁After traversal, get

A data window, deriving the data in all windows to form a data set DS to be analyzed_i，i∈n；

S3, piecewise linearization of sequence data: providing a piecewise linearization algorithm of sequence data, and combining a variable number of points in online data to form a multi-group data point set; the grouping of data points is normalized in that the error between the line segment fitted to all points and the actual data points is less than a threshold, and the slope and line segment span of the line segment used characterize the fitted line segment;

s4, constructing a model for describing the similarity of different line segments: constructing a similarity model based on the slope and span of the line segments, classifying the line segments by using a K-means clustering algorithm improved based on the maximum and minimum distances, giving symbols to the line segments of the same class, and completing the symbolization of sequence data;

s5, mining the relevance among different sequences: based on the idea of Apriori algorithm, setting minimum confidence and support degree, mining frequent item sets existing among different sequences, and quantifying the relevance among different sequences;

s6, extracting and screening abnormal values existing in DGA online monitoring data: according to the strength of the correlation among the sequences, separating data of different abnormal modes from the abnormal numerical value types in the judged data;

s7, improving particle swarm optimization support vector regression: defining the distance between the particle solution sets, dividing different particle categories based on the distance, and defining a particle updating mode; and optimizing the key parameters supporting vector regression by using an algorithm to complete the processing of DGA online data.

2. The processing method of the on-line monitoring data of the DGA of the transformer as claimed in claim 1, wherein: the specific steps of the piecewise linearization algorithm of the sequence data set forth in S3 are:

1) for the online monitoring data of the equipment indexes similar to DGA, equivalent to time sequence data;

2) for time series X_K＝{x₁,x₂,…,x_kIntercepting data points by a window with the length of L (L < k), and carrying out piecewise linear fitting on the data points contained in the intercepted window on the basis of the idea of a sliding window;

3)the first data point in the window is taken as the fitting starting point of the initial line segment, and the point is taken as x_iAssuming that the fitting end point of the initial line segment is x_i+m(m > 1), fitting the m +1 data points to a line segment;

4) then for such a line segment, it is expressed by the following equation:

my-(X_i+m-1-X_i)X-(m-1)X_i+X_i+m-1＝0 (2)

taking the distance from the actual data point to the fitting line segment as a fitting error; calculating the distances from all actual data points in the step length of the fitted line segment to the line segment, and taking the sum of the distances as the overall fitting error ER of the line segment:

5) setting the fitting error threshold to ER_rIf ER < ER_rIf so, the line segment can still continue to increase the fitting point, let m be m +1, and repeat the above steps; if ER > ER is present_rIf the line segment can not be fitted, the fitting end point of the current line segment is stored as X_end＝X_i+m-1And recording the data sampling time, returning to the step 3), resetting the parameter m, and fitting the next part of data by taking the current fitting endpoint as the fitting starting point of the next line segment until all data points in the sequence are fitted.

3. The processing method of the on-line monitoring data of the DGA of the transformer as claimed in claim 1, wherein: the similarity model is constructed in S4, and the main steps of cluster analysis based on the similarity model are as follows:

1) form all line segment attributes present in the same sequence as

The standardization operation of (2);

2) during cluster analysis, establishing a standard for measuring the similarity of the line segments; extracting two key parameters of the slope and the span of the line segment, describing the similarity between the line segments by using Euclidean distance, and expressing the consideration degree of different attributes of the line segment in a weight mode; the established line segment similarity model is shown as the following formula:

3) based on the line segment similarity model, the line segment set is subjected to clustering analysis by using a K-means algorithm improved based on the maximum and minimum distances, and similar line segments are divided into the same category.

4. The method for processing the on-line monitoring data of the DGA of the transformer as claimed in claim 3, wherein: in S4, the improved K-means algorithm based on the maximum and minimum distance mainly comprises the following steps:

1) the maximum and minimum distances are also based on Euclidean distances, and the difference between the maximum and minimum distances and the K-means algorithm is that an object with a maximum distance is taken as a clustering center; for the sample set, a proportion coefficient theta (0 < theta < 1) is given, and the sample set s is taken arbitrarily_nIs the initial clustering center, denoted as z₁；

2) Optionally taking the distance z of the remaining n-1 samples₁The farthest sample is the second cluster center, denoted as z₂；

3) Calculate the remaining n-2 samples and z₁And z₂And finding the minimum value among them, namely:

D_ij＝||x_i-z_j||,j＝1,2 (6)

D_i＝min(D_i1,D_i2),i＝1,2,…,n (7)

4) if it is

D_i＝max{D_i}＞θ×||z_i-z₂|| (8)

Then select the corresponding sample s_iAs a third cluster center z₃；

5) Assuming that there are K cluster centers, the distance from the remaining n-K samples to the cluster centers is calculated, and the following steps are carried out:

D_r＝max{min(D_i1,D_i2,…D_ik)}＞θ×||z₁-z₂|| (9)

then the corresponding sample x_rIs the K +1 cluster center and is marked as z_K+1(ii) a The process is continuously circulated until no new clustering center appears;

6) when no new cluster center is present, the samples are assigned to each class according to the minimum distance principle.

