Transformer inspection system and inspection method

Abstract

To provide a system and method for inspecting a state of a transformer without using gas chromatography mass spectrometry every time insulating oil is sampled from the transformer.SOLUTION: An inspection method for an inspection system 1 for a transformer comprises air purification filters 2, 9, a sample container 3, electromagnetic valves A, B, a gas mixing container 4, a smell sensor 5, a pump 6, a device control unit 7, and a maintenance management standard estimation unit 8. The air purification filters remove volatile organic chemicals. The sample container encloses insulating oil sampled from the transformer. The gas mixing container mixes gas discharged from the carrier gas side with gas discharged from the sample container side. Each electromagnetic valve adjusts a valve opening and closing degree according to a valve adjustment command from the device control unit. The pump draws gas from the gas mixing container. The smell sensor outputs a sensor signal corresponding to the gas from the gas mixing container. The maintenance inspection standard estimation unit estimates which section of internal abnormality sections for the transformer the transformer's state corresponds to on the basis of the sensor signal from the smell sensor.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an inspection system and an inspection method used for the maintenance and inspection of a transformer. [Background technology]

[0002] Conventionally, methods have been proposed for inspecting the condition of transformers by detecting abnormalities in the insulating oil used for insulation and cooling. For example, as shown in Non-Patent Document 1, a widely used method is to perform gas analysis in oil to check the amount of gas components contained in the insulating oil and compare it with maintenance management standards to determine the condition of the transformer.

[0003] Furthermore, Patent Document 1 proposes a method for predicting abnormalities that focuses on the concentration of benzyl sulfide, a substance that causes copper sulfide precipitation. Furthermore, Patent Document 2 discloses a method for determining abnormalities in oil-filled electrical equipment by collecting a sample of the insulating oil in the body of a transformer and analyzing the presence or absence of methylvinylacetylene in the sample oil.

[0004] In addition, Non-Patent Document 2 describes the artificial insulation oil gas of a transformer. sense of smell The data measured by the system is used to evaluate whether or not the insulating oil itself has deteriorated. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Electrical Engineering Research Association, Vol. 65, No. 1, 2009 [Non-patent document 2] Sosuke Akagawa, Shintaro Furuno, Yuanchang Liu, Rui Yatabe, Takeshi Onodera, Nobuyuki Fujiwara, Shuichi Takeda, Kiyoshi Toko, "Detection of Deterioration of Insulating Oil Using an Artificial Olfactory System," Institute of Electrical Engineers of Japan, 2020 E Division General Research Meeting [Non-patent document 3] Masamichi Kato, Nobuyuki Ohta, Hidenobu Koide, "Decomposition Gas Behavior of Various Ester-Based Insulating Oils During Local Heating," IEEJ Transactions on Electrical Engineering, Vol. 37, No. 4, pp. 208-214, 2017 [Non-patent document 4] "Porta Test 100A-2 Electrical Insulating Oil Withstand Voltage Test Device" Online, Internet, searched on March 1, 2022 <URL:https: / / manualzilla.com / doc / 6538370 / %E9%9B%BB%E6%B0%97%E7%B5%B6%E7%B8%81%E6%B2%B9%E8%80%90%E9%9B%BB%E5%9C%A7%E8%A9%A6%E9%A8%93%E8%A3%85%E7%BD%AE-%E3%83%9D%E3%83%AB%E3%82%BF%E3%83%86%E3%82%B9%E3%83%88100a-2> [Non-Patent Document 5] Masaru Wakimoto and Kazuki Shigemori, "DGA (Dissolved Gas Analysis) Criteria for Palm Oil-Immersed Transformers," Meiden Jiho, Vol. 372, 2021, No. 3 [Patent documents]

[0006] [Patent Document 1] Patent No. 4623334 [Patent Document 2] Patent No. 5239193 [Patent Document 3] Patent No. 6849791 Summary of the Invention [Problem to be solved by the invention]

[0007] The technique of Non-Patent Document 1 is widely used as a method for inspecting the condition of transformers, but in order to determine the gas components, oil sampled from the transformer must be analyzed using a gas chromatography mass spectrometer.

[0008] This means that gas chromatography must be performed every time oil is sampled, which can take a long time. In particular, if a company does not have its own analytical equipment, it must outsource the analysis of gas components to an analytical company, which takes even more time.

[0009] Furthermore, the methods of Patent Documents 1 and 2, like the method of Non-Patent Document 1, predict and determine abnormalities based on the results of analysis using a gas chromatography mass spectrometer each time oil is sampled from the transformer, so it may take a long time to obtain results.

[0010] On the other hand, the method in Non-Patent Document 2 is capable of evaluating whether or not the insulating oil itself has deteriorated, but it does not reach a detailed evaluation level that would enable the state of the transformer to be judged in accordance with the maintenance management standards, and therefore may lack practicality.

