JP2011185880A - Reliability evaluation device, and program and method of the same - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a reliability evaluation device 10 or the like capable of immediately calculating the reliability of the accuracy of an estimated value based on an output factor of a neural network section 41, for example a residual life estimated value of an oil immersed transformer, even when data out of the range of the data used for modeling the neural network section 41 is input. <P>SOLUTION: When an average degree of polymerization is estimated using the neural network section 41 using the operation history, maintenance history, and design items in addition to the analysis result of insulating oil, a statistics calculation section 25 determines a score matrix t consisting of a principal component score based on a loading matrix P and a measured value recorded on a learning information DB30, calculates T<SP>2</SP>statistics of Hotelling from the determined score matrix t or the like, and determines Q statistics from the measured value. An estimated reliability calculation section 26 calculates an estimated reliability ER for evaluating the reliability of the residual life of the oil immersed transformer calculated by an estimated value calculation section 24 based on a predetermined function using only the T<SP>2</SP>statistics of Hotelling and Q statistics. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a reliability evaluation apparatus that evaluates the reliability of a predetermined estimated value for a predetermined object obtained using a neural network, and in particular, evaluates the reliability of an estimated remaining life for an oil-filled electrical device. The present invention relates to a reliability evaluation device and the like.

There are various types of neural networks. For example, a hierarchical neural network has an input layer, an intermediate layer, and an output layer. In the input layer, the p-dimensional input vector (input factor) x = (x ₁ ,..., X _p ) is converted by the connection weight in the neural network and the sigmoid function in the node constituting the neural network, and the output layer To obtain an r-dimensional output vector (output factor) y = (y ₁ ,..., Y _r ) (see Non-Patent Document 1). For example, in the case of supervised learning, the nodes of the neural network learn and are optimized by changing the coupling strength depending on the input of a correct signal. Neural networks are often applied to problems that have multidimensional data such as statistical processing and cannot be linearly separated.

  Neural networks can accurately model nonlinear relationships between input and output factors, but on the other hand, data outside the range of the data used for modeling is estimated based on the output factors of the neural network. There is a problem that the accuracy of the value is deteriorated. This is generally called an extrapolation problem. Therefore, when prediction is performed using a model based on a neural network, it is necessary to determine whether or not extrapolation is performed and to use it under conditions that are not extrapolated as much as possible. As a method of determining whether or not this extrapolation, generally, whether or not each of the input factors used when obtaining the estimated value is included in the range of data values used for model construction. The upper and lower limit checks are used to check each factor.

  However, in the above upper and lower limit checks, even if each factor falls within the upper and lower limits, it may be out of the data range when viewed as a combination of factors. There was a problem that the accuracy of the estimated value deteriorated. For example, if there are two factors A and B, the data of the combination of both factors is (factor A, factor B) = (0, 0), (1, Assume that there are four types of data 1), (10, 1), and (1, 10). In this case, the upper and lower limit ranges of both factor A and factor B are 0-10. Therefore, in the above-described conventional upper / lower limit check, if factor A and factor B are included in the range of 0 to 10, it is determined to be normal. As a result, for example, the data (factor A, factor B) = (10, 10) is also determined to be normal. However, since there is no combination of (factor A, factor B) = (10, 10) in the data range used for modeling, there is a problem that the accuracy of the estimated value based on the output factor of the neural network cannot be guaranteed at all. there were. If it is about two factors, it can be judged on the spot by humans, but as the number of factors increases, the judgment on the spot by humans becomes impossible.

  Therefore, an object of the present invention is to solve the above problem, and even when data outside the range of the data used for modeling the neural network is input, the output factor of the neural network It is an object to provide a reliability evaluation apparatus and the like that can immediately calculate the reliability of the accuracy of the estimated value based on the above.

The reliability evaluation apparatus according to the present invention is a reliability evaluation apparatus that evaluates the reliability of a predetermined estimated value for a predetermined object obtained using a neural network, and is an input factor used for learning of the neural network. A learning information recording unit that records a predetermined measurement value measured from the predetermined object and a predetermined output factor used for learning of the neural network;
Diagnostic information to be diagnosed by the neural network, a diagnostic information recording unit that records a predetermined measurement value measured from the predetermined object, a predetermined measurement value and a predetermined output factor recorded in the learning information recording unit Learning means for learning by the neural network, and a loading matrix for obtaining a loading matrix composed of coupling coefficients of principal components using principal component analysis based on predetermined measurement values recorded in the learning information recording unit A calculating means; a diagnostic means for inputting a predetermined measurement value recorded in the diagnostic information recording section to a neural network learned by the learning means; and outputting a predetermined output factor as a diagnostic result; and the diagnostic means An estimated value calculating means for calculating a predetermined estimated value based on the predetermined output factor outputted by Calculating a T ² statistics Hotelling from the scoring matrix calculated a score matrix consisting of principal component scores based on the recorded predetermined measurement value to the learning information recording unit and the loading matrix calculated by, the predetermined measurement Statistic calculation means for obtaining a Q statistic that is a square prediction error from a value; and the estimated value calculation means by a predetermined function based on Hotelling's T ² statistic and Q statistic obtained by the statistic calculation means And an estimated reliability calculating means for calculating an estimated reliability for evaluating the reliability of the predetermined estimated value calculated by the above.

