Transformer oil chromatographic false alarm feature extraction and comprehensive identification method

Abstract

The application discloses a transformer oil chromatogram false alarm feature extraction and comprehensive identification method, comprising the following steps: screening typical false alarms and real alarm data in the online monitoring historical data of oil chromatogram; setting a sliding time window according to the screened data, and constructing a standardized alarm history library; based on Shapelets transformation, extracting features of each state quantity in the standardized alarm history library; based on the extracted feature data, using a double-layer LightGBM algorithm to comprehensively judge the authenticity of the transformer oil chromatogram alarm of all state quantities. By extracting the time sequence features of the oil chromatogram false alarm data section, combining the double-layer regressor, the real-time and comprehensive discrimination of the false alarms of eight gas components is realized, which is suitable for power companies in the process of transformer equipment management, reduces the number of online chromatogram false alarms, reduces the work burden of the grassroots, and improves the quality and efficiency of maintenance.

Description

Technical Field

[0001] This invention belongs to the field of transformer oil chromatography alarm, and relates to the technology of feature extraction and recognition of false alarms in transformer oil chromatography, specifically to a method for feature extraction and comprehensive recognition of false alarms in transformer oil chromatography. Background Technology

[0002] Oil-immersed transformers are prone to aging and decomposition under the influence of factors such as temperature and electric fields, producing dissolved gases in the oil, such as hydrogen, carbon dioxide, and methane. When the transformer becomes abnormal or malfunctions, the gas content increases significantly. With the increasing prevalence of online oil chromatography monitoring devices, remote monitoring of the dissolved gas content in transformer oil allows for timely understanding of the transformer's operating status and the detection of potential faults.

[0003] To detect changes in dissolved gases in oil as early as possible, the power grid company has implemented alarm rules for gases such as acetylene. These rules use standardized thresholds to determine whether an alarm should be issued and, if so, at what level. These gas thresholds are a set of quantitative rules developed by transformer experts based on years of field experience. Theoretically, threshold analysis can determine whether there are any abnormalities in the dissolved gases in the oil of most transformers. However, due to limitations in monitoring environment and data transmission, coupled with inherent flaws in some online monitoring devices, problems such as excessive fluctuations in online chromatograms and inconsistencies with offline chromatograms frequently occur, leading to a proliferation of false alarms from online monitoring devices.

[0004] Therefore, a new technical solution is needed to solve this problem. Summary of the Invention

[0005] Purpose of the invention: In order to overcome the shortcomings of the existing technology, a method for feature extraction and comprehensive identification of false alarms in transformer oil chromatography is provided. This method establishes a series of mathematical models for the false alarm curve of transformer oil chromatography, realizes feature extraction of real alarm and false alarm data, and then realizes real-time identification of false alarm data in oil chromatography.

[0006] Technical Solution: To achieve the above objectives, this invention provides a method for extracting and comprehensively identifying false alarm features from transformer oil chromatography, comprising the following steps:

[0007] S1: Screen typical false alarms and real alarm data from historical data of online oil chromatography monitoring;

[0008] S2: Based on the filtered data, set a sliding time window and construct a standardized alarm history database;

[0009] S3: Feature extraction of each state variable in the standardized alarm history database based on Shapelets transformation;

[0010] S4: Based on the extracted feature data, the authenticity of the transformer oil chromatographic alarm is determined by comprehensively considering all state variables using the dual-layer LightGBM algorithm.

[0011] Furthermore, in step S1, the data is divided according to the manufacturer and product model of the online oil chromatography monitoring device, and the historical false alarm and true alarm data are screened separately. Among them, the true alarm data mainly consists of equipment alarm data segments that have historically experienced actual failures, while the false alarms consist of data segments that have historically experienced typical sudden changes and fluctuations caused by external factors such as environmental and communication interference.

[0012] Furthermore, the alarm data selected in step S1 must also meet the following three requirements:

[0013] The data length should be no less than 20 consecutive monitoring values, while ensuring that it reflects the basic characteristics of alarm data.

[0014] In the selected typical false alarm and real alarm data, each set of alarm data contains eight gas components, namely methane, ethane, acetylene, ethylene, total hydrocarbons, hydrogen, carbon monoxide, and carbon dioxide, which correspond to the attributes of correct alarm (i.e., 0) or false alarm (i.e., 1), respectively.

