Inference apparatus, method for inference, program, and method for creating trained model

Abstract

To provide an inference apparatus capable of inferring future states of a transformer.SOLUTION: In a transformer that contains oil and can take a first state and a second state having a higher probability of an internal abnormality than the first state, an inference apparatus for inferring a state of the transformer includes: an acquisition unit configured to acquire gas-concentration information relating to concentrations of gases contained in the oil; and an inference unit configured to input, into a trained model, the gas-concentration information relating to the transformer in the first state and to infer whether the transformer is in a progression state in which it will progress to the second state before a predetermined elapsed period elapses or in a maintenance state in which it will remain in the first state even after the elapsed period has elapsed. The trained model is trained to classify and output, when the gas-concentration information is input, whether the transformer in the first state is in the progression state or in the maintenance state.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an estimation device, an estimation method, a program, and a method for creating a trained model. [Background technology]

[0002] Abnormalities such as overheating and discharge can occur inside transformers (oil-filled transformers). When such an internal abnormality occurs, a chemical reaction caused by the internal abnormality generates gas in the oil (e.g., insulating oil) contained in the transformer. Therefore, by sampling the oil contained in the transformer and analyzing the sampled oil, it is possible to determine whether or not an internal abnormality has occurred in the transformer.

[0003] Non-Patent Document 1 discloses a technology for determining whether an internal abnormality has occurred in a transformer by measuring the concentration of gas contained in oil (gas-in-oil concentration). This technology uses pairs of the gas-in-oil concentration in the transformer and the results of expert-level condition determination as training data, and prepares a trained model that learns the relationship between the gas-in-oil concentration and the results of the condition determination. Here, the "condition determination result" refers to the result of determining the condition of the transformer. Specifically, it is determined whether the transformer falls into one of five conditions: "normal," "overheating," "discharge," "overheating + micro-discharge," or "insulating oil contamination." Then, by inputting the gas-in-oil concentration of a transformer whose condition is unknown (unknown condition transformer) into the trained model, the result of the condition determination for the unknown condition transformer is estimated. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Central Research Institute of Electric Power Industry Report (Central Research Institute of Electric Power Industry Report), Report No. R10030, "Method for Determining the Condition of Power Transformers Using Gas-in-Oil Analysis Data," 2011 Summary of the Invention [Problem to be solved by the invention]

[0005] The technology of Non-Patent Document 1 is a technology that aims to improve the accuracy of state discrimination by using a trained model. In the technology of Non-Patent Document 1, the state (state) of the transformer at the time when the gas concentration in the oil is measured is estimated, but the future state of the transformer is not estimated.

[0006] The present invention has been made in consideration of the above circumstances, and aims to provide an estimation device, an estimation method, a program, and a method for creating a trained model that are capable of estimating the future state of a transformer. [Means for solving the problem]

[0007] In order to solve the above problem, an estimation device according to a first aspect of the present invention is an estimation device for estimating the state of a transformer that contains oil and can be in a first state and a second state in which the probability of an internal abnormality occurring is higher than in the first state, and includes an acquisition unit that acquires gas concentration information related to the concentration of gas contained in the oil, and an estimation unit that, by inputting the gas concentration information related to the transformer in the first state into a trained model, estimates whether the transformer is in a developed state in which it progresses to the second state before a predetermined elapsed period has elapsed, or a maintained state in which it maintains the first state even after the elapsed period has elapsed, and the trained model is trained to classify the transformer in the first state as either the developed state or the maintained state when the gas concentration information is input, and output the result.

[0008] A second aspect of the present invention relates to the estimation device of the first aspect, wherein the trained models include a first trained model and a second trained model, and the first trained model is trained to classify and output, when the gas concentration information is input, whether the transformer in the first state is in a first progression state in which the transformer progresses to the second state before a first elapsed period has elapsed, or a first maintenance state in which the transformer remains in the first state even after the first elapsed period has elapsed; and the second trained model is trained to classify and output, when the gas concentration information is input, whether the transformer in the first state is in a first progression state in which the transformer progresses to the second state before a second elapsed period has elapsed. The trained model is trained to classify and output whether the transformer is in a second progressing state in which the transformer progresses to the second state, or a second maintaining state in which the first state is maintained even after the second elapsed period has elapsed, and the estimation unit estimates whether the transformer is in the first progressing state or the first maintaining state by inputting the gas concentration information related to the transformer in the first state into the first trained model, and estimates whether the transformer is in the second progressing state or the second maintaining state by inputting the gas concentration information related to the transformer in the first state into the second trained model.

[0009] Furthermore, in a third aspect of the present invention, in the estimation device of the first or second aspect, the trained model is trained to further output a certainty of the classification, and the estimation unit outputs the certainty together with the result of the estimation.

[0010] Furthermore, in a fourth aspect of the present invention, in the estimation device of the third aspect, the trained models include a plurality of trained models having different elapsed periods, and the estimation unit outputs the certainty factor for each elapsed period.

[0011] Furthermore, in a fifth aspect of the present invention, in the estimation device of any one of the first to fourth aspects, the first state and the second state are states that are distinguished from each other based on discrimination information related to the concentration of the gas.

[0012] Furthermore, in a sixth aspect of the present invention, in the estimation device of any one of the first to fifth aspects, the trained model is a model trained by a random forest.

[0013] In order to solve the above problem, an estimation method according to aspect 7 of the present invention is an estimation device for estimating the state of a transformer that contains oil and can be in a first state or a second state in which the probability of an internal abnormality occurring is higher than in the first state, the estimation method acquiring gas concentration information related to the concentration of gas contained in the oil, and inputting the gas concentration information related to the transformer in the first state into a trained model to estimate whether the transformer is in a developed state in which it will progress to the second state within a predetermined elapsed period, or a maintained state in which it will maintain the first state even after the elapsed period, the trained model being trained to classify the transformer in the first state as either the developed state or the maintained state when the gas concentration information is input, and outputting the result.

[0014] In order to solve the above problem, an estimation device according to aspect 8 of the present invention is a program for estimating the state of a transformer that contains oil and can be in a first state or a second state in which the probability of an internal abnormality occurring is higher than in the first state, the program causing a computer to acquire gas concentration information relating to the concentration of gas contained in the oil, and inputting the gas concentration information relating to the transformer in the first state into a trained model to estimate whether the transformer is in a developed state in which it will progress to the second state within a predetermined elapsed period, or a maintained state in which it will maintain the first state even after the elapsed period, and the trained model is trained to classify the transformer in the first state as either the developed state or the maintained state when the gas concentration information is input and output the result.