  Here, in the reliability evaluation apparatus of the present invention, the predetermined target is an oil-filled electrical device, the predetermined estimated value is a remaining life of the oil-filled electrical device, and the predetermined measurement value is an oil-filled electrical device. The deterioration index contained in the insulating oil, the operating state of the oil-filled electrical device, and the design specifications of the oil-filled electrical device, the predetermined output factor is the average degree of polymerization of the insulating paper of the oil-filled electrical device can do.

Here, in the reliability evaluation apparatus of the present invention, the oil-filled electrical device is an oil-filled transformer, the deterioration index includes a furfural amount and CO ₂ + CO, and the operation state includes an insulating oil replacement history. The design specifications may include the amount of insulating paper and the amount of insulating oil.

A reliability evaluation program according to the present invention is a reliability evaluation program executed by a reliability evaluation apparatus that evaluates the reliability of a predetermined estimated value for a predetermined object, which is obtained using a neural network. A learning information recording unit that records a predetermined measurement value measured from the predetermined object and a predetermined output factor used for learning of the neural network, and a diagnosis for causing the neural network to diagnose And a diagnostic information recording unit that records a predetermined measurement value measured from the predetermined target, and uses the computer of the reliability evaluation device to store the predetermined measurement value recorded in the learning information recording unit. Learning means for inputting a predetermined output factor to cause the neural network to learn, the predetermined information recorded in the learning information recording unit A loading matrix calculation means for obtaining a loading matrix consisting of coupling coefficients of principal components using principal component analysis based on measurement values, and a neural network in which predetermined learning values recorded in the diagnostic information recording unit are learned by the learning means A diagnostic means for inputting and outputting a predetermined output factor as a diagnostic result; an estimated value calculating means for calculating a predetermined estimated value based on the predetermined output factor output by the diagnostic means; and the loading matrix calculating means Based on the calculated loading matrix and a predetermined measurement value recorded in the learning information recording unit, a score matrix composed of principal component scores is obtained, and a Hotelling T ² statistic is calculated from the score matrix, and the predetermined measurement value Statistic calculation means for obtaining a Q statistic that is a square prediction error from the above, Hotelling's T ² statistic obtained by the statistic calculation means Reliability for functioning as an estimated reliability calculating means for calculating an estimated reliability for evaluating the reliability of the predetermined estimated value calculated by the estimated value calculating means by a predetermined function based on the quantity and the Q statistic It is an evaluation program.

  Here, in the reliability evaluation program of the present invention, the predetermined target is an oil-filled electrical device, the predetermined estimated value is a remaining life of the oil-filled electrical device, and the predetermined measurement value is an oil-filled electrical device. The deterioration index contained in the insulating oil, the operating state of the oil-filled electrical device, and the design specifications of the oil-filled electrical device, the predetermined output factor is the average degree of polymerization of the insulating paper of the oil-filled electrical device can do.

The reliability evaluation method according to the present invention is a reliability evaluation method for causing a reliability evaluation apparatus to evaluate the reliability of a predetermined estimated value for a predetermined object obtained using a neural network, and learning the neural network A learning information recording unit for recording a predetermined measurement value measured from the predetermined object and a predetermined output factor used for learning of the neural network, and diagnostic information for diagnosing the neural network And a diagnostic information recording unit that records a predetermined measurement value measured from the predetermined target, and the computer of the reliability evaluation device uses the predetermined measurement value recorded in the learning information recording unit and a predetermined value. A learning step of inputting an output factor to cause the neural network to learn, and a predetermined measurement value recorded in the learning information recording unit. A loading matrix calculation step for obtaining a loading matrix composed of principal component coupling coefficients using principal component analysis, and a predetermined measurement value recorded in the diagnostic information recording unit is input to the neural network learned in the learning step. A diagnostic step for outputting a predetermined output factor as a diagnostic result, an estimated value calculating step for calculating a predetermined estimated value based on the predetermined output factor output in the diagnostic step, and the loading matrix calculating step. Based on the calculated loading matrix and a predetermined measurement value recorded in the learning information recording unit, a score matrix composed of principal component scores is obtained, and a Hotelling T ² statistic is calculated from the score matrix, and the predetermined measurement value Statistic calculating step for obtaining a Q statistic that is a square prediction error from the by a predetermined function based on the ng of T ² statistic and Q statistic, the estimated reliability calculation step of calculating the estimated reliability of evaluating the reliability of a predetermined estimation value calculated by the estimated value calculation step It is characterized by having.