[0015] Typical false alarm and real alarm data are preprocessed as follows: For time-series data of eight gas components, considering factors such as amplitude and range, a comprehensive false alarm level is labeled, including Level 1, Level 2, Level 3, and Level 4. Among them, Level 1 represents the lowest probability of a false alarm, and Level 4 represents the highest probability of a false alarm.

[0016] Furthermore, the construction of the standardized alarm history database in step S2 includes the construction of the first layer of data and the first layer of data tags, as well as the construction of the second layer of data and the second layer of data tags.

[0017] Furthermore, the method for constructing the first layer of data and the first layer of data labels is as follows: considering the time-series characteristics of online monitoring data, a sliding window is used to process the selected typical false alarms and real alarms, and 18 consecutive chromatographic data (approximately 1 to 6 days) are grouped together as the first layer of data, while retaining the attribute labels of their correct alarms or false alarms as the labels of the first layer of data.

[0018] Furthermore, the construction method of the second layer data and the second layer data labels is as follows: For 18 time series data, the maximum value, standard deviation, average value, coefficient of variation (standard deviation / average value), range value (maximum value - minimum value), and proportion of 0 values ​​are calculated, and the labels of the first layer data are added to form a total of 6 features, which are used as the second layer data; the comprehensive false alarm level marked in step S1 is used as the label of the second layer data, which can form a two-layer standardized alarm history library.

[0019] Furthermore, in step S2, if the number of consecutive monitoring values ​​in the data filtered out in step S1 is greater than 18, multiple sets of standardized alarm data can be constructed by shifting the time window and added to the standardized alarm history database.

[0020] Furthermore, the specific process of feature extraction in step S3 is as follows:

[0021] For the first-level time series data, Shapelets transformation was used to extract features from eight gaseous components: methane, ethane, acetylene, ethylene, total hydrocarbons, hydrogen, carbon monoxide, and carbon dioxide.

[0022] For a single-component gas time series dataset, limiting parameters such as the length *l* and the number *k* of Shapelets subsequences are set. The minimum Euclidean distance between all candidate Shapelets sequences and the time series dataset *T* is calculated, and information gain (IG) is used as the standard for calculating the quality of Shapelets sequences.

[0023] H(T)=-p(X)log[p(X)]-p(Y)log[p(Y)]

[0024]

[0025] In the formula, p(X) and p(Y) are the proportions of data of type X and type Y in the total data, respectively;

[0026] The optimal candidate set of Shapelets for output gas components {S i,1 ,S i,2 ,…,S i,k Regardless of the length of the time series, select the data that corresponds to S. i,j Subsequences of equal length (1≤j≤k) such that they are equal to S i,j The Euclidean distance between them is the smallest, and this Euclidean distance is denoted as d. i,jmin d i,jmin That is, S in the optimal candidate set of Shapelets i,j For the features extracted from this sequence, the optimal Shapelets candidate set can extract k features {d} for any sufficiently long time series. i,1min ,d i,2min ,…,d i,k min}

[0027] Furthermore, the specific process of step S4 is as follows:

[0028] For the first layer of data, based on step S3, the original first layer of data is transformed into first layer feature data, and alarm attribute labels are attached at the same time.

[0029] The LightGBM algorithm is used to regress the first layer of feature data one by one according to the gas components to form the first layer regressor, thereby realizing the calculation of false alarm confidence of single-component gas.

[0030] The LightGBM algorithm is used to perform unified regression training on all data (8×6=48 data points) in the second layer of step S2 to form the second-layer regressor, thereby realizing the comprehensive false alarm level calculation for multi-component gases;

[0031] The sigmoid activation function is used to convert the overall false alarm level into an overall false alarm confidence level, which is used to comprehensively evaluate the probability of transformer oil chromatography false alarms. By setting a threshold (such as 60%), the authenticity of the transformer oil chromatography alarm is judged, and it is decided whether to issue an online chromatography alarm.

[0032] This invention utilizes artificial intelligence algorithms to analyze and extract the temporal characteristics of false alarm curves from online oil chromatography monitoring devices. Combined with classification and detection algorithms, it eliminates or reduces false alarms while ensuring that genuine alarms are issued normally, thereby reducing the workload of frontline teams and achieving both workload reduction and efficiency improvement.