[0015] In order to solve the above problem, a method for creating a trained model according to aspect 9 of the present invention is a method for creating a trained model for predicting the state of a transformer that contains oil and can be in a first state and a second state in which the probability of an internal abnormality is higher than in the first state, the method comprising: acquiring a plurality of pieces of gas concentration information related to the concentration of gas contained in the oil; acquiring classification information linked to the acquired plurality of pieces of gas concentration information and indicating whether the transformer in the first state related to the linked gas concentration information is in an advanced state in which it progresses to the second state before a predetermined elapsed period has elapsed, or a maintained state in which it maintains the first state even after the elapsed period has elapsed; and using a pair of the gas concentration information and the classification information as training data, creating a trained model that has learned the relationship between the gas concentration information and the classification information. [Effects of the Invention]

[0016] According to the above aspects of the present invention, it is possible to provide an estimation device, an estimation method, a program, and a method for creating a trained model that are capable of estimating the future state of a transformer. [Brief explanation of the drawings]

[0017] [Figure 1] 1 is a block diagram showing a system configuration of an estimation device according to an embodiment of the present invention. [Figure 2] 10 is a table showing the conditions applicable to each abnormality level. [Figure 3] FIG. 4 is a schematic diagram showing an example of a change in gas concentration. [Figure 4] FIG. 2 is a block diagram showing estimation data according to an embodiment of the present invention. [Figure 5] FIG. 10 is a schematic diagram showing an example of a method for creating training data used in learning. [Figure 6] FIG. 1 is a schematic diagram showing a learning process using a random forest. [Figure 7] 10 is a flowchart illustrating an example of processing performed in the estimation device according to the embodiment of the present invention. [Figure 8] 10 is a table showing an example of an estimation result. [Figure 9A] FIG. 8(b) is a diagram showing details of the results of FIG. 8(a). [Figure 9B] FIG. 8(b) is a diagram showing details of the results of FIG. 8(b). [Figure 9C] FIG. 8(c) is a diagram showing details of the results of FIG. [Figure 9D] FIG. 8(d) shows the details of the results of FIG. [Figure 10] FIG. 10 is a block diagram showing estimation data according to a modified example of the present invention. [Figure 11] FIG. 10 is a diagram showing an example of information output by an estimation unit according to a modified example of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0018] Hereinafter, an estimation device, an estimation method, a program, and a method for creating a trained model according to an embodiment of the present invention will be described with reference to the drawings.

[0019] <Guessing device> FIG. 1 is a block diagram showing the system configuration of an estimation device 1 according to this embodiment. The estimation device 1 is a device for estimating the future state of a transformer (oil-filled transformer) 2. The transformer 2 includes a housing 21. The transformer 2 has a housing 21 filled with, for example, oil 24. The housing 21 contains, for example, an iron core 22 around which a coil 23 is wound. The oil 24 may be, for example, insulating oil for cooling the transformer 2. However, the configuration of the transformer 2 shown in FIG. 1 is merely an example. As long as the transformer 2 contains oil 24, the specific configuration of the transformer 2 to which the estimation device 1 is applied can be changed as appropriate.

[0020] In this embodiment, gas concentration information G is used to predict the future state of the transformer 2. The gas concentration information G is information related to the concentration of gas contained in the oil 24. The gas concentration information G is also information used as an explanatory variable in a trained model M described later (details will be described later).

[0021] The gas concentration information G may include, for example, concentration values ​​of predetermined gases (gas concentration values). For example, the gas concentration information G may include the gas concentration values ​​of CO, CO2, H2, CH4, C2H6, C2H4, and C2H2. The gas concentration values ​​can be obtained, for example, by sampling oil 24 from the transformer 2 and analyzing the sampled oil 24. The gas concentration information G may also include information indicating the time the oil 24 was sampled.

[0022] The gas concentration information G may also include values ​​(calculated values) calculated based on these gas concentration values. For example, the gas concentration information G may include the total amount of combustible gases (the sum of the gas concentration values ​​of CO2, H2, CH4, C2H6, C2H4, and C2H2), the total amount of hydrocarbon gases (the sum of the gas concentration values ​​of CH4, C2H6, C2H4, and C2H2), and various gas ratios (C2H2 / C2H4, C2H2 / C2H6, and C2H4 / C2H6). The gas concentration information G may also include values ​​obtained by taking the logarithm of the gas concentration values ​​and calculated values ​​(the total amount of combustible gases, the total amount of hydrocarbon gases, and various gas ratios).

[0023] The value adopted as the gas concentration information G can be changed as appropriate as long as it is a value related to the concentration of the gas contained in the oil 24. The gas concentration information G only needs to include at least one of a gas concentration value and a calculated value. In addition, the type of gas for which the gas concentration value is measured can be changed as appropriate.

[0024] 1, the estimation device 1 according to this embodiment includes an input unit 110, an output unit 120, a storage unit 130, and a processing unit 140. The estimation device 1 is configured using an information device such as a smartphone, a tablet, a personal computer, or a dedicated device.

[0025] The input unit 110 is configured using existing input devices such as a keyboard, a pointing device (mouse, tablet, etc.), buttons, a touch panel, etc. The input unit 110 is operated by a user when inputting the user's instructions to the estimation device 1. The input unit 110 may be an interface for connecting the input device to the estimation device 1. In this case, the input unit 110 inputs an input signal generated in the input device in response to the user's input to the estimation device 1. The input unit 110 may be configured in any way as long as it is capable of inputting the user's instructions to the estimation device 1.

[0026] The input unit 110 may function as an acquisition unit for acquiring gas concentration information G (at least one of a gas concentration value and a calculated value). That is, for example, a user may operate the input unit 110 to input the gas concentration information G to the estimation device 1 as an explanatory variable for the trained model M.

[0027] When the calculated value is used as an explanatory variable of the trained model M (i.e., gas concentration information G), the input unit 110 and the processing unit 140 may cooperate to function as an acquisition unit for acquiring the gas concentration information G. That is, for example, a user may operate the input unit 110 to input a gas concentration value to the estimation device 1, and the processing unit 140 may calculate a calculated value based on the input gas concentration value. That is, the processing unit 140 may function as a calculation unit (not shown) for calculating a calculated value based on the gas concentration value. The input unit 110 and the calculation unit may cooperate to function as an acquisition unit for acquiring the gas concentration information G.

[0028] The above-mentioned acquisition unit (input unit 110, etc.) outputs the acquired gas concentration value and gas concentration information G to processing unit 140 (determination unit 141 and estimation unit 142).

[0029] The output unit 120 outputs information in a form that can be recognized by the user. The output unit 120 may be, for example, an image display device such as a liquid crystal display or an organic EL (Electro Luminescence) display. The output unit 120 may be an interface for connecting an image display device to the estimation device 1. In this case, the output unit 120 generates a video signal for displaying image data and outputs the video signal to the image display device connected to the output unit 120. The output unit 120 may be configured as a touch panel integrated with the input unit 110. The output unit 120 may be configured in any way as long as it is capable of outputting information in a form that can be recognized by the user.

[0030] The storage unit 130 is configured using a storage device such as a magnetic hard disk drive or a semiconductor storage device. The storage unit 130 stores data used by the processing unit 140. The storage unit 130 stores data required when the processing unit 140 performs processing.