[0011] The present invention has been made to solve such conventional problems, and aims to solve the problem of inspecting the condition of a transformer when inspecting the transformer in accordance with maintenance management standards without using gas chromatography-mass spectrometry every time insulating oil is sampled from the transformer. [Means for solving the problem]

[0012] One aspect of the present invention is a system for inspecting the condition of a transformer, comprising: a sample container for enclosing insulating oil collected from the transformer; a mixing vessel for mixing the gas discharged from the sample vessel side with the gas discharged from the carrier gas side; A sensor signal corresponding to the gas flowing in from the mixing vessel is output. sense of smell A sensor, The aforementioned sense of smell a maintenance and inspection standard estimation unit that analyzes the state of the transformer based on the sensor signal of the sensor and estimates which of the internal abnormality categories of the transformer the abnormality corresponds to; Equipped with The maintenance and inspection standard estimation unit includes a feature extraction unit that extracts feature values ​​from the signal values ​​of the sensor signals; a maintenance and inspection standard identification unit that identifies the internal abnormality classification based on the feature amount; Equipped with the maintenance and inspection standard identification unit identifies the internal abnormality category using parameters previously learned based on a learning sample; The training samples are A pseudo-abnormal oil is prepared according to each aspect of the classification, and a learning mixed oil of a group of patterns is prepared by blending the pseudo-abnormal oil with insulating oil, based on the results of measurement by the olfactory sensor. The transformer condition is diagnosed by gas-in-oil analysis based on the amount of gas components in the learning mixed oil, and the diagnosis result is set as the internal abnormality classification. It is characterized by the following.

[0013] Another aspect of the present invention is a sample container for enclosing insulating oil collected from the transformer; a mixing vessel for mixing the gas discharged from the sample vessel side with the gas discharged from the carrier gas side; A sensor signal corresponding to the gas flowing in from the mixing vessel side is output. sense of smell A sensor, 1. A method for inspecting the condition of a transformer performed by a system comprising: a maintenance and inspection standard estimation step of analyzing the state of the transformer based on the sensor signal of the olfactory sensor and estimating which of the internal abnormality categories of the transformer the state corresponds to; The maintenance and inspection standard estimation step includes: a feature extraction step of extracting a feature from the signal value of the sensor signal; a maintenance and inspection standard identification step of identifying the internal abnormality classification based on the feature amount; and In the maintenance inspection criteria identification step, identifying the internal anomaly category using pre-trained parameters based on training samples; The training samples are A pseudo-abnormal oil is prepared according to each aspect of the classification, and a learning mixed oil of a group of patterns is prepared by blending the pseudo-abnormal oil with insulating oil, based on the results of measurement by the olfactory sensor. The transformer condition is diagnosed by gas-in-oil analysis based on the amount of gas components in the learning mixed oil, and the diagnosis result is set as the internal abnormality classification. It is characterized by the following. [Effects of the Invention]

[0014] According to the present invention, when inspecting a transformer in accordance with maintenance management standards, it is possible to inspect the condition of the transformer without using gas chromatography-mass spectrometry each time insulating oil is sampled from the transformer. [Brief explanation of the drawings]

[0015] [Figure 1] FIG. 1 is a configuration diagram of a transformer inspection system according to a first embodiment. [Figure 2] FIG. [Figure 3] FIG. 10 is a schedule diagram of the data measurement operation. [Figure 4] 10 is a graph showing an example of sensor signal data. [Figure 5] FIG. 2 is a block diagram of the maintenance management standard estimation unit of the same. [Figure 6] 5 is a graph showing how feature data is created from the example sensor signal data of FIG. 4; [Figure 7] A diagram showing how feature data from each sensor is organized into a high-dimensional vector. [Figure 8] FIG. 10 is a chart showing the transformer state estimation process performed by the maintenance management standard identification unit. [Figure 9] (a) is a schematic diagram showing the creation of a learning mixed oil based on abnormality diagnosis diagram A, and (b) is a schematic diagram showing the creation of a learning mixed oil based on abnormality diagnosis diagram B. [Figure 10] (a) is an abnormality diagnosis diagram A from Non-Patent Document 1, and (b) is an abnormality diagnosis diagram B from the same document. [Figure 11] A diagram of the learning sample configuration. [Figure 12] FIG. 10 is a chart showing the processing steps of a method for creating learning data. [Figure 13] (a) is a chart showing the diagnostic process for the transformer condition, and (b) is an explanatory diagram showing the reference values ​​for each level. [Figure 14] FIG. 1 is a conceptual diagram showing the preparation of a mixed oil for study in Example 2. [Figure 15] A diagram of the "Duval Triangle" from Non-Patent Document 7. [Figure 16] 1 is a conceptual diagram showing the structure of deep learning implemented in the maintenance management standard identification unit of the first embodiment. [Figure 17] 1 is a conceptual diagram showing the structure of deep learning implemented in the maintenance management standard identification unit of the first embodiment. [Figure 18] FIG. 10 is a configuration diagram of a simple transformer condition inspection system according to a third embodiment. [Figure 19] FIG. 2 is a block diagram of the maintenance management standard estimation unit of the same. [Figure 20] FIG. 2 is a chart showing the transformer status processing and state estimation processing performed by the maintenance management standard identification unit of the same. DETAILED DESCRIPTION OF THE INVENTION