[0033] Beneficial effects: Compared with the prior art, this invention extracts the temporal characteristics of oil chromatography false alarm data segments and combines them with a two-layer regressor to achieve real-time and comprehensive discrimination of false alarms for eight gas components, including methane, ethane, acetylene, ethylene, total hydrocarbons, hydrogen, carbon monoxide, and carbon dioxide. It is applicable to power companies in the process of transformer equipment management, reducing the number of online chromatography false alarms, reducing the workload of grassroots staff, and improving maintenance efficiency. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0035] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0036] like Figure 1 As shown, the present invention provides a method for extracting and comprehensively identifying false alarm features from transformer oil chromatography, comprising the following steps:

[0037] S1: Screen typical false alarms and true alarms from historical data of online oil chromatography monitoring:

[0038] The online oil chromatography monitoring device is installed on an operating transformer and is exposed to a complex electromagnetic environment for extended periods, frequently experiencing problems such as fixed deviations, data jumps, and accuracy distortion. When the measured gas concentration deviates significantly from the normal value, the threshold alarm rules no longer apply, as the output results are often false alarms.

[0039] This invention categorizes data according to the manufacturer and product model of the online oil chromatography monitoring device, and separately filters historical false alarms and true alarms. The true alarm data mainly consists of alarm data segments from equipment that experienced actual faults in the past, while false alarms consist of data segments from historical data segments that showed sudden changes or fluctuations caused by external factors such as environmental and communication interference.

[0040] The selected alarm data must also meet the following three requirements:

[0041] The data length should be no less than 20 consecutive monitoring values, while ensuring that it reflects the basic characteristics of alarm data;

[0042] In the selected typical false alarm and real alarm data, each set of alarm data contains eight gas components, namely methane, ethane, acetylene, ethylene, total hydrocarbons, hydrogen, carbon monoxide, and carbon dioxide, which correspond to the attributes of correct alarm (i.e., 0) or false alarm (i.e., 1), respectively.

[0043] Typical false alarm and real alarm data are preprocessed as follows: For time-series data of eight gas components, considering factors such as amplitude and range, a comprehensive false alarm level is labeled, including Level 1, Level 2, Level 3, and Level 4. Among them, Level 1 represents the lowest probability of a false alarm, and Level 4 represents the highest probability of a false alarm.

[0044] S2: Based on the filtered data, set a sliding time window and construct a standardized alarm history database:

[0045] The construction of a standardized alarm history database includes the construction of first-level data and first-level data tags, as well as second-level data and second-level data tags;

[0046] The method for constructing the first layer of data and the first layer of data labels is as follows: Considering the time-series characteristics of online monitoring data, a sliding window is used to process the typical false alarm and real alarm data selected. 18 consecutive chromatographic data (about 1 to 6 days) are grouped together as the first layer of data, while retaining the attribute labels of the correct alarm or false alarm as the labels of the first layer of data.

[0047] The method for constructing the second layer of data and the second layer of data labels is as follows: For 18 time series data, calculate the maximum value, standard deviation, average value, coefficient of variation (standard deviation / average value), range (maximum value - minimum value), and percentage of 0 values. Add the labels of the first layer of data to form a total of 6 features, which are used as the second layer of data. The comprehensive false alarm level marked in step S1 is used as the label of the second layer of data, which can form a two-layer standardized alarm history library.

[0048] If the number of consecutive monitoring values ​​in the data filtered in step S1 is greater than 18, multiple sets of standardized alarm data can be constructed by shifting the time window and added to the standardized alarm history database.

[0049] S3: Feature extraction of each state variable in the standardized alarm history database based on Shapelets transformation:

[0050] For the first-level time series data, Shapelets transformation was used to extract features from eight gaseous components: methane, ethane, acetylene, ethylene, total hydrocarbons, hydrogen, carbon monoxide, and carbon dioxide.

[0051] For a single-component gas time series dataset, limiting parameters such as the length *l* and the number *k* of Shapelets subsequences are set. The minimum Euclidean distance between all candidate Shapelets sequences and the time series dataset *T* is calculated, and information gain (IG) is used as the standard for calculating the quality of Shapelets sequences.

[0052] H(T)=-p(X)log[p(X)]-p(Y)log[p(Y)]

[0053]

[0054] In the formula, p(X) and p(Y) represent the proportions of data of type X and type Y in the total data, respectively, and the data is divided into T. x T y Given two datasets, H(T) represents the entropy of dataset T.