[0031] The storage unit 130 according to this embodiment stores, for example, determination data 131 and estimation data 132. The determination data 131 is data used by the determination unit 141 included in the processing unit 140. The estimation data 132 is data used by the estimation unit 142 included in the processing unit 140. The estimation data 132 includes a trained model M (see also FIG. 4). The determination data 131 and the estimation data 132 (trained model M) will be described in detail below.

[0032] The processing unit 140 is configured using a processor such as a CPU (Central Processing Unit) and a memory (main storage device). The processing unit 140 according to this embodiment functions as a determination unit 141 and an estimation unit 142 by the processor executing a program. Note that all or part of the functions of the processing unit 140 may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The program may be recorded on a computer-readable recording medium. Examples of computer-readable recording media include portable media such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, and a semiconductor storage device (e.g., a solid-state drive (SSD)), as well as storage devices such as a hard disk or semiconductor storage device built into a computer system. The program may be transmitted via a telecommunications line.

[0033] The determination unit 141 determines the state of the transformer 2 based on the gas concentration value acquired by the input unit 110. Specifically, the determination unit 141 according to this embodiment determines whether the transformer 2 is in the first state, the second state, the third state, or the fourth state at the time when the oil 24 is sampled for analysis of the gas concentration value.

[0034] Here, each of the first to fourth states is a state that the transformer 2 can be in. The second state is a state in which there is a higher probability that an abnormality has occurred inside the transformer 2 (i.e., that an internal abnormality has occurred in the transformer 2) than in the first state. The third state is a state in which there is a higher probability that an internal abnormality has occurred in the transformer 2 than in the second state. The fourth state is a state in which there is a higher probability that an internal abnormality has occurred in the transformer 2 than in the third state. The fourth state may be a state in which an internal abnormality has occurred in the transformer 2.

[0035] In this embodiment, the first to fourth states correspond to the abnormality levels defined in the "Power Transformer Renovation Guideline Expert Committee: Power Transformer Renovation Guideline, Electrical Cooperative Research, Vol. 65, No. 1 (2009)." Figure 2 is a table showing the conditions for each level defined in the guideline.

[0036] The "level" is an index corresponding to the probability of an internal abnormality occurring. Specifically, the "normal level" refers to a state in which it is believed that no abnormality (internal abnormality) has occurred in the transformer 2. The "caution level 1" refers to a state in which it cannot be determined that an abnormality has occurred, but there is some internal change that deviates from normal. The "caution level 2" refers to a state in which signs of an abnormality are observed. The "abnormal level" refers to a state in which an abnormality has clearly occurred. In this embodiment, the first state corresponds to the "normal level," the second state corresponds to the "caution level 1," the third state corresponds to the "caution level 2," and the fourth state corresponds to the "abnormal level." Note that in the practical inspection of the transformer 2, a level of "caution level 1" or higher is treated as "tracking management," and the inspection interval is often shortened. In the practical inspection of the transformer 2, a level of "caution level 2" or higher is often considered to be an internal abnormality.

[0037] As shown in FIG. 2, the "normal level (first state)," "caution level 1 (second state)," "caution level 2 (third state)," and "abnormal level (fourth state)" are states that are distinguished from one another based on information related to gas concentration (also referred to as "discrimination information"; in the example of FIG. 2, this is the gas concentration value). For example, if any one of conditions C1 to C3 shown in FIG. 2 is met, the transformer 2 corresponds to the "abnormal level (fourth state)." If the transformer 2 does not correspond to the "abnormal level" and any one of conditions B1 to B2 is met, the transformer 2 corresponds to the "caution level 2 (third state)." If the transformer 2 does not correspond to either the "abnormal level" or the "caution level 2" and any one of conditions A1 to A7 is met, the transformer 2 corresponds to the "caution level 1 (second state)." If the transformer 2 does not correspond to any of the "abnormal level," "caution level 2," or "caution level 1," the transformer 2 corresponds to the "normal level (first state)." For example, the inequality in condition A2 means that "the H2 gas concentration value is 400 ppm or more." The same applies to other inequalities. Also, "TCG" refers to the total amount of combustible gases mentioned above (the sum of the gas concentration values ​​of CO2, H2, CH4, C2H6, C2H4, and C2H2). "TCG increase rate" refers to the increase in the total amount of combustible gases per unit time.

[0038] The above-mentioned determination data 131 (see FIG. 1) may include information indicating the applicable conditions in FIG. 2. The determination unit 141 may then compare the gas concentration value acquired by the input unit 110 with the applicable conditions included in the determination data 131 to determine whether the state of the transformer 2 is "normal level (first state)," "caution level 1 (second state)," "caution level 2 (third state)," or "abnormal level (fourth state)."

[0039] FIG. 3 is a schematic diagram showing an example of changes in gas concentration inside the transformer 2 (oil 24). It is generally known that when a change (internal change) toward an abnormality (internal abnormality) such as overheating or discharge occurs inside the transformer 2, the gas concentration in the oil 24 increases due to a chemical reaction. FIG. 3 shows, on a two-dimensional plane, a gradual increase in gas concentration due to internal changes in case C, where the gas concentration was initially low. Note that "Gas A" and "Gas B" in FIG. 3 are examples of gases generated in the oil 24. Line L3 indicates the boundary between the normal level and the caution 1 level.

[0040] Here, case C surrounded by line L1 is far from boundary L3, while case C surrounded by line L2 is close to boundary L1. Therefore, case C surrounded by line L2 is considered to have a higher probability of progressing to level 1 requiring caution within a predetermined period of time than case C surrounded by line L1. The inventors of the present application have considered that there is a correlation between information related to the gas concentration of oil 24 (i.e., gas concentration information G) and "whether the level (state) of transformer 2 will progress to the next level (state) before a predetermined period of time has elapsed." Based on this consideration, the estimation unit 142 (and the trained model M) is configured to predict whether the level (state) of transformer 2 will progress based on the gas concentration information G.

[0041] The estimation unit 142 (see FIG. 1) inputs the gas concentration information G into the trained model M included in the estimation data 132, thereby estimating whether the level (state) of the transformer 2 will progress to the next level (state) before a predetermined elapsed period has elapsed. The trained model M has been trained so that, when the gas concentration information G is input, it classifies and outputs whether the level of the transformer 2 will progress to the next level before a predetermined elapsed period has elapsed.

[0042] For example, the estimation unit 142 estimates whether the state of the transformer 2, which is currently at the normal level (first state), will progress to the caution level 1 (second state) before a predetermined elapsed time has elapsed. In other words, the estimation unit 142 estimates whether the transformer 2, which is currently at the normal level (first state), is in an advanced state or a maintained state. Here, the "advanced state" refers to a state that will progress to the next level (state) before a predetermined elapsed time has elapsed. For example, if the transformer 2 is currently at the normal level (first state), the "advanced state" refers to a state that will progress to the caution level 1 (second state) before a predetermined elapsed time has elapsed. The "maintained state" refers to a state in which the current level (state) is maintained even after a predetermined elapsed time has elapsed. For example, if the transformer 2 is currently at the normal level (first state), the "advanced state" refers to a state that will progress to the normal level (first state) even after a predetermined elapsed time has elapsed.