  Here, in the reliability evaluation method of the present invention, the predetermined target is an oil-filled electrical device, the predetermined estimated value is a remaining life of the oil-filled electrical device, and the predetermined measurement value is an oil-filled electrical device. The deterioration index contained in the insulating oil, the operating state of the oil-filled electrical device, and the design specifications of the oil-filled electrical device, the predetermined output factor is the average degree of polymerization of the insulating paper of the oil-filled electrical device can do.

According to the reliability evaluation apparatus and the like of the present invention, in addition to the analysis result of the insulating oil, when the average polymerization degree is accurately estimated by using a neural network that uses an operation history, a maintenance history, and design specifications, Hotelling's T ² statistic, which is an indicator of how far the data calculator compares with the amplitude (variation from the average value) of the data used to create the model, and the data variables used to create the model A Q statistic that is an index representing how much the data of the oil-immersed transformer to be estimated for the remaining life deviates from this correlation. Estimation reliability calculation unit, by a predetermined function based on the T ² statistic and Q statistic of Hotelling obtained by statistic calculation unit, reliability of the remaining life of the oil-filled transformers which are calculated by the estimated value calculation unit The estimated reliability for evaluating the degree is calculated. Predetermined function is given by using the T ² statistic of Hotelling representing the degree of deviation from the data used in the model creation and Q statistic. Since the remaining life of oil-filled transformers is estimated using a number of factors, the reliability of oil-filled transformer remaining life estimation results can be expressed in ^two statistics: Hotel ² 's T ² statistics and Q statistics. Judgment can be easily made based on the estimated reliability using only the target data.

  From the above, even when data outside the range of the data used for modeling the neural network is input, the estimated value based on the output factor of the neural network, that is, the accuracy of the estimated remaining life of the oil-filled transformer There is an effect that the reliability can be calculated immediately. Furthermore, the extrapolation problem of the neural network can be determined in advance by the estimated reliability calculating unit calculating the estimated reliability. For this reason, it becomes possible to improve the reliability of the estimation result of the remaining life, and a sense of security can be given to the user with respect to the estimation result. In addition, there is an effect that it is possible to present what kind of data is necessary for learning in order to increase the accuracy of the above-described model.

It is a figure which shows the chemical structural formula of a cellulose. It is a figure which shows the evaluation reference | standard of the coil insulation paper average polymerization degree of the oil-filled transformer and oil-filled reactor which exceed 1000 kVA which Japan Electrical Manufacturers' Association standard JEM1463-1993 sets. It is a figure which shows the relationship between the amount of furfural and the average degree of polymerization (Source: Electric Joint Research Vol. 54, No. 5 (Part 1)). It is a figure which shows the reliability evaluation apparatus 10 in Example 1 of this invention. It is a figure which shows each data of the inputs 1 and 2 used for the learning of the neural network part 41. FIG. It is a figure which shows each data of the points A and B used for the diagnosis by the neural network part 41. FIG. It is a figure which shows the learning result and diagnostic result of the neural network part 41. FIG. It is a flowchart which shows the flow of a process of the reliability evaluation program and the reliability evaluation method which the computer 11 in Example 1 of this invention performs. It is a block diagram which shows the internal circuit 50 of the computer 11 which performs the computer program of this invention.

  Hereinafter, each embodiment will be described in detail with reference to the drawings.

  In the first embodiment, an example in which the reliability evaluation device of the present invention is applied to the evaluation of the estimated value of the remaining life of the oil-filled electrical device will be described. In the following, first, the outline of the deterioration of the oil-filled electrical equipment will be described, and then the problem of applying the reliability evaluation device etc. of the present invention will be specifically shown, and then the application example of the reliability evaluation device etc. of the present invention Will be described in detail.

In general, the following materials are used in oil-filled electrical equipment.
1) Conductive materials such as copper and aluminum 2) Insulating materials such as insulating oil, insulating paper and press board 3) Iron core material of silicon steel strip 4) Structural materials such as iron and stainless steel Among these materials, oil-filled In electrical equipment, for example, oil-filled transformers, deterioration over time is recognized in insulating materials such as insulating oil and insulating paper.