[0016] The following describes a transformer inspection system according to an embodiment of the present invention. This inspection system simply estimates the internal abnormality classification of the maintenance management standard, which is used as a guideline for indicating the state of a transformer, and aims to speed up and simplify the maintenance and inspection work of transformers. [Example]

[0017] A first embodiment of the inspection system will be described with reference to FIGS. 1 to 13. Here, a sensor array type sense of smell The sensor is used to measure the oil and gas in the insulating oil. However, other sensors can be used as long as they can detect and understand the characteristics of the oil and gas. For example, a sensing element made of a piezoelectric thin film of lead zirconate titanate coated with a sensitive film is mounted on a sensor chip. sense of smell A sensor may be used.

[0018] Like this sense of smell Using the sensor measurement results, a pre-trained deep learning classifier is used to directly estimate the classification corresponding to the transformer maintenance inspection standards described in Non-Patent Document 1, i.e., the internal abnormality classification of the transformer (Tables 1 to 3), and the transformer condition is inspected.

[0019] [Table 1]

[0020] [Table 2]

[0021] [Table 3]

[0022] Table 1 shows the "Level 1 Caution Required" internal abnormality classifications, and if any one of them is met, the transformer will be in a "Level 1 Caution Required" state. Table 2 shows the "Level 2 Caution Required" internal abnormality classifications, and if either A or B is met, the transformer will be in a "Level 2 Caution Required" state.

[0023] Table 3 shows the "abnormal level" of the internal abnormality classification, and if any one of A to C is applicable, the transformer state is at the "abnormal level."

[0024] In the inspection system, it is important to create training data for training the classifier. Deep learning (DNN), a type of multi-class classifier, is used for this purpose (details will be described later).

[0025] <Configuration example> An example of the configuration of the inspection system will be described with reference to Figure 1. Here, the inspection system 1 is composed of air purification filters 2 and 9, a sample container 3, solenoid valves A and B, a gas mixing container 4, an olfactory sensor 5, a pump 6, an equipment control unit 7, and a maintenance management standard estimation unit 8.

[0026] The air purification filters 2 and 9 are filters that remove volatile organic compounds from the air or oil gas, and are, for example, activated carbon filters. The sample container 3 is a container that contains insulating oil collected from a transformer. The solenoid valves A and B adjust their opening and closing degrees A and B in response to valve adjustment commands from the device control unit 7.

[0027] The gas mixing vessel 4 is used to mix the gas discharged from the solenoid valve A on the carrier gas side that has passed through the air purification filter 2 with the gas discharged from the solenoid valve B on the sample vessel 3 side.

[0028] The olfactory sensor 5 is a sensor that outputs a sensor signal corresponding to the gas flowing in from the gas mixing container 4, and uses an artificial olfactory system configured with a sensor array equipped with multiple sensors, such as the gas sensor in Patent Document 3.

[0029] The pump 6 sucks in the gas and circulates the gas within the system 1. Here, the pump 6 is equipped with a flow meter, and the flow rate of the gas circulating within the system 1 can be controlled.

[0030] The device control unit 7 and the maintenance management standard estimation unit 8 constitute the main components of the inspection system 1 and function as an inspection device implemented in a computer. These two units 7 and 8 may be configured as the same computer or may be configured as separate computers.

[0031] The device control unit 7 serves as a control unit that controls the entire system, and controls the operation processing of the solenoid valves A and B, the pump 6, and the maintenance management standard estimation unit 8 based on control information that has been preset using control parameters.

[0032] The maintenance management standard estimation unit 8 records the input analysis parameters in the memory unit 23 (see FIG. 5), and records sensor signal data that chronologically summarizes the sensor signals based on the control signal from the device control unit 7 and the sensor signal input from the olfactory sensor 5. It also performs data analysis to estimate which of the internal abnormality classifications of transformers the state of the transformer corresponds to.

[0033] (1) Details of the device control unit 7 The details of the device control unit 7 will be described with reference to Fig. 2. Here, the device control unit 7 includes a parameter input unit 10, a diagnostic schedule management unit 11, and a storage unit 12. This storage unit 12 is constructed in a storage device (RAM, ROM, HDD, SSD, etc.) of a computer.

[0034] Pre-set control parameters are input to the parameter input unit 10. The control parameters input here are output to the diagnostic schedule management unit 11 and stored in the storage unit 12 via the diagnostic schedule management unit 11.

[0035] (A) The diagnosis schedule management unit 11 creates a data measurement operation schedule according to the input control parameters. According to each schedule of the created data measurement operation schedule, it outputs a gas flow rate command to the pump 6, outputs valve adjustment commands for valve opening / closing degrees A and B to the solenoid valves A and B, and outputs a sensor signal recording command to the maintenance management standard estimation unit 8.

[0036] That is, the above-mentioned commands are output with values ​​corresponding to the gas flow rate setting value, sensor cleaning time setting value, sample ON time, sample OFF time, number of sample ON / OFF repetitions, sample ON opening / closing degree setting value, and sample OFF opening / closing degree setting value included in the input control parameters.