[0043] In this embodiment, multiple elapsed periods (a first elapsed period and a second elapsed period) are set as the "predetermined elapsed period." Specifically, the first elapsed period is three years, and the second elapsed period is five years. The second elapsed period is longer than the first elapsed period. Whether the level (state) of the transformer 2 will progress to the next level (state) is then predicted for each elapsed period. A state in which the level (state) progresses to the next level (state) before the first elapsed period has elapsed may be referred to as a "first maintenance state." A state in which the current level (state) is maintained even after the first elapsed period has elapsed may be referred to as a "first maintenance state." A state in which the level (state) progresses to the next level (state) before the second elapsed period has elapsed may be referred to as a "second maintenance state." A state in which the current level (state) is maintained even after the second elapsed period has elapsed may be referred to as a "second maintenance state."

[0044] The estimation unit 142 according to this embodiment also estimates whether the transformer 2 in the caution level 1 (second state) will progress to the caution level 2 (third state) before a predetermined period of time has elapsed. That is, the estimation unit 142 according to this embodiment makes the following four types of estimations: <<Conjecture 1-1>> It is predicted whether the state of transformer 2, which is currently at a normal level (first state), will progress to level 1 requiring attention (second state) within three years. That is, it is predicted whether the normal level (first state) will progress to level 1 requiring attention (second state) within three years. In other words, it is predicted whether transformer 2, which is currently at a normal level (first state), will be in a progressed state (first progressed state) in which it will progress to level 1 requiring attention (second state) within three years, or in a maintained state (first maintained state) in which it will remain at the normal level (first state) even after three years have passed. <<Conjecture 1-2>> This program predicts whether the state of transformer 2, which is currently at a normal level (first state), will progress to level 1 requiring attention (second state) within five years. That is, this program predicts whether the normal level (first state) will progress to level 1 requiring attention (second state) within five years. In other words, this program predicts whether transformer 2, which is currently at a normal level (first state), will be in a progressed state (second progressed state) in which it will progress to level 1 requiring attention (second state) within five years, or in a maintained state (second maintained state) in which it will remain at the normal level (first state) even after five years have passed. <Conjecture 2-1> This program predicts whether the state of transformer 2, which is currently at level 1 requiring attention (second state), will progress to level 2 requiring attention (third state) within three years. That is, it predicts whether the state will progress from level 1 requiring attention (second state) to level 2 requiring attention (third state) within five years. In other words, it predicts whether transformer 2, which is currently at level 1 requiring attention (second state), will be in a progressed state (third progressed state) in which it will progress to level 2 requiring attention (third state) within three years, or in a maintained state (third maintained state) in which it will remain at a normal level (first state) even after three years have passed. <Conjecture 2-2> This program predicts whether the state of transformer 2, which is currently at level 1 requiring attention (second state), will progress to level 2 requiring attention (third state) within five years. That is, this program predicts whether the state will progress from level 1 requiring attention (second state) to level 2 requiring attention (third state) within five years. In other words, this program predicts whether transformer 2, which is currently at level 1 requiring attention (second state), will be in a progressed state (fourth progressed state) in which it will progress to level 2 requiring attention (third state) within five years, or in a maintained state (fourth maintained state) in which it will remain at a normal level (first state) even after five years have passed.

[0045] 4 is a block diagram showing the estimation data 132 according to this embodiment. As shown in FIG. 4, the estimation data 132 according to this embodiment includes four trained models M (trained models M1-1, M1-2, M2-1, and M2-2). The number of trained models M included in the estimation data 132 matches the number of types of inferences made by the estimation unit 142, and there is a one-to-one correspondence between the trained models M and the types of inferences. That is, the trained model M1-1 is a model for making the 1-1 inference, the trained model M1-2 is a model for making the 1-2 inference, the trained model M2-1 is a model for making the 2-1 inference, and the trained model M2-2 is a model for making the 2-2 inference.

[0046] The trained model M1-1 is trained to classify and output whether the state of the transformer 2 will progress from the normal level (first state) to the caution level 1 (second state) within three years when gas concentration information G is input. In other words, the trained model M1-1 is trained to classify and output whether the transformer 2, which is at the normal level (first state), is in the first maintenance state or the first progression state when gas concentration information G is input. In other words, the trained model M1-1 is a model trained using the gas concentration information G as an explanatory variable and the first maintenance state and the first progression state as target variables.

[0047] The trained model M1-2 is trained to classify and output whether the state of the transformer 2 will progress from the normal level (first state) to the caution level 1 (second state) within five years when gas concentration information G is input. In other words, the trained model M1-2 is trained to classify and output whether the transformer 2, which is at the normal level (first state), is in the second maintenance state or the second progression state when gas concentration information G is input. In other words, the trained model M1-2 is a model trained using the gas concentration information G as an explanatory variable and the second maintenance state and the second progression state as target variables.

[0048] The trained model M2-1 is trained to classify and output whether the state of the transformer 2 will progress from Caution Level 1 (Second State) to Caution Level 2 (Third State) within three years when gas concentration information G is input. In other words, the trained model M2-1 is trained to classify and output whether the transformer 2, which is in Caution Level 1 (Second State), is in the Third Maintenance State or the Third Progression State when gas concentration information G is input. In other words, the trained model M2-1 is a model trained using the gas concentration information G as an explanatory variable and the Third Maintenance State and the Third Progression State as objective variables.

[0049] The trained model M2-2 is trained to classify and output whether the state of the transformer 2 will progress from Caution Level 1 (Second State) to Caution Level 2 (Third State) within five years when gas concentration information G is input. In other words, the trained model M2-1 is trained to classify and output whether the transformer 2, which is in Caution Level 1 (Second State), is in the Fourth Maintenance State or the Fourth Progression State when gas concentration information G is input. In other words, the trained model M2-2 is a model trained using the gas concentration information G as an explanatory variable and the Fourth Maintenance State and the Fourth Progression State as objective variables.

[0050] The estimation unit 142 (see FIG. 1) may determine the learned model M to which the gas concentration information G is input based on the determination result by the determination unit 141. For example, if the determination unit 141 determines that the state of the transformer 2 is at the normal level (first state), the estimation unit 142 may input the gas concentration information G to the learned model M1-1 and the learned model M1-2. This allows the above-described 1-1 estimation and 1-2 estimation (i.e., estimation of whether or not there is a progression from the normal level (first state) to the caution level 1 (second state)). If the determination unit 141 determines that the state of the transformer 2 is at the caution level 1 (second state), the estimation unit 142 may input the gas concentration information G to the learned model M2-1 and the learned model M2-2. This allows the above-described 2-1 estimation and 2-2 estimation (i.e., estimation of whether or not there is a progression from the caution level 1 (second state) to the caution level 2 (third state)). If the determination unit 141 determines that the state of the transformer 2 is at the caution level 2 (third state) or the abnormal level (fourth state), the estimation unit 142 may not input the gas concentration information G to the learned model M. In this case, the processing unit 140 may output information indicating that the state of the transformer 2 is not subject to estimation.