  For insulating oil, there is also the function of an oil deterioration prevention device (open type, air-sealed type, nitrogen-sealed type, etc.), the deterioration is very slow, and the degree of decrease in dielectric breakdown voltage, which is an important characteristic, is small. On the other hand, with regard to insulating paper, the degree of decrease in dielectric breakdown voltage due to deterioration over time is small, but the degree of decrease in mechanical strength is large. That is, the paper becomes shabby due to aging. As the mechanical strength of the insulating paper decreases (deteriorates), the insulating paper cracks or breaks due to an inrush current or mechanical stress due to electromagnetic force generated at the time of an external short circuit, increasing the risk of dielectric breakdown. Therefore, the life of the oil-filled electrical device is strongly influenced by the mechanical strength of the insulating paper, in particular, the degree of deterioration of the winding conductor insulating paper. That is, it may be considered that the remaining life of the oil-filled electrical device is determined by the dielectric breakdown of the winding conductor insulating paper, that is, the state of deterioration of the insulating paper (decrease in the degree of polymerization).

(1) Degree of polymerization of insulating paper Insulating paper is a polymer made by polymerizing many cellulose molecules. FIG. 1 shows the chemical structural formula of cellulose. The number of basic molecules constituting the cellulose shown in FIG. 1 is called the degree of polymerization. The degree of polymerization in the case of new kraft paper is about 1000. As for the degree of polymerization, when the insulating paper is oxidized and deteriorated, the molecular chain of the cellulose molecule is lowered, that is, the average degree of polymerization is lowered, by breaking the chain of the cellulose molecule. For example, in a transformer used for 30 years, the average degree of polymerization is said to decrease to about 40-60% of the initial value (degree of polymerization 400-600).

(2) Life of Oil-filled Transformer FIG. 2 shows the evaluation criteria for the average degree of polymerization of coil insulation paper of oil-filled transformer and oil-filled reactor exceeding 1000 kVA defined by the Japan Electrical Manufacturers' Association Standard JEM1463-1993. As shown in FIG. 2, the degree of polymerization 450 is a life level at which it is determined that the reliability of the transformer is deteriorated due to deterioration of the insulating paper and needs to be updated, and the degree of polymerization 250 is that of the insulating paper itself. The mechanical strength is lost, and it is a dangerous level where the shape as insulating paper cannot be maintained. In general, according to the Japan Electrical Manufacturers' Association standard JEM1463-1993, the time point at which the degree of polymerization becomes 450 is defined as the life of the oil-filled transformer.

(3) Current oil-filled transformer deterioration diagnosis method In transformer life diagnosis, it is necessary to estimate the degree of deterioration of coil insulation paper. However, it is difficult to measure the coil insulation paper tensile strength and average polymerization degree of the transformer in operation because the coil insulation paper cannot be collected easily. Therefore, the average degree of polymerization of the insulating material (press board, lead insulating paper) that can be collected inside the transformer, furfural which is a kind of aromatic aldehyde which is a product in the decomposition process of insulating paper, and CO ₂ + CO Degradation diagnosis is performed using these results (for example, IEEJ Technical Report No. 922 “Current Status and Trends in Aging Transformer Reliability Maintenance Technology”, Electric Joint Research Vol. 54, No. 5) (See Part 1). Hereinafter, a typical oil-filled transformer deterioration diagnosis method will be described.

1) Degree of polymerization method During inspections such as when the operation is stopped, a method of collecting the press board or lead insulation paper from the transformer that does not affect insulation and diagnosing the degree of deterioration of the insulation paper is called “Polymerization degree method. " The polymerization degree method is a method for estimating the life by estimating the degree of deterioration of the coil insulating paper at the hottest portion (hot spot portion) of the winding coil from the degree of polymerization of the collected insulating paper.

2) CO ₂ + CO method Insulating paper generates various organic components such as water, CO ₂ , and CO due to deterioration. Ingredients that are effective as degradation indicators include CO ₂ + CO, furfural, etc., which are also correlated with the average degree of polymerization. In the CO ₂ + CO method, gas analysis in oil is performed, and the degree of average polymerization is estimated from the amount of CO ₂ + CO, which is the final deterioration product of insulating paper, and deterioration diagnosis is performed.

3) Furfural method Furfural, an aldehyde component, is produced during the decomposition of cellulose. Even when the insulating oil is deaerated, 85% of the furfural remains in the insulating oil and does not diffuse into the gas. Since the adsorption rate to the insulating paper does not depend on the temperature and is constant at about 85%, it is said that the diagnosis can be performed with higher accuracy than the CO ₂ + CO method. FIG. 3 shows the relationship between the amount of furfural and the average degree of polymerization (Source: Electric Joint Research Vol. 54, No. 5 (Part 1)). In FIG. 3, the horizontal axis represents the average degree of polymerization remaining (%), the vertical axis represents the amount of furfural (mg / g), the black circles indicate the actual values, and the hatched portions indicate the experimental values. In the furfural method, the average residual degree of polymerization is determined from the measured amount of furfural according to the relationship shown in FIG. However, as shown in FIG. 3, the average residual degree of polymerization has a considerable range (about 20%) with respect to the measured amount of furfural. It is very difficult to estimate the lifetime.