[0055] Based on information gain, the optimal candidate set of Shapelets {S} can be calculated. i,1 ,S i,2 ,…,S i,k Regardless of the length of the time series, select the data that corresponds to S. i,j Subsequences of equal length (1≤j≤k) such that they are equal to S i,j The Euclidean distance between them is the smallest, and this Euclidean distance is denoted as d. i,jmin d i,jmin That is, S in the optimal candidate set of Shapelets i,jFor the features extracted from this sequence, the optimal Shapelets candidate set can extract k features {d} for any sufficiently long time series. i,1min ,d i,2min ,…,d i,kmin}

[0056] S4: Based on the extracted feature data, the authenticity of the transformer oil chromatogram alarm is determined by comprehensively considering all state variables using a two-layer LightGBM algorithm.

[0057] For the first layer of data, based on step S3, the original first layer of data is transformed into first layer feature data, and alarm attribute labels are attached at the same time.

[0058] The LightGBM algorithm is used to regress the first layer of feature data one by one according to the gas components to form the first layer regressor, thereby realizing the calculation of false alarm confidence of single-component gas.

[0059] The LightGBM algorithm is used to perform unified regression training on all data (8×6=48 data points) in the second layer of step S2 to form the second-layer regressor, thereby realizing the comprehensive false alarm level calculation for multi-component gases;

[0060] The sigmoid activation function S(x) is used to convert the overall false alarm level into an overall false alarm confidence level, which is used to comprehensively evaluate the probability of transformer oil chromatography false alarms. By setting a threshold (such as 60%), the authenticity of the transformer oil chromatography alarm is judged, and it is decided whether to issue an online chromatography alarm.

[0061]

[0062] Based on the above, in order to verify the effectiveness of the present invention, this embodiment uses historical data from an online oil chromatography monitoring device of a 500kV transformer in Jiangsu Province as an example for a calculation test, as follows:

[0063] First, based on the manufacturer and product model of the online monitoring device for this transformer, historical false alarm and real alarm data were screened to construct a two-layer standardized alarm history database. Then, features were extracted using Shapelets transformation to construct a two-layer LightGBM false alarm evaluation regression engine to achieve a comprehensive evaluation of the false alarm level for this type of device.

[0064] In July 2023, the device detected one alarm data. The alarms were arranged according to a standardized alarm history database using a sliding time window, as shown in Table 1:

[0065] Table 1

[0066]

[0067]

[0068] The false alarm confidence levels of the eight gas components were calculated using the first-level regressor, and the results are shown in Table 2.

[0069] Table 2

[0070] False Alarm Confidence 98.21 36.85 94.04 87.39 79.67 84.21 21.50 25.83

[0071] The maximum value, standard deviation, mean, coefficient of variation (standard deviation / mean), range (maximum value - minimum value), and percentage of zero values ​​were calculated for each of the eight gases. These, along with the labels from the first-layer data, resulted in a total of six major features and 48 sub-features, which were then used as the second-layer data. This data was then imported into the second-layer regressor, and a sigmoid activation function was applied to output the overall false alarm confidence score. The results are shown in Table 3.

[0072] Table 3

[0073]

[0074]

[0075]

[0076] As shown in Table 3, the overall false alarm confidence level in this embodiment is 84.15%, and the threshold set in this embodiment is 60%, so it is judged as a false alarm.

[0077] In the 30 days following the initial alarm, a total of 14 alarms occurred, of which 11 were false alarms and 3 were genuine alarms, as determined by experts. Based on the method of this invention, a total of 10 false alarms were detected, and no genuine alarms were misclassified as false alarms, achieving a false alarm identification rate of 91% for online chromatography.