[0051] The estimation unit 142 outputs the result of estimation (classification) obtained using the determination data 131 (trained model M) to the output unit 120. As a result, the result of estimation (classification) is provided to the user via the output unit 120.

[0052] Each of the trained models M1-1, M1-2, M2-1, and M2-2 may be trained to further output a confidence level of the classification in addition to the above-described classification result. In this case, the estimation unit 142 may be configured to output the confidence level to the output unit 120 together with the estimation (classification) result. As a result, the confidence level is provided to the user together with the estimation (classification) result. The confidence level is a value indicating the reliability of the classification by the trained model M. When the trained models M1-1, M1-2, M2-1, and M2-2 are trained using a random forest, the confidence level may be a classification probability, which will be described later.

[0053] <How to create a trained model> An example of a method for creating the trained model M (trained models M1-1, M1-2, M2-1, and M2-2) will be described below. In this embodiment, learning using random forests is used as an example of a method for creating the trained models M1-1, M1-2, M2-1, and M2-2. However, the method for creating the trained models M1-1, M1-2, M2-1, and M2-2 is not limited to learning using random forests and can be changed as appropriate. Furthermore, the methods for creating the trained models M1-1, M1-2, M2-1, and M2-2 are similar to each other. For this reason, in the following description, only the method for creating the trained model M1-1 will be described, and this will be substituted for the description of the methods for creating the other trained models M1-2, M2-1, and M2-2.

[0054] First, training data to be used for learning is acquired. FIG. 5 is a diagram showing an example of a method for creating training data. In this embodiment, data is first acquired for a plurality of case studies C. Specifically, each case study C is associated with gas concentration information G obtained by sampling the oil 24, the time the oil 24 was sampled, and information indicating the level to which the case study C corresponds. FIG. 5 is a graph plotting case studies C on a two-dimensional plane based on this information, with the horizontal axis representing the time the oil 24 was sampled and the vertical axis representing the level.

[0055] Next, case C, which is the target of trained model M1-1 (1-1 inference), is extracted. That is, from all cases C, only cases C that fall into the normal class are extracted. In FIG. 7, cases C belonging to period P are the cases to be extracted. Then, for each extracted case C, a response variable (i.e., either the first maintenance state or the first progression state) is set. Specifically, the response variable is set based on whether the class progresses to the caution 1 class within three years from each case C. In FIG. 7, "first maintenance state" is set as the response variable for case C belonging to period P1, and "first progression state" is set as the response variable for case C belonging to period P2. As a result, multiple training data (cases C) are prepared, each including a pair of gas concentration information G and classification information (information indicating whether it is the "first maintenance state" or the "first progression state").

[0056] Next, learning is performed using a random forest based on the prepared training data (case C). That is, a trained model M1-1 is created that has learned the relationship between the gas concentration information G and the classification information. Figure 6 is a schematic diagram showing the learning process using a random forest.

[0057] As shown on the left side of Figure 6, multiple decision trees T are used in learning using random forests. Each decision tree T is a flowchart that assigns multiple cases C to multiple leaf nodes LN by repeating assignment based on a decision D multiple times. Here, each decision D is made based on an explanatory variable (i.e., gas concentration information G). For example, the content of decision D may be "whether the concentration value of gas A (an example of an explanatory variable) is less than 180 ppm" or "whether the concentration value of gas B (an example of an explanatory variable) is less than 25 ppm," etc.

[0058] When an example C is input to such a decision tree T, in each decision tree T, the example C is assigned to a certain leaf node LN. Each leaf node LN may be assigned either an example C labeled as a "first maintenance state" or an example C labeled as a "first progress state." For example, in FIG. 7, there is a leaf node LN' labeled "80% maintenance, 20% progress." This means that 80% of all the examples C assigned to the leaf node LN' fall into the first maintenance state and 20% fall into the first progress state. The "proportion of the examples C assigned to the leaf node LN' that fall into that class (first maintenance state / first progress state)" is referred to as the classification probability (in the decision tree T'). In other words, the classification probability (in a certain decision tree T') is defined by the following equation (1): (Classification probability in a decision tree T') = (number of cases C that belong to the class among cases C assigned to the leaf node LN') / (total number of cases C assigned to the leaf node LN) ... (1)

[0059] Conversely, if case C, whose class is unknown, is assigned to this leaf node LN', it can be interpreted that case C corresponds to the first maintenance state with a probability of 80% and to the first progress state with a probability of 20%. In other words, the above-mentioned classification probability can be interpreted as representing the probability that case C corresponds to that class. Based on this classification probability, in the decision tree T' to which leaf node LN' belongs, case C is interpreted as corresponding to the first maintenance state with a probability of 80% and to the first progress state with a probability of 20%. Here, since the probability of corresponding to the first maintenance state exceeds 50%, case C is determined to correspond to the first maintenance state in this decision tree T', and the classification probability of 80% is output.

[0060] Here, in learning using a random forest, the content of the determination D is different for each decision tree T. Specifically, in each decision tree T, a combination of explanatory variables used for the determination D is randomly selected. That is, when the total number of explanatory variables is p, for each decision tree T, m (m < p) explanatory variables are randomly selected as the explanatory variables used for the determination D. And this random selection is performed so that the combination of explanatory variables is different for each decision tree T. Note that the value of m is preferably on the order of p 1 / 2 or so.

[0061] Since the content of the determination D is different for each decision tree T, even for the same case C, the classification result (which category it belongs to, the first maintenance state or the first progression state) and the classification probability can be different for each decision tree T. By integrating the classification results and classification probabilities for all decision trees T, the final classification result and classification probability can be obtained. For example, by averaging the classification probabilities for all decision trees T, the final classification result and classification probability can be calculated.

[0062] The right side of FIG. 6 is a schematic diagram showing an example of the classification result two-dimensionally. The vertical axis and the horizontal axis respectively correspond to a certain explanatory variable. Region AR1 corresponds to the region where case C is classified into the first maintenance state, and region AR2 corresponds to the region where case C is classified into the first progression state. The classification boundary BD is the boundary between region AR1 and region AR2. In the example shown on the right side of FIG. 6, the classification accuracy is not perfect. That is, there are cases C (hereinafter also referred to as error cases) where the classification result is incorrect. Specifically, a part of case C labeled as the first progression state (i.e., should exist in region AR1) exists in region AR2. Similarly, a part of case C labeled as the first maintenance state (i.e., should exist in region AR2) exists in region AR1.

[0063] Here, the shape of the classification boundary BD depends on the content of the decision D included in the decision tree T. In other words, when the content of the decision D is changed, the shape of the classification boundary BD also changes accordingly. In learning using a random forest, the content of the decision D is optimized to improve classification accuracy. This corresponds to optimizing the shape of the classification boundary BD to reduce the number of error cases included in the areas AR1 and AR2.