The problems of the conventional oil-filled transformer deterioration diagnosis method described above can be summarized as follows. 1) Degree of polymerization is a method in which the average degree of polymerization of insulating paper representing the degree of deterioration of oil-filled electrical equipment is directly collected and measured. However, there is a problem that the operation cannot be performed while the device is in operation, and the operation must be stopped. 2) In the CO ₂ + CO method and 3) the furfural method, there is a large range in the estimated degree of polymerization, and accurate deterioration diagnosis is difficult. For example, in the furfural method shown in FIG. 3, there is a range of about 20% in terms of the average degree of polymerization residual ratio with respect to the measured amount of furfural of the oil-filled equipment. This is presumably because the average residual degree of polymerization is also affected by differences in the operating conditions and design specifications of oil-filled electrical equipment. As described above, the conventional oil-filled transformer deterioration diagnosis method basically performs deterioration diagnosis using a single index representing the deterioration of insulating paper, and comprehensively compares the operating conditions and design specifications. There was a problem that it was not a method to judge automatically.

4) Neural network method (refer to JP 2006-308515 A)
An invention for solving the above-described problem is disclosed in Japanese Patent Laid-Open No. 2006-308515. In the invention disclosed in Japanese Patent Application Laid-Open No. 2006-308515, measured values of furfural amount, carbon dioxide and carbon monoxide, moisture content, oxygen content, and hydrogen content in insulating oil, operation history of oil-filled electrical equipment, maintenance history , And a combination of all or some of the design specifications of oil-filled electrical equipment as input factor groups, and the average degree of polymerization of insulating paper as an output factor, by identifying or learning an average degree of polymerization estimation model, a different average Several models for estimating the degree of polymerization have been constructed. Input the input factor group of the oil-filled electrical equipment to be diagnosed into the average polymerization degree estimation model, process the obtained multiple average polymerization degree estimates, and estimate the final average polymerization degree of insulating paper is doing. According to the invention disclosed in Japanese Patent Laid-Open No. 2006-308515, not only the amount of furfural and carbon dioxide and carbon monoxide that have been conventionally used alone, but also other measured values, equipment operation history, maintenance history, and design. Factors related to deterioration of insulating paper such as specifications can be comprehensively considered, and it is actually used in the field for the remaining life diagnosis of oil-filled electrical appliances. In the invention disclosed in Japanese Patent Laid-Open No. 2006-308515, a neural network is used as the average polymerization degree estimation model. However, the extrapolation problem in the neural network described above remains unsolved. The problem of applying the reliability evaluation apparatus or the like in the first embodiment of the present invention is to solve the extrapolation problem when estimating the remaining life of the oil-filled electrical device using a neural network. Below, the application example of the reliability evaluation apparatus etc. in Example 1 of this invention is demonstrated concretely.

  FIG. 4 shows the reliability evaluation apparatus 10 according to the first embodiment of the present invention. The reliability evaluation device 10 is a device that evaluates the reliability of an estimated value (predetermined value) of the remaining life of an oil-filled electrical device (predetermined object) obtained using a neural network. Hereinafter, an oil-filled transformer will be described as an example of an oil-filled electrical device, but the oil-filled electrical device is not limited to the oil-filled transformer. In FIG. 4, reference numeral 11 denotes a computer on which the reliability evaluation program of the present invention is installed, 12 denotes a display unit such as a display of the computer 11 that appropriately displays various output values described later, and 20 denotes the reliability evaluation apparatus 10. Functional blocks 30 indicating functions are input factors used for learning of the neural network, and predetermined measured values measured from an oil-filled transformer (not shown) and predetermined output factors used for learning of the neural network are recorded. A learning information database (DB) (learning information recording unit) 31 is diagnostic information to be diagnosed by the neural network, and a diagnostic information DB (diagnosis information recording unit) in which predetermined measurement values measured from the oil-filled transformer are recorded, Reference numeral 40 denotes a recording device that records the neural network unit 41. The learning information DB 30, the diagnostic information DB 31, and the recording device 40 are connected to the computer 11, and each function shown in the function block 20 is recorded in the recording device 40 as a reliability evaluation program executed by the computer 11. The reliability evaluation apparatus 10 is comprised by the above whole.