[0037] The diagnostic schedule management unit 11 also outputs a sensor signal recording command to the maintenance management standard estimation unit 8. sense of smell It commands the sensor 5 to record the sensor data, and also outputs a data analysis command to command data analysis.

[0038] (B) The data measurement operation schedule is composed of the values ​​of gas flow rate, valve opening / closing degree A, B, and sensor signal recording command against time, as shown in Figure 3. In this case, the time other than the sensor cleaning time and data measurement time in the data measurement operation schedule is standby time.

[0039] The data measurement time indicates the time required to repeat a measurement cycle consisting of a set of a sample ON time indicated by an arrow P in FIG. 3 and a sample OFF time indicated by an arrow Q in FIG.

[0040] The gas flow rate is set to "0" during the standby time, while the gas flow rate setting value is set to "gas flow rate value" during the sensor cleaning time and data measurement time.

[0041] The valve opening / closing degree A is set as follows: during standby time, "valve opening / closing degree A value = 0" and during sensor cleaning time, "sample OFF opening / closing degree set value = valve opening / closing degree A value." Also, during the sample ON time of the data measurement time, the valve opening / closing degree A value is the value obtained by subtracting the "sample ON opening / closing degree set value" from the "sample OFF opening / closing degree set value," while during sample OFF time, "sample OFF opening / closing degree set value = valve opening / closing degree A value."

[0042] The valve opening / closing degree B is set as follows. That is, during standby time and sensor cleaning time, "valve opening / closing degree B value = 0" is used. Also, during the sample ON time of the data measurement time, "sample ON opening / closing degree setting value = valve opening / closing degree B value" is used, while during the sample OFF time, "valve opening / closing degree B value = 0" is used.

[0043] The sensor signal recording command is set to "sensor signal recording command value = 0" during the standby time and the sensor cleaning time, and to "sensor signal recording command value = 1" during the data measurement time. Furthermore, the analysis command is set to "data analysis command value = 1" when the maintenance management standard estimation unit 8 is to perform data analysis, and to "data analysis command value = 0" in other cases.

[0044] In this way, by controlling the data analysis of the pump 6, solenoid valves A and B, and maintenance management standard estimator 8 according to the data measurement operation schedule, sawtooth-shaped sensor signal data is obtained, which records the sensor signals of each sensor mounted on the olfactory sensor 5 in chronological order. During sample ON time, the olfactory sensor 5 generates an output that reacts to the components of the insulating oil, as shown by X in Figure 4. Meanwhile, during sample OFF time, the output decreases as the sensor is cleaned by outside air that has passed through the air purification filter 2, as shown by Y in Figure 4.

[0045] (2) Details of Maintenance Management Standard Estimation Section 8 As shown in Figure 5, the maintenance management standard estimation unit 8 includes a parameter input unit 20, a command setting unit 21, a sensor signal data creation unit 22, a memory unit 23, a feature extraction unit 24, a maintenance management standard identification unit 25, a result output unit 26, and a sensor signal data output unit 27.

[0046] Analytical parameters prepared in advance and control parameters output by the device control unit 7 are input to the parameter input unit 20. The analytical parameters and control parameters input here are output to and stored in the storage unit 23.

[0047] The command setting unit 21 also receives the sensor signal recording command and data analysis command output from the device control unit 7. Here, the command setting unit 21 sets the value of the analysis control state to prioritize the sensor signal recording command, and outputs the analysis control state to the sensor signal data creation unit 22.

[0048] The analysis control status is broadly divided into the following status values: standby status value, sensor signal recording status value, and data analysis status value. In this case, when "sensor signal recording command = 0" and "data analysis command = 0", the analysis control status indicates the standby status value. Furthermore, when "sensor signal recording command = 1", the analysis control status indicates the sensor signal recording status value. Furthermore, when "sensor signal recording command = 0" and "data analysis command = 1", the analysis control status indicates the data analysis status value.

[0049] When the analysis control state is the sensor signal recording state value, the sensor signal data creation unit 22 receives the sensor signal, organizes the sensor signal into sensor signal data recorded in time series, and stores the sensor signal data in the storage unit 23.

[0050] When the analysis control state is the data analysis state value, the feature extraction unit 24 receives the control parameters and sensor signal data from the storage unit 23 and extracts feature amounts from the sensor signal values. The extracted feature amounts are organized as feature amount data and stored in the storage unit 23. Specifically, the feature amount data is created in the following manner.

[0051] First, as shown in Fig. 6, the average value of the sensor signal during the sample OFF time is calculated, and then the difference values ​​of the sensor signal from that average value are calculated discretely and arranged to create feature data for each sensor. Next, as shown in Fig. 7, the feature data calculated from each sensor is connected and organized into high-dimensional vector data. The high-dimensional vector data created in this way is used as feature data.