[0064] Specifically, for example, the content of decision D may be optimized so as to maximize the precision of each class. Note that "precision" refers to the proportion of cases C classified into that class that are correctly classified. Alternatively, the content of decision D may be optimized so as to maximize the recall of each class. Note that "recall" refers to the proportion of cases C that actually fall into that class that are correctly classified as falling into that class. Note that the value maximized in optimization does not have to be the precision or recall, and can be changed as appropriate. The multiple decision trees T obtained as a result of such optimization are the learning results of the random forest (i.e., the trained model M-1).

[0065] The trained model M described above may be created by the estimation device 1. In this case, the processing unit 140 may function as a learning processing unit for creating (learning) the trained model M. Alternatively, the trained model M may be created (learned) by a learning device different from the estimation device 1, and then the created trained model M may be stored in the memory unit 130 of the estimation device 1.

[0066] <Processing performed by the estimation device (estimation method)> 7 is a flowchart showing an example of processing (estimation method) performed in the estimation device 1 according to this embodiment. The processing of the flowchart shown in FIG. 7 may be started, for example, when the user operates the input unit 110 to issue a start instruction to the estimation device 1.

[0067] (Step S101) When the processing of the flowchart shown in FIG. 7 is initiated, the processing of step S101 is first performed. In the processing of step S101, the above-mentioned acquisition unit (input unit 110, etc.) acquires gas concentration information G. For example, the user may operate the input unit 110 to input the gas concentration information G to the estimation device 1 (acquisition unit). The acquisition unit outputs the acquired gas concentration information G to the processing unit 140 (determination unit 141 and estimation unit 142). After the processing of step S101 is performed, the processing of step S102 is performed.

[0068] (Step S102) In the process of step S102, the determination unit 141 compares the gas concentration information G (gas concentration value) acquired by the acquisition unit with the corresponding conditions (see also FIG. 2) included in the determination data 131 to determine whether the state of the transformer 2 corresponds to "Caution Level 2 (Third State)" or "Abnormal Level (Fourth State)." If the state of the transformer 2 corresponds to "Caution Level 2 (Third State)" or "Abnormal Level (Fourth State)" (step S102; YES), the process of the flowchart ends. At this time, the processing unit 140 may cause the output unit 120 to output information indicating that the state of the transformer 2 is not subject to estimation. If the state of the transformer 2 does not correspond to either "Caution Level 2 (Third State)" or "Abnormal Level (Fourth State)" (step S102; NO), the process of step S103 is performed.

[0069] (Step S103) In the processing of step S103, the determination unit 141 determines whether the state of the transformer 2 corresponds to the "Caution Level 1 (second state)" by comparing the gas concentration information G (gas concentration value) acquired by the acquisition unit with the corresponding conditions (see also FIG. 2) included in the determination data 131. If the state of the transformer 2 corresponds to the "Caution Level 1 (second state)" (step S103; YES), the processing of step S104 is performed. If the state of the transformer 2 does not correspond to the "Caution Level 1 (second state)" (i.e., corresponds to the "Normal Level (first state)") (step S103; NO), the processing of step S105 is performed.

[0070] If the gas concentration value used in the processes of steps S102 and S103 is not included in the gas concentration information G, the acquisition unit may acquire the gas concentration value in step S101. For example, the user may operate the input unit 110 to input the gas concentration value to the estimation device 1 (acquisition unit). Then, the acquisition unit may output the acquired gas concentration value to the processing unit 140 (determination unit 141).

[0071] (Step S104) In the processing of step S104, the estimation unit 142 inputs the gas concentration information G into the learned model M2-1 and the learned model M2-2 to perform estimation 2-1 and estimation 2-2. That is, it estimates (classifies) whether or not there will be progression to caution level 2 (third state) within three years and within five years, respectively. The estimation unit 142 outputs the result of the estimation (classification) to the output unit 120. As a result, the result of the estimation (classification) is provided to the user via the output unit 120. In addition, the estimation unit 142 performs the estimation (classification) and outputs the above-mentioned classification probability (certainty factor) to the output unit 120. As a result, the classification probability (certainty factor) is provided to the user together with the result of the estimation (classification). After the processing of step S104 is performed, the processing of the flowchart ends.

[0072] (Step S105) In the processing of step S105, the estimation unit 142 inputs the gas concentration information G into the learned model M1-1 and the learned model M1-2 to perform estimation 1-1 and estimation 1-2. That is, it estimates (classifies) whether or not there will be progression to caution level 1 (second state) within three years and within five years, respectively. The estimation unit 142 outputs the result of the estimation (classification) to the output unit 120. As a result, the result of the estimation (classification) is provided to the user via the output unit 120. In addition, the estimation unit 142 performs the estimation (classification) and outputs the above-mentioned classification probability (certainty factor) to the output unit 120. As a result, the classification probability (certainty factor) is provided to the user together with the result of the estimation (classification). After the processing of step S105 is performed, the processing of the flowchart ends.

[0073] <Summary> As described above, the estimation device 1 of this embodiment is an estimation device for estimating the state of a transformer 2 that contains oil 24 and can be in a first state (normal level) or a second state (caution level 1) in which the probability of an internal abnormality is higher than in the first state.The estimation device 1 includes an acquisition unit (input unit 110, etc.) that acquires gas concentration information G related to the concentration of the gas contained in the oil 24, and an estimation unit 142 that inputs gas concentration information G related to the transformer 2 in the first state into a trained model M (trained models M1-1, M1-2) to estimate whether the transformer 2 is in a progressed state in which it progresses to the second state within a predetermined elapsed period (3 years, 5 years), or a maintained state in which it maintains the first state even after the elapsed period has elapsed.The trained model M is trained to classify the transformer 2 in the first state as either a progressed state or a maintained state when gas concentration information G is input, and output the result.

[0074] With this configuration, it is possible to estimate the state of the transformer 2 after a predetermined period of time (for example, three or five years) has passed based on the gas concentration information G. By knowing whether the state of the transformer 2 will progress during the predetermined period of time, the user can adjust the inspection schedule (inspection interval) of the transformer 2.

[0075] The trained model M further includes a trained model M1-1 (first trained model) and a trained model M1-2 (second trained model), and the trained model M1-1 is trained to classify and output, when the gas concentration information G is input, whether the transformer 2 in the first state is in a first progression state in which it progresses to a second state before the first elapsed period (3 years) has elapsed, or a first maintenance state in which it maintains the first state even after the first elapsed period has elapsed. The trained model M1-2 is trained to classify and output, when the gas concentration information G is input, whether the transformer 2 in the first state is in a first progression state in which it progresses to a second state before the first elapsed period (3 years) has elapsed, or a first maintenance state in which it maintains the first state even after the first elapsed period has elapsed. The trained model M1-1 is trained to classify and output a state as either a second progressed state, in which the transformer 2 progresses to the second state within a certain period of time, or a second maintenance state, in which the transformer 2 maintains the first state even after the second elapsed period has elapsed. The estimation unit 142 inputs gas concentration information G related to the transformer 2 in the first state into the trained model M1-1 to estimate whether the transformer 2 is in the first progressed state or the first maintenance state, and inputs gas concentration information G related to the transformer 2 in the first state into the trained model M1-2 to estimate whether the transformer 2 is in the second progressed state or the second maintenance state. This configuration allows for more detailed estimation of the future state of the transformer 2. For example, outputting information such as "the transformer will not progress to the second state within three years, but will progress to the second state within five years" makes it easier for the user to adjust the inspection schedule for the transformer 2.