(1) Past measurement data input In order to construct the neural network unit 41 for estimating the average degree of polymerization of the insulation paper of the oil-filled transformer, the actual measurement of the oil-filled transformer measured in the past is used as learning information for the neural network unit 41. Input data (input factor, predetermined measured value). The items of data collection are various deterioration indicators obtained by analyzing insulating oil (for example, the amount of furfural, CO ₂ + CO, moisture, oxygen, hydrogen, etc.) and the oil-filled transformer. Operating conditions (for example, average load factor, insulating oil replacement history, insulating oil degassing history, etc.) and oil-filled transformer design specifications (for example, insulating oil deterioration prevention method, cooling System, the amount of insulating paper, the amount of insulating oil, etc.) and the average degree of polymerization of insulating paper that is the answer to the neural network unit 41 for learning the measured data (output of the neural network unit 41). Predetermined output factor).

(2) Construction of Average Degree of Polymerization Estimation Model A learning unit (learning means) 21 shown in FIG. 4 inputs the measured data recorded in the learning information DB 30 and the average degree of polymerization of insulating paper, and a neural network unit. Let 41 learn. That is, an average polymerization degree estimation model is constructed using the input actual measurement data as shown in (1) above. As the average polymerization degree estimation model, a neural network is used as described above. As the neural network, the hierarchical neural network described in the background art is suitable. However, the neural network is not limited to this, and various neural networks shown in Non-Patent Document 1, for example, may be used. The neural network unit 41 is recorded in a recording device 40 such as a hard disk, and parameters such as learned connection weights are recorded. During learning and during later-described diagnosis, the RAM 53 (described later) is loaded and used. In the following description, the neural network unit 41 is assumed to be loaded in the RAM 53 unless otherwise specified.

(3) Data Range Model Construction The loading matrix calculation unit (loading matrix calculation means) 22 shown in FIG. 4 is composed of principal component coupling coefficients using principal component analysis based on the actual measurement data recorded in the learning information DB 30. Find the loading matrix. That is, a model (coefficient matrix P composed of loading vectors) that aggregates a small number of statistical data using the principal component analysis, which is one method of multivariate analysis, for the actual measurement data input in (1) above. Is calculated. Principal component analysis automatically aggregates correlations among variables from a large number of variables. For the principal component analysis and the loading matrix P, see Non-Patent Documents 2 to 4. Hereinafter, description will be made with reference to Non-Patent Documents 2 to 4 as appropriate.

When the linear combination y _{1 of K} variables {x ₁ ,..., X _K } is given by Equation 1,

When the variance of y ₁ is the maximum among the variances of the primary expression of x _K (k = 1,..., K), y ₁ is called the first principal component. Here, ω _K1 (k = 1,..., K) is called a coupling coefficient. Refer to Non-Patent Documents 2 to 4 for how to obtain the maximum dispersion. Now, if there are samples for M oil-filled transformers for K variables (number of input and output factors described above), the measured value {x _mk } (m = 1,..., M; k = 1,..., K) is assumed to be X shown in Equation 2. The coupling coefficient ω _K1 (k = 1,..., K) is expressed by Equation 3.

That is, the data matrix X is data for M samples with the above-described input factor and output factor as a set. Here, the m-th sample X _m shown in Equation 4

A value t _m1 of the first principal component y _{1 represented} by the equation 5 corresponding to

Is called the first principal component score. When the first principal component score corresponding to M samples is set to t ₁ as shown in Equation 6, Equation 7 is obtained.

The second main component or less is also obtained in the same manner. Here, taking into account the R-th principal component, a score matrix t composed of principal component scores is expressed by Equation 8,

And a loading matrix P composed of coupling coefficients as shown in Equation 9,

Then, Equation 10 is obtained.

  For convenience of explanation, Expression 10 has been described, but Expression 10 is used in the latter half. The loading matrix calculation unit 22 calculates a loading matrix P expressed by Equation 9.

(4) Data Input for Diagnosis Target and Average Degree of Polymerization Estimation The diagnosis unit (diagnostic means) 23 shown in FIG. 4 inputs the actual measurement data recorded in the diagnosis information DB 31 to the neural network unit 41 learned by the learning unit 21. Then, the average degree of polymerization of the insulation paper of the oil-filled transformer is output as a diagnosis result. The diagnosis unit 23 inputs a corresponding input factor to the learned neural network unit 41 for the oil-filled transformer to be diagnosed. Specifically, various degradation indexes obtained by analyzing insulating oil (for example, furfural amount, CO ₂ + CO amount, moisture amount, oxygen amount, hydrogen amount, etc. are suitable) and the oil-filled transformer Operating conditions (for example, average load factor, insulating oil replacement history, insulating oil degassing history, etc.) and oil-filled transformer design specifications (for example, insulating oil deterioration prevention method, cooling method) The input factor used in the construction of the above-described (2) average polymerization degree estimation model is input to the neural network unit 41, such as the amount of insulating paper, the amount of insulating oil, etc.). Thereafter, the diagnosis unit 23 calculates and outputs the average degree of polymerization corresponding to the input actual measurement data using the neural network unit 41.