[0052] When the analysis control state is the data analysis state value, the maintenance management standard identification unit 25 receives the analysis parameters and feature data from the storage unit 23. At this time, deep learning (DNN), a type of multi-class classifier, is used to identify which of the internal abnormality classifications of transformers the transformer state corresponds to. Here, the identification results of "normal," "Caution level 1," "Caution level 2," and "Abnormal level" are saved in the storage unit 23 as the internal abnormality classification of the transformer.

[0053] The analysis parameters include learning parameters learned by deep learning (DNN) through offline processing using previously prepared learning samples. As shown in FIG. 8, the maintenance management standard identification unit 25 reads the input analysis parameters (S01) and prepares a multi-class classifier using deep learning (DNN) (S02). It then reads the feature data created by the feature extraction unit 24 (S03) and identifies the transformer's status (level) (S04). After this identification, the transformer's status, i.e., the internal anomaly classification of the transformer, is output to the memory unit 23 and stored.

[0054] The result output unit 26 reads out the internal abnormality classification of the transformer from the storage unit 23 and outputs it to an external device such as a monitor (not shown). In addition, the sensor signal data output unit 27 reads out the sensor signal data from the storage unit 23 and similarly outputs it to an external device.

[0055] <Offline learning of analysis parameters> As mentioned above, in transformer state estimation by the maintenance management standard identification unit 25, deep learning (DNN), a type of multi-class classifier, is used to identify which category of internal abnormality the transformer state corresponds to, so the learning parameters must be learned by offline processing.

[0056] In offline learning, the preparation of training samples is particularly important. Classifiers that apply deep learning (DNN) have the advantage of being able to automatically adjust learning parameters through learning, but on the other hand, they require the preparation of training samples consisting of a large amount of training data.

[0057] In particular, in the state estimation by the maintenance management standard identification unit 25, sense of smell A large amount of training samples is required, which are training data that are a set of feature amount data created from the sensor signal of the sensor 5 and the internal abnormality classification of the transformer.

[0058] In this case, a sufficient amount of training data must be prepared without bias at each of the levels of "normal," "level 1 requiring caution," "level 2 requiring caution," and "abnormal." In other words, it is necessary to collect a large number of diverse oil samples, but it is difficult to collect a large number of oil samples, including those with defects, from transformers in actual operation.

[0059] Therefore, assuming that an abnormality has occurred inside a transformer, we intentionally created pseudo-abnormal oil that corresponds to "each aspect of the abnormality diagnosis diagram within the aspect diagnosis of a faulty transformer" described in Non-Patent Document 1 (Chapter 5: Diagnosis of Faulty Transformer Features).More than 5,000 patterns of learning mixed oils were created by blending insulating oil with this pseudo-abnormal oil, and learning samples were created based on the data measured for the learning mixed oils.

[0060] (1) How to make mixed oil for learning An example of a method for creating learning oil mixtures is explained based on Figures 9(a) and 9(b). First, insulating oil is subjected to overheating and discharge, etc., to create several types of abnormal crude oil in which flammable gases are generated in the insulating oil. The amount of gas components in each of the abnormal crude oils is then measured using a gas chromatography mass spectrometer.

[0061] In this case, the method for heating the insulating oil may be a local heating device with a heating conductor installed in a container that can heat the oil to a temperature of 300 to 700°C. For example, a local heating device with a heating conductor installed in a container, such as that described in Non-Patent Document 3, can be used to heat the oil to a temperature of 300 to 700°C.

[0062] Furthermore, as a method for applying a discharge to insulating oil, for example, a commercially available electrical insulating oil breakdown voltage test device (see Non-Patent Document 4) is used to generate a discharge pulse in the oil.

[0063] Next, the abnormal crude oils are blended to create the following pseudo-abnormal oils A and B, each with more than 25 patterns. ·Pseudo abnormal oil A The combination of gas ratios of "C2H4 (ethylene) / C2H6 (ethane)" and "C2H2 (acetylene) / C2H4 (ethylene)" was adjusted to correspond evenly to each aspect of the abnormality diagnosis diagram A (see Figure 10(a)) in Non-Patent Document 1. ·Pseudo abnormal oil B The combination of gas ratios of "C2H4 (ethylene) / C2H6 (ethane)" and "C2H2 (acetylene) / C2H6 (ethane)" was adjusted to correspond evenly to each aspect of the abnormality diagnosis diagram B (see Figure 9(b)) in Non-Patent Document 1. Then, the pseudo-abnormal oils A and B are blended with normal oil in over 200 blending ratios to create over 500 patterns of mixed oils for learning.

[0064] (2) Example of the configuration of the learning sample An example of the configuration of a training sample is explained based on Figure 11. As mentioned above, the training sample contains a large amount of training data. The training data consists of the training mixed oil ID, phase classification, amount of gas components in the oil, transformer condition level (internal abnormality classification of the transformer), and feature data.

[0065] The learning oil mixture ID is a six-digit number that identifies each learning oil mixture. The first two digits of this number indicate the number of times the learning oil mixture was created, and the last four digits indicate a serial number assigned to each learning oil mixture in the order in which it was created.