[0076] Furthermore, the trained model M is trained to further output the classification certainty (classification probability), and the estimation unit 142 outputs the certainty together with the estimation result. This configuration makes it easier for the user to decide whether or not to actually trust the estimation by the estimation unit 142.

[0077] Furthermore, the first state (normal level) and the second state (caution level 1) are states that are distinguished from each other based on discrimination information related to gas concentration. As described above, the trained models M1-1 and M1-2 use the presence or absence of progression from the first state (normal level) to the second state (caution level 1) as the dependent variable. Therefore, when acquiring the dependent variable as training data, it is necessary to distinguish whether the state of the transformer 2 is the first state (normal level) or the second state (caution level 1). On the other hand, in the technology of Non-Patent Document 1, the dependent variable is the state of the transformer 2. Therefore, in order to acquire the dependent variable as training data, it is necessary to perform state discrimination. State discrimination is generally performed by an expert and also involves disassembling the transformer 2, which results in high implementation costs. In contrast, in this embodiment, state discrimination is not necessary to acquire the dependent variable. Furthermore, the first state (normal level) and the second state (caution level 1) can be distinguished based on the gas concentration value that can be measured by sampling oil 24 from the transformer 2. Therefore, the dependent variable can be easily created without disassembling the transformer 2. In this way, by setting the first state and the second state related to the dependent variable to states that can be distinguished from each other based on the discrimination information related to the gas concentration (for example, "normal level" and "caution level 1" based on the power transformer renovation guidelines), it is possible to easily create training data in a learning scenario.

[0078] The trained model M is a model trained by a random forest. This configuration can improve the accuracy of predictions. [Example]

[0079] The above embodiment will be described below using specific examples, but the present invention is not limited to the following examples.

[0080] The trained models M1-1, M1-2, M2-1, and M2-2 described in the above embodiment were created. Then, actual predictions were made using the trained models M1-1, M1-2, M2-1, and M2-2. The parameters listed below were used as explanatory variables (i.e., gas concentration information G). Because gas concentration values ​​often vary by orders of magnitude depending on the type of gas, logarithmic values ​​were used to smooth these out. Furthermore, random forests were used to train the trained models M1-1, M1-2, M2-1, and M2-2, and precision and recall were calculated using cross-validation using the leave-one-out method. Logarithm of CO gas concentration value Logarithm of CO2 gas concentration value Logarithm of H2 gas concentration value Logarithm of CH4 gas concentration value Logarithm of C2H6 gas concentration value Logarithm of C2H4 gas concentration value Logarithm of C2H2 gas concentration value Logarithm of the total amount of flammable gas Logarithm of the total amount of hydrocarbon gas Logarithm of gas ratio (C2H2 / C2H4) Logarithm of gas ratio (C2H2 / C2H6) Logarithm of gas ratio (C2H4 / C2H6)

[0081] FIG. 8 is a table showing the results of the inferences. FIG. 8(a) corresponds to the above-mentioned inference 1-1, FIG. 8(b) corresponds to the above-mentioned inference 1-2, FIG. 8(c) corresponds to the above-mentioned inference 2-1, and FIG. 8(d) corresponds to the above-mentioned inference 2-2. Also, FIG. 9A is a diagram showing details of the results of FIG. 8(a), FIG. 9B is a diagram showing details of the results of FIG. 8(b), FIG. 9C is a diagram showing details of the results of FIG. 8(c), and FIG. 9D is a diagram showing details of the results of FIG. 8(d).

[0082] As shown in Figure 8, the precision and recall rates of 1-1, 1-2, 2-1, and 2-2 were all greater than 0.5. This means that the predictions made in this embodiment can be made with higher accuracy than completely random classification. Specifically, the precision values ​​are generally in the range of 0.7 to 0.8.

[0083] This estimation, with a precision rate of approximately 0.7 to 0.8, is more accurate than completely random classification, but classification errors still exist. As described above, in this embodiment, the precision rate (classification probability) is output along with the classification results. The tables on the right side of each of Figures 9A to 9D show the precision rate when extracting case C, which has a classification probability equal to or greater than a certain value. As can be seen from these tables, extracting cases with high classification probabilities improves the precision rate. In other words, classification results output with high classification probabilities can be said to be highly reliable.

[0084] By checking the classification probability output along with the classification result, the user can decide whether to trust the classification result. Furthermore, the user can refer to the classification result and classification probability to adjust the inspection schedule for transformer 2. For example, if safety is prioritized, it is possible to operate in a way that "even if the output judges that "normality is maintained," the inspection interval is shortened if the classification probability is low." Alternatively, if minimizing inspection costs is prioritized, it is possible to operate in a way that "if the output judges that "normality is maintained," the inspection interval is not shortened even if the classification probability is low." It is expected that more accurate inferences can be made by performing learning using larger amounts of data.

[0085] <Modification> The technical scope of the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present invention.

[0086] For example, in the above embodiment, estimation is performed for two elapsed periods, 3 years and 5 years, but estimation may be performed for three or more elapsed periods. Fig. 10 is a block diagram showing estimation data 132A according to a modified example in which estimation is performed for three or more elapsed periods.

[0087] As shown in FIG. 10, the estimation data 132A according to this modification includes trained models M1-1, M1-2, M1-3, ..., M1-N and trained models M2-1, M2-2, M2-3, ..., M2-N (N≧3). Here, the trained models M1-1, M1-2, M1-3, ..., M1-N are models for estimating whether or not a progression from "normal level (first state)" to "caution level 1 (second state)" occurs for different elapsed periods. The trained models M2-1, M2-2, M2-3, ..., M1-N are models for estimating whether or not a progression from "caution level 1 (second state)" to "caution level 2 (third state)" occurs for different elapsed periods. Furthermore, each trained model M is trained to output a confidence level (classification probability) along with the classification result, as in the above embodiment.

[0088] In this case, the estimation unit 142 (see FIG. 1) may output the classification certainty (classification probability) for each elapsed period. FIG. 11 is a diagram illustrating an example of information output by the estimation unit 142 according to this modification. In the example of FIG. 11, the estimation unit 142 outputs the relationship R between the elapsed period and the abnormality progression probability in the form of a two-dimensional graph. The "abnormal progression probability" is the probability that the state of the transformer 2 progresses to the next level (state) and corresponds to the classification probability of being classified as a progression state. By presenting this relationship R to the user via the output unit 120 or the like, the user can use the relationship R to help determine the inspection schedule for the transformer 2. For example, the user can make a decision such as, "Since the abnormality progression probability is low until five years from now, the next inspection will be performed five years from now," or "Since we want to minimize the risk, the next inspection will be performed three years from now." The relationship R between the elapsed period and the abnormality progression probability may be output in a format other than a graph (e.g., a table, etc.).