(5) Remaining Life Estimation The estimated value calculation unit (estimated value calculation means) 24 shown in FIG. 4 calculates the remaining life of the oil-filled transformer based on the average degree of polymerization output by the diagnosis unit 23. In calculating the remaining life, the time point when the average polymerization degree 450 defined as the life level of the average polymerization degree shown in FIG. 2 is reached is regarded as the life time point, and the number of years from the current time point to the life time point is defined as the remaining life time. Specifically, a widely used method represented by the following formula 11 was adopted (see “Diagnosis and Update Technology for Factory Electrical Equipment”, IEEJ Technical Report, No. 831 (2001), page 33). ).

Here, n is the number of years of operation (years), DP is the average degree of polymerization after n years of operation, DP ₀ is the initial degree of polymerization, and r is a constant (life reduction rate). First, the lifetime reduction rate r is obtained by substituting the following values in Equation 11 respectively. The average degree of polymerization DP after n years of operation is the average degree of polymerization (estimated value) obtained in (4) above, 1000 for the initial average degree of polymerization DP ₀ , and the introduction of a transformer for the number of years of operation n. The life reduction rate r is obtained by substituting the operation years from the current time to the present time. Next, 450 representing the life level shown in FIG. 2 in the average degree of polymerization DP after n years of operation, and the average degree of polymerization (estimated value) determined in (4) above in the initial average degree of polymerization DP ₀ Then, each of the r calculated as described above is substituted into the life reduction rate r, and the number of operating years n that will be the life is obtained.

(7) Calculation of estimation result reliability The reliability of the estimation result is calculated using the data range model constructed in the above (3). Using the data range model, the input data (data matrix X shown in Expression 2) input in (4) described above is aggregated into a small number of statistical data (score vectors) as shown in Expression 10. From now on, we will summarize the two indicators of Hotelling's T ² statistic and Q statistic.

4 is based on the loading matrix P calculated by the loading matrix calculation unit 22 and the measurement values recorded in the learning information DB 30, as shown in Expression 10. A score matrix t composed of principal component scores is obtained. Next, a standard deviation σ _r of principal component scores from the first principal component to the R-th principal component is obtained from the score matrix t. Further, from the following equation 12, the loading matrix P and the measurement value X _d (1 row × K column) to be estimated recorded in the diagnostic information DB 31 are used to estimate from the first principal component to the R-th principal component. A principal component score t ′ (1 row × R column) is obtained. The Hotelling T ² statistic represented by Equation 13 is calculated from the obtained standard deviation σ _r and the principal component score t ′ to be estimated.

  Subsequently, the statistic calculator 25 obtains a Q statistic that is a square prediction error from the measured value. Equation 14 shows the Q statistic.

Here, x _{k in} the first term of Expression 14 is a measurement value to be diagnosed recorded in the diagnosis information DB 31, and x ^ _{k in} the second term (here, the symbol “^” is an expression for convenience of description) 15 is an adjacent value instead of the upper part of x, which is expressed in the same way in the text below.) Is an estimated value obtained from the loading matrix P and a measurement value to be estimated. Note that a matrix composed of the estimated values x ^ _k from k = 1 to k = K is expressed as X ^ _d and can be obtained by Expression 15. The Hotelling T ² statistic represented by Equation 13 is an index that represents how far the data is compared with the amplitude (variation from the average value) of the data used for model creation. The Q statistic represented by Equation 14 is an index representing how much the data of the oil-immersed transformer to be estimated for remaining life deviates from this correlation with respect to the correlation between the variables of the data used for model creation. Since residual life estimation of the oil-filled transformer in which the remaining service life estimation using multiple factors is performed, the remaining service life estimation result of the oil-filled transformers only two statistical data of T ² statistics and Q statistic of Hotelling Judgment can be made easily using.

The estimated reliability calculation unit (estimated reliability calculation means) 26 shown in FIG. 4 calculates an estimated value by a predetermined function based on the Hotelling T ² statistic and the Q statistic obtained by the statistic calculation unit 25. The estimated reliability ER for evaluating the reliability of the remaining life of the oil-filled transformer calculated by the unit 24 is calculated. Given function by using the T ² statistic of Hotelling representing the degree of deviation from the data used in the model creation and Q statistic is given by equation 16.