[0066] The aspect classification indicates a numerical value indicating each aspect in the abnormality diagnosis diagram B (see FIG. 10(b)) of Non-Patent Document 1, and is set as follows in this embodiment. 0: "Normal" 1: "Overheating low" 2: "Overheating" 3: "Overheating High" 4: "Partial discharge" 5: "Overheating above 700°C or overheating + discharge" 6: "Discharge" 7: Arc Flash Here, the abnormality diagnosis diagram of Non-Patent Document 1 does not have a "normal" category, but in this embodiment, "normal" is provided as a phase category when the internal abnormality category of the transformer is normal.

[0067] The amount of gas components in oil indicates the amount (ppm) of each gas, "H2", "CH4", "C2H6", "C2H4", "C2H2", and "CO", as well as the total amount of combustible gas TCG (Total Combustible Gas), which is the sum of these gas amounts.

[0068] The internal abnormality classification of the transformer is a numerical value indicating the internal abnormality classification of the transformer. In this embodiment, it is set as follows. 0: "Normal" 1: "Level 1 Caution" 2: "Caution Level 2" 3: "Abnormal Level" The feature data indicates feature data obtained as a result of extracting features from the sensor signal measured by the olfactory sensor for the learning mixed oil using a method similar to that used by the feature extraction unit 24 of the maintenance management standard estimation unit 8.

[0069] (3) How to create individual training data A method for creating learning data will be described with reference to FIG.

[0070] S11, S12: First, when the creation process is started, the learning oil mixture ID and the aspect classification recorded when the learning oil mixture was created are set as learning data (S11, S12).

[0071] S13, S14: The learning oil mixture to be used for creating learning data is subjected to a component analysis using a gas chromatography mass spectrometer (S13). This determines the amount of gas components in the oil, which is an item of the learning data, and sets the amount of gas components in the oil as learning data (S14).

[0072] S15, S16: The transformer condition is diagnosed by gas-in-oil analysis based on the amount of gas components in the learning oil mixture (S15), and the diagnosis result is set as learning data as an internal abnormality classification of the transformer (S16).

[0073] Here, the transformer condition diagnosis in S15 uses the method of Non-Patent Document 1, i.e., the judgment method based on the internal abnormality classifications in Tables 1 to 3, and judges the internal abnormality classification of the transformer based on the gas components in the oil of the learning mixed oil.

[0074] Details will be explained based on Figures 13(a) and 13(b). First, it is determined whether the amount of gas components in the learning oil mixture corresponds to "Level 1 Caution Required" (S21). In this case, the gas components in the learning oil mixture are compared with the criteria for determining "Level 1 Caution Required," namely, the threshold values ​​for "TCG (total variable gases)," "H2 (hydrogen)," "CH4 (methane)," "C2H6 (ethane)," "C2H4 (ethylene)," "C2H2 (acetylene)," and "CO (carbon monoxide)" (each reference value in Figure 13(b)). If any of the comparison results apply, it is determined to be "Level 1 Caution Required," and otherwise it is determined to be "Normal."

[0075] Next, if it is judged to be "Caution Level 1", proceed to S22 to determine whether it corresponds to "Caution Level 2". In S22, the gas components in the learning mixed oil are compared with the threshold values ​​of "C2H2", "C2H4", and "TCG" which are the judgment criteria for "Caution Level 2". If any one of the comparison results applies, it is judged to be "Caution Level 2", otherwise it is judged to be "Caution Level 1".

[0076] Finally, if the result is "Caution Level 2," the process proceeds to S23, where an "Abnormal Level" determination is made. In S23, the gas components in the learning mixed oil are compared with the threshold values ​​of "C2H2," "C2H4," and "TCG," which are the criteria for determining "Abnormal Level." If any one of the comparison results applies, the result is determined to be "Abnormal Level," otherwise it is determined to be "Caution Level 2," and the transformer condition diagnosis is terminated.

[0077] In the method of Non-Patent Document 1 (a determination method based on internal abnormality classifications in Tables 1 to 3), the monthly increase in "TCG" is included in the criteria for determining the abnormality level, but this is not dealt with in this embodiment.

[0078] S17~S19: Mixed oil for learning sense of smell Measurement is performed by the sensor 5 (S17), and feature values ​​are extracted (S18). Here, the same processing as that of the sensor signal data creation unit 22 and feature value extraction unit 24 of the maintenance management standard estimation unit 8 is performed to create feature value data for the learning mixed oil. The created learning data is set (S18), and the creation process ends.

[0079] According to the inspection system 1 of this embodiment, there is no need to use a gas chromatography mass spectrometer when maintaining and inspecting a transformer (S01 to S05), and the condition of the transformer can be easily evaluated in a short period of time, such as a few minutes, using an inexpensive device configuration.

[0080] This can prevent delays in transformer maintenance and inspection work and also contribute to cost reduction. By preparing a large amount of intentionally created learning oil mixtures and using learning parameters pre-trained using learning samples consisting of a large amount of learning data, it is possible to judge the condition of the transformer in accordance with the maintenance management standards. [Example]

[0081] A second embodiment of the inspection system will be described below. The inspection system 1 of this embodiment is basically the same as the first embodiment, but differs in the following respects.