[0089] Furthermore, in the above embodiment, estimation is performed for two elapsed periods, but estimation may be performed for only one elapsed period.

[0090] Furthermore, in the above embodiment, the determination unit 141 determines the current state (class) of the transformer 2 based on the gas concentration value, but this is not limited to this. For example, information indicating the current state of the transformer 2 may be linked to the gas concentration information G input to the estimation device 1. Then, the trained model M to be used by the estimation unit 142 may be selected based on this linked information. In this case, the processing unit 140 does not need to have the function of the determination unit 141.

[0091] Furthermore, in the above embodiment, the prediction of whether or not there will be progression is performed for each of "Normal Level (First State) to Caution Level 1 (Second State)" and "Caution Level 1 (Second State) to Caution Level 2 (Third State)." However, the prediction of whether or not there will be progression may also be performed for "Caution Level 2 (Third State) to Abnormal Level (Fourth State)." Alternatively, the prediction of whether or not there will be progression may be performed for only one of "Normal Level to Caution Level 1," "Caution Level 1 to Caution Level 2," and "Caution Level 2 to Abnormal Level." In this case, of the two levels for which the prediction is performed, the level before progression corresponds to the first state, and the level after progression corresponds to the second state. In other words, the classes to which the first and second states (and the third and fourth states) correspond can be changed as appropriate. Furthermore, the first to fourth states do not have to be based on the above-mentioned power transformer renovation guidelines.

[0092] The estimation device 1 may further include a sampling means for sampling the oil 24 from the transformer 2, an analysis means for analyzing the gas concentration value from the sampled oil 24, and the like.

[0093] Furthermore, the estimation device 1 may be configured to output the classification result only when the classification probability exceeds a predetermined threshold.

[0094] Furthermore, the estimation device 1 may be implemented using a plurality of information processing devices. For example, the estimation device 1 may be implemented using a device such as a cloud. For example, in the estimation device 1, the storage unit 130 and the processing unit 140 may be implemented in different information processing devices. For example, the storage unit 130 of the estimation device 1 may be implemented in a distributed manner in a plurality of information processing devices.

[0095] In addition, it is possible to replace the components in the above-described embodiments with well-known components as appropriate, without departing from the spirit of the present invention, and the above-described embodiments and variations may be combined as appropriate. [Explanation of symbols]

[0096] 1... Estimation device 2... Transformer 24... Oil 110... Input unit (acquisition unit) 142... Estimation unit M... Trained model

Claims

1. 1. An estimation device for estimating a state of a transformer that contains oil and can be in a first state and a second state in which there is a higher probability that an abnormality has occurred inside the transformer than in the first state, comprising: an acquisition unit that acquires gas concentration information related to the concentration of gas contained in the oil; an estimation unit that estimates whether the transformer is in a progressed state in which the transformer progresses to the second state before a predetermined elapsed period has elapsed, or a maintained state in which the transformer remains in the first state even after the elapsed period has elapsed, by inputting the gas concentration information related to the transformer in the first state into a trained model; The trained model is trained to classify and output whether the transformer in the first state is in the development state or the maintenance state when the gas concentration information is input. Guessing device.

2. The trained model includes a first trained model and a second trained model, the first trained model is trained to classify and output, when the gas concentration information is input, whether the transformer in the first state is in a first progressing state in which the transformer progresses to the second state before a first elapsed period has elapsed, or a first maintaining state in which the transformer maintains the first state even after the first elapsed period has elapsed; the second trained model is trained to classify, when the gas concentration information is input, whether the transformer in the first state is in a second progression state in which the transformer progresses to the second state before a second elapsed period has elapsed, or a second maintenance state in which the transformer maintains the first state even after the second elapsed period has elapsed, and output the classification result; The estimation unit The gas concentration information related to the transformer in the first state is input to the first trained model to estimate whether the transformer is in the first development state or the first maintenance state; and The gas concentration information related to the transformer in the first state is input to the second trained model to estimate whether the transformer is in the second development state or the second maintenance state. The estimation device according to claim 1 .

3. The trained model is trained to further output a confidence level of the classification, the estimation unit outputs the certainty factor together with the result of the estimation. The estimation device according to claim 1 or 2.

4. The trained model includes a plurality of trained models having different elapsed periods, The estimation unit outputs the certainty factor for each elapsed period. The estimation device according to claim 3.

5. The first state and the second state are states that are distinguished from each other based on discrimination information related to the concentration of the gas. The estimation device according to claim 1 or 2.

6. The trained model is a model trained by a random forest. The estimation device according to claim 1 or 2.

7. 1. An estimation device for estimating a state of a transformer that contains oil and can be in a first state and a second state in which there is a higher probability that an abnormality has occurred inside the transformer than in the first state, comprising: Acquire gas concentration information relating to the concentration of gas contained in the oil; By inputting the gas concentration information related to the transformer in the first state into a trained model, it is possible to estimate whether the transformer is in a progressed state in which the transformer progresses to the second state before a predetermined elapsed period has elapsed, or a maintained state in which the transformer remains in the first state even after the elapsed period has elapsed; The trained model is trained to classify and output whether the transformer in the first state is in the development state or the maintenance state when the gas concentration information is input. Guessing method.

8. A program for estimating a state of a transformer that contains oil and can be in a first state and a second state in which there is a higher probability that an abnormality has occurred inside the transformer than in the first state, the program comprising: On the computer, acquiring gas concentration information relating to the concentration of gas contained in the oil; inputting the gas concentration information related to the transformer in the first state into a trained model, thereby making the trained model infer whether the transformer is in a progressed state in which the transformer progresses to the second state before a predetermined elapsed period has elapsed, or a maintained state in which the transformer remains in the first state even after the elapsed period has elapsed; The trained model is trained to classify and output whether the transformer in the first state is in the development state or the maintenance state when the gas concentration information is input. program.

9. A method for creating a trained model for estimating a state of a transformer that contains oil and can be in a first state and a second state in which there is a higher probability that an abnormality has occurred inside the transformer than in the first state, comprising: acquiring a plurality of pieces of gas concentration information relating to the concentration of gas contained in the oil; acquire classification information associated with the acquired plurality of pieces of gas concentration information, and indicating whether the transformer in the first state associated with the associated gas concentration information is in a progressed state in which the transformer progresses to the second state before a predetermined elapsed period has elapsed, or in a maintained state in which the transformer maintains the first state even after the elapsed period has elapsed; creating a trained model that has learned the relationship between the gas concentration information and the classification information using a set of the gas concentration information and the classification information as training data; How to create a trained model.

Images