5. The processing method of the on-line monitoring data of the DGA of the transformer as claimed in claim 1, wherein: the main process of sequence association mining in S5 is as follows:

1) setting parameters of minimum support degree and minimum confidence degree; the confidence coefficient and the support threshold are the basis for judging the sequence association and the frequent item set, and the minimum support threshold of the frequent-1 and frequent-2 item sets is recorded as min₁And min₂The minimum confidence threshold in the sequence association mining is mincon;

2) generating a frequent item set; using the summed two-signed sequence as a transaction set, denoted

Wherein

All symbol categories corresponding to the two sequences are: { A₁,A₂,…,A_CAAnd { B }₁,B₂,…,B_CBObtaining a frequent item set of the sequence by scanning the transaction set in two stages based on the basic idea of an Apriori algorithm; the confidence for each symbol in the sequence is calculated according to equation (10):

in the formula N^tRepresenting the number of transaction sets, namely the number of elements in the sequence, representing the proportion of items in the transaction sets by the support degree, and when a frequent-1 item set is mined, the support degree is greater than the minimum₁The items of (a) are divided into a set of frequent-1 item sets;

the collection of frequent-1 item sets of two sequences in the association mining is recorded as P_A、P_BPairing the items in the set according to the index parameters to form the form (P)_Ai,P_Bi) Form 2-item set, calculating the support degree of each item in the 2-item set, and enabling the support degree to be greater than min₂Is divided into a frequent-2 item set, denoted as { P_A,P_B}_freq；

3) Mining sequence relevance; combining all the sequences pairwise, and respectively counting the support degree of the frequent-2 item concentrated items in the sequences and the confidence degree between the corresponding association mining sequences;

accumulating the support degrees of all frequent-2 item sets between two index parameters according to the formula (11), and taking the accumulated support degrees as the support degree counts of the two parameter sequences in all multivariate sequences;

σ(X^A)＝sum(σ(P_A)) (12)

σ(X^B)＝sum(σ(P_B)) (13)

wherein m is CA + CB, and is the total number of line segment categories divided after the two-sequence clustering analysis; while recording the minimum of the index sequence levelSupport degree threshold value is minisu₃If the support degree of the parameter index level is larger than the set threshold value, calculating the confidence degree con (X) of the combination of the symbol item sets in the two sequences^A→X^B) As shown in formula (14):

when the confidence is greater than the set minimum confidence threshold, the association rule X is reserved^A→X^BAnd describing the strength of the association between the two indexes by using the confidence coefficient, and judging that the two indexes have strong association.

6. The processing method of the on-line monitoring data of the DGA of the transformer as claimed in claim 1, wherein: the improved particle swarm optimization support vector regression in S7, which is mainly: for vacant numerical points caused by deletion of abnormal values, repairing the vacant numerical points by using a support vector regression algorithm for improving particle swarm optimization; the method mainly comprises the following steps:

1) defining the number m of variables, generating N m-dimensional particles in a feasible solution space, S^tIs the t-th generation particle in the iteration, wherein the element is

Wherein the elements are expressed as

2) Determining an inertia weight, wherein the inertia weight represents the inheritance degree of the particle to the speed during the last iteration; the specific expression is as follows:

wherein, w_aAnd w_zRepresenting maximum and minimum values of inertial weight, f_z,f_pjRespectively representThe fitness value of the particle, the minimum fitness value of all the particles and the average fitness value of all the particles;

3) the types of the particle populations are divided, and the distance between each particle is expressed by Euclidean distance:

define a standard distance:

wherein r is a dividing radius, and the density c of i particles is calculated_i：

n_iThe number of particles in the particle swarm is i, and N is the number of generated solution concentration particles;

4) two learning factors mu of the initialization algorithm₁、μ₂(ii) a When the particle density c_iWhen the value is larger than a certain threshold value, the updating mode is as follows:

when the particle density c_iWhen the value is less than a certain threshold value, the updating mode is as follows:

Images