[0082] (1) The method for creating pseudo-abnormal oil in the method for creating learning oil mixtures differs from that in Example 1. That is, as shown in Figure 14, instead of pseudo-abnormal oils corresponding to each aspect of the abnormality diagnosis diagram in Non-Patent Document 1, 25 or more patterns of pseudo-abnormal oils corresponding to each aspect of the "Duval Triangle" are created.

[0083] This pseudo-abnormal oil is blended with normal oil in over 200 different ratios to create over 500 different learning oil mixtures. The pseudo-abnormal oils are created by blending abnormal crude oil so that the gas ratios of "CH4", "C2H4", and "C2H2" correspond evenly to each aspect of the "Duval Triangle".

[0084] The Duval Triangle is used overseas as an index for abnormality diagnosis, and is also being discussed in Japan as an index for abnormality diagnosis. For example, Non-Patent Document 5 discusses it as one of the criteria for using palm oil as insulating oil.

[0085] (2) The configuration of the training sample in this embodiment is the same as in the first embodiment, but the content of the modal classification among the elements constituting the training data is different. As mentioned above, the modal classification in this embodiment uses numerical values ​​indicating each modal classification in the "Duval Triangle" also mentioned in Non-Patent Document 5 (see FIG. 15). In this embodiment, the following settings are made:

[0086] 0: "Normal" 1: “PD (Partial Discharges)” 2: “D1 (Discharges of High Energy)” 3: “D2 (Discharges of Low Energy)” 4: “DT(Discharges with Thermal)” 5: "T1 (Low Temperature t≦300℃)" 6:"T2(Medium Temperature 300℃ <t≦700℃)」 7: "T3 (High Temperature 700℃ <t)」 Here, the "Duval Triangle" does not have a defined "normal" category, but in this embodiment, "normal" is set as a phase category for when the transformer is normal in the internal abnormality category.

[0087] According to the inspection system 1 of this embodiment, in addition to the effect of embodiment 1, it is possible to obtain learning parameters for estimating the state of a transformer using learning samples corresponding to foreign abnormality diagnosis indicators. [Example]

[0088] A third embodiment of the inspection system 1 will be described. In the inspection system 1 of this embodiment, similarly to the first embodiment, the measurement results of oil and gas measured by the olfactory sensor 5 are used to directly estimate the internal abnormality classification of the transformer among the transformer maintenance management standards using a classifier based on deep learning that has been trained in advance, and further, the aspect classification of the abnormality diagnosis diagram is simultaneously estimated.

[0089] (1) The maintenance management standard identification unit 8 of the first embodiment receives analysis parameters and feature data as input, and estimates the internal abnormality classification of the transformer using deep learning (DNN), which is a type of multi-class classifier.

[0090] Deep learning (DNN) is structured as a multi-layer neural network. This structure is roughly composed of a feature extraction layer that extracts features from input data and a recognition layer that compiles the feature extraction results and outputs the recognition results.

[0091] Here, as shown in FIG. 16, the maintenance management standard identification unit of Example 1 is composed of a feature extraction layer 30 that further processes feature data created from the sensor signal data of the olfactory sensor 5, and a recognition layer 31 that estimates the internal abnormality classification of the transformer.

[0092] In contrast to this, the inspection system 1 of this embodiment simultaneously estimates the internal abnormality classification and the aspect classification of the transformer in the maintenance management standard identification unit 8. Here, as shown in Fig. 17 , a recognition layer 32 is added to the deep learning structure of the first embodiment.

[0093] This recognition layer 32 estimates the aspect classification, and shares the recognition layer 31 with the feature extraction layer 30. Here, the aspect classification to be estimated uses the numerical values ​​indicating each aspect of the abnormality diagnosis diagram B in the first embodiment.

[0094] (2) The inspection system 1 of this embodiment is basically configured in the same way as in the first embodiment, as shown in Fig. 18. However, it differs in that the maintenance management standard estimation unit 8 simultaneously estimates the internal abnormality classification and the aspect classification of the transformer and outputs each estimation result.

[0095] The maintenance management estimation unit 8 of this embodiment records the input analysis parameters, and receives the control parameters from the device control unit 7 and the sensor signal from the olfactory sensor 5. At this time, as shown in FIG. 19, the sensor signal data creation unit 22 sense of smell The sensor signals from the sensor 5 are stored in the storage unit 23 as sensor signal data in a time series.

[0096] In addition, the data analysis of the maintenance management standard identification unit 25 simultaneously identifies the transformer state in S04 (determines which category of the internal abnormality category of the transformer it corresponds to) and estimates the aspect category (which category of the abnormality diagnosis diagram B it corresponds to).

[0097] 20, the processing content of the maintenance management standard identification unit 25 will be explained. When the analysis control state is the data analysis state value, the processing starts and analysis parameters and feature data are input. The input analysis parameters are read (S51), and a multi-class classifier using deep learning (DNN), which is a type of multi-class classifier, is prepared (S32).