Claims

1. A method for extracting and comprehensively identifying false alarm features from transformer oil chromatography, characterized in that, Includes the following steps: S1: Screen typical false alarms and real alarm data from historical data of online oil chromatography monitoring; S2: Based on the filtered data, set a sliding time window and construct a standardized alarm history database; S3: Feature extraction of each state variable in the standardized alarm history database based on Shapelets transformation; S4: Based on the extracted feature data, the authenticity of the transformer oil chromatographic alarm is determined by comprehensively considering all state variables using the dual-layer LightGBM algorithm; The specific process of feature extraction in step S3 is as follows: For the first-level time series data, Shapelets transformation was used to extract features from eight gaseous components: methane, ethane, acetylene, ethylene, total hydrocarbons, hydrogen, carbon monoxide, and carbon dioxide. For a single-component gas time series dataset, limiting parameters are set, and the minimum Euclidean distance between all candidate Shapelets sequences and the time series dataset T is calculated. Information gain (IG) is used as the standard for calculating the quality of Shapelets sequences. ； ； In the formula, p(X) and p(Y) represent the proportions of data of type X and type Y in the total data, respectively, and the data is divided into T. x T y Given two datasets, H(T) represents the entropy of dataset T; The optimal candidate set of Shapelets for output gas components {S i,1 , S i,2 , … , S i , k Regardless of the length of the time series, select the data that corresponds to S. i,j Subsequences of equal length (1≤j≤k) such that they are equal to S i,j The Euclidean distance between them is the smallest, and this Euclidean distance is denoted as d. i,j min d i,j min That is, S is the candidate set of the optimal Shapelets. i,j For the features extracted from this sequence, the optimal Shapelets candidate set can extract k features {d} for any sufficiently long time series. i,1 min , d i,2 min , …, d i,k min }; The specific process of step S4 is as follows: For the first layer of data, based on step S3, the original first layer of data is transformed into first layer feature data, and alarm attribute labels are attached at the same time. The LightGBM algorithm is used to regress the first layer of feature data one by one according to the gas components to form the first layer regressor, thereby realizing the calculation of false alarm confidence of single-component gas. The LightGBM algorithm is used to perform unified regression training on all the data in the second layer in step S2 to form the second-layer regressor, thereby realizing the comprehensive false alarm level calculation for multi-component gases. The sigmoid activation function is used to convert the overall false alarm level into an overall false alarm confidence level, which is used to comprehensively evaluate the probability of transformer oil chromatography false alarms. By setting a threshold, the authenticity of transformer oil chromatography alarms is judged, and it is decided whether to issue an online chromatography alarm.

2. The method for extracting and comprehensively identifying false alarm features from transformer oil chromatography according to claim 1, characterized in that, In step S1, the typical false alarms and real alarm data selected each contain eight gas components: methane, ethane, acetylene, ethylene, total hydrocarbons, hydrogen, carbon monoxide, and carbon dioxide, which correspond to the attributes of correct alarms or false alarms, respectively.

3. The method for extracting and comprehensively identifying false alarm features from transformer oil chromatography according to claim 2, characterized in that, In step S1, the typical false alarm and real alarm data are preprocessed. Specifically, for the time series data of the eight gas components, the comprehensive false alarm level is labeled, including level 1, level 2, level 3 and level 4. Among them, level 1 represents the lowest probability of false alarm and level 4 represents the highest probability of false alarm.

4. The method for extracting and comprehensively identifying false alarm features from transformer oil chromatography according to claim 1, characterized in that, The construction of the standardized alarm history database in step S2 includes the construction of the first layer of data and the first layer of data tags, as well as the construction of the second layer of data and the second layer of data tags.

5. The method for extracting and comprehensively identifying false alarm features from transformer oil chromatography according to claim 4, characterized in that, The method for constructing the first layer of data and the first layer of data labels is as follows: Considering the time-series characteristics of online monitoring data, a sliding window is used to process the typical false alarm and real alarm data selected. 18 consecutive chromatographic data are grouped together as the first layer of data, while the attribute labels of the correct alarm or false alarm are retained as the labels of the first layer of data.

6. The method for extracting and comprehensively identifying false alarm features from transformer oil chromatography according to claim 5, characterized in that, The construction method of the second layer data and the second layer data label is as follows: For 18 time series data, the maximum value, standard deviation, average value, coefficient of variation, range value, and proportion of 0 values ​​are calculated, and the labels of the first layer data are added to form a total of 6 features, which are used as the second layer data; the comprehensive false alarm level marked in step S1 is used as the label of the second layer data, which can form a two-layer standardized alarm history library.

7. The method for extracting and comprehensively identifying false alarm features from transformer oil chromatography according to claim 6, characterized in that, In step S2, if the number of continuous monitoring values ​​in the data filtered in step S1 is greater than 18, multiple sets of standardized alarm data can be constructed by shifting the time window and added to the standardized alarm history database.

8. The method for extracting and comprehensively identifying false alarm features from transformer oil chromatography according to claim 1, characterized in that, The expression for the sigmoid activation function is: 。

Images