[0098] Thereafter, the input feature data is read (S33), and the deep learning-based multi-class classifier simultaneously estimates the classification of the internal abnormality category and the appearance category of the transformer (S34). That is, a numerical value indicating the internal abnormality category of the transformer is acquired, and the internal abnormality category corresponding to the acquired numerical value is stored in memory unit 23 (S34a, S34b). At the same time, a numerical value indicating the appearance category of abnormality diagnosis diagram B is acquired, and the appearance category corresponding to the acquired numerical value is stored in memory unit 23 (S34c, S34d), and the process ends. The internal abnormality category and appearance category of the transformer stored in memory unit 23 are read out by result output unit 27 and output to the outside.

[0099] The analysis parameters include learning parameters obtained by deep learning (DNN) through offline processing using previously prepared learning samples. In this example, a set of the mode classification, the transformer internal anomaly classification, and feature data is used. In other words, the feature data of the learning samples is used as input, and learning parameters are obtained offline by training a deep learning network that simultaneously outputs the transformer internal anomaly classification and the mode classification.

[0100] According to this embodiment, in addition to the effects of the first embodiment, the phase classification of the abnormality diagnosis diagram B can be simultaneously estimated, which makes it easier to evaluate the state of the transformer. In this respect, it becomes possible to evaluate the state of the transformer more accurately in a shorter time.

[0101] The present invention is not limited to the above-described embodiment, and can be modified and implemented within the scope of the claims. For example, the "Duval Triangle" may be used for the aspect classification in Example 3, instead of the abnormality diagnosis diagram B. [Explanation of symbols]

[0102] 1...Transformer inspection system 2...Air purification filter 3...Sample container 4...Gas mixer 5… sense of smell Sensor 6...Pump 7...Device control section 8…Maintenance management standards estimation department 9...Air purification filter

Claims

1. 1. A system for inspecting the condition of a transformer, comprising: a sample container for enclosing insulating oil collected from the transformer; a mixing vessel for mixing the gas discharged from the sample vessel side with the gas discharged from the carrier gas side; an olfactory sensor that outputs a sensor signal according to the gas that flows in from the mixing container; a maintenance and inspection standard estimation unit that analyzes the state of the transformer based on the sensor signal of the olfactory sensor and estimates which of the internal abnormality classifications of the transformer the state corresponds to; Equipped with The maintenance and inspection standard estimation unit includes a feature extraction unit that extracts feature values ​​from the signal values ​​of the sensor signals; a maintenance and inspection standard identification unit that identifies the internal abnormality classification based on the feature amount; Equipped with the maintenance and inspection standard identification unit identifies the internal abnormality category using parameters previously learned based on a learning sample; The training samples are A pseudo-abnormal oil is prepared according to each aspect of the classification, and a learning mixed oil of a group of patterns is prepared by blending the pseudo-abnormal oil with insulating oil, based on the results of measurement by the olfactory sensor. The transformer condition is diagnosed by gas-in-oil analysis based on the amount of gas components in the learning mixed oil, and the diagnosis result is set as the internal abnormality classification. A transformer inspection system comprising:

2. The pseudo-abnormal oil is prepared by blending the insulating oil with abnormal crude oil that has been heated and discharged to generate flammable gas.

2. The transformer inspection system according to claim 1.

3. The learning samples were prepared by creating a group of patterns of pseudo-abnormal oils corresponding to the various aspects of "Duval Triangle," and then creating learning mixed oils of a group of patterns by blending the pseudo-abnormal oils with normal oils.

2. The transformer inspection system according to claim 1, wherein the test oil mixture is prepared based on the results of measurement by the olfactory sensor.

4. 4. The transformer inspection system according to claim 1, wherein the classification and the aspect classification are estimated simultaneously.

5. a sample container for enclosing insulating oil collected from the transformer; a mixing vessel for mixing the gas discharged from the sample vessel side with the gas discharged from the carrier gas side; an olfactory sensor that outputs a sensor signal according to the gas that has flowed in from the mixing container side; 1. A method for inspecting the condition of a transformer performed by a system comprising: a maintenance and inspection standard estimation step of analyzing the state of the transformer based on the sensor signal of the olfactory sensor and estimating which of the internal abnormality categories of the transformer the state corresponds to; The maintenance and inspection standard estimation step includes: a feature extraction step of extracting a feature from the signal value of the sensor signal; a maintenance and inspection standard identification step of identifying the internal abnormality classification based on the feature amount; and In the maintenance inspection criteria identification step, identifying the internal anomaly category using pre-trained parameters based on training samples; The training samples are A pseudo-abnormal oil is prepared according to each aspect of the classification, and a learning mixed oil of a group of patterns is prepared by blending the pseudo-abnormal oil with insulating oil, based on the results of measurement by the olfactory sensor. The transformer condition is diagnosed by gas-in-oil analysis based on the amount of gas components in the learning mixed oil, and the diagnosis result is set as the internal abnormality classification. A transformer inspection method comprising:

Images