As f (Q, T ² ), for example, various calculation formulas as shown in the following formulas 17 to 20 can be considered. The greater the ^{T 2} statistic and Q statistic of Hotelling In Equation 17-20, estimation reliability ER increases. In this case, it is considered that the reliability of the estimation result increases as the estimated reliability ER decreases. On the contrary, the smaller is the T ² statistic and Q statistic of Hotelling, may be used estimation reliability ER increases expression. In this case, it is considered that the reliability of the estimation result increases as the estimated reliability ER increases.

5 to 7 show results of experiments using the reliability evaluation apparatus 10 according to the first embodiment of the present invention. For convenience of explanation, the input measurement values are two (inputs 1 and 2). FIG. 5 shows each data of inputs 1 and 2 used for learning of the neural network unit 41. FIG. 6 shows data of points A and B (described later) used for diagnosis by the neural network unit 41. FIG. 7 shows a learning result and a diagnosis result of the neural network unit 41. In FIG. 7, the horizontal axis is input 1, the vertical axis is input 2, the diamond mark is the learning data of the neural network unit 41, and the triangle mark is the point A (data used for unlearned prediction, which is within the range of the learning data Data) and circles are points B (data used for unlearned prediction and outside the range of the learning data). In the learning results shown in FIG. 7, the maximum value of the Q statistic is 0.634556, and the maximum value of Hotelling's T ² statistic is 4.625113. As shown in FIGS. 6 and 7, the Q statistic for the point A, which is data within the learning data range, was 0.0187, and the Hotelling T ² statistic was 0.6984. On the other hand, for point B, which is data outside the range of the learning data, as shown in FIGS. 6 and 7, the Q statistic is 2.165, and the Hotelling T ² statistic is 2.5955. The smaller the T ² statistic and Q statistic of Hotelling as estimation reliability ER, estimation reliability ER were used increases expression. In this case, it is considered that the reliability of the estimation result increases as the estimated reliability ER decreases. Since the estimated reliability ER for the point A that is the prediction result is smaller than the estimated reliability for the point B that is the prediction result, the reliability for the point A that is the prediction result is high, while the reliability for the point B that is the prediction result is low. . Thus, it is possible to easily determine the two statistical data only estimation reliability ER using the T ² statistic Hotelling confidence for residual life estimation result and Q statistic for an oil-filled transformer . In other words, it is possible to present that data such as point B is necessary for learning in order to increase the accuracy of the above-described model.

FIG. 8 is a flowchart showing the processing flow of the reliability evaluation program and the reliability evaluation method executed by the computer 11 according to the first embodiment of the present invention. As shown in FIG. 8, the measured value and the output factor recorded in the learning information DB 30 are input and the neural network unit 41 learns (learning step, step S10). Based on the measurement value recorded in the learning information DB 30, a loading matrix P composed of the coupling coefficients of the principal components is obtained using principal component analysis (loading matrix calculation step, step S12). The measured value recorded in the diagnosis information DB 31 is input to the neural network unit 41 learned in the learning step (step S10), and an output factor as a diagnosis result is output (diagnosis step, step S14). Based on the output factor output in the diagnosis step (step S14), an estimated value of the remaining life of the oil-filled transformer is calculated (estimated value calculation step, step S16). Based on the loading matrix P calculated in the loading matrix calculation step (step S12) and the measurement values recorded in the learning information DB 30, a score matrix t including the principal component scores is obtained, and recorded in the loading matrix P and the diagnostic information DB 31. A principal component score t ′ from the first principal component to the R-th principal component of the estimation target is obtained from the measurement value _Xd of the estimation target, and Hotelling's T ² statistic is obtained from the score matrix t and the principal component score t ′. And a Q statistic that is a square prediction error is obtained from the estimated value _Xd recorded in the diagnostic information DB 31 and the estimated value X ^ _d converted inside the prediction model (statistic calculation step, step). S18). By a predetermined function based on the statistic calculation step T ² statistics Hotelling obtained in (Step S18) and the Q statistic, the reliability of a predetermined estimation value calculated by the estimated value calculation step (step S16) An estimated reliability ER to be evaluated is calculated (estimated reliability calculation step, step S20).

As described above, according to the first embodiment of the present invention, in addition to the analysis result of the insulating oil, the average degree of polymerization can be accurately obtained by using the neural network unit 41 using the operation history, the maintenance history, and the design specifications. At the time of estimation, the statistic calculator 25 obtains a score matrix t composed of principal component scores based on the loading matrix P calculated by the loading matrix calculator 22 and the measured values recorded in the learning information DB 30. Next, a standard deviation σ _r of principal component scores from the first principal component to the R-th principal component is obtained from the score matrix t. Further, a principal component score t ′ from the first principal component to the R-th principal component to be estimated is obtained from the loading matrix P and the measurement value _Xd to be estimated from Equation 12. The Hotelling T ² statistic represented by Equation 13 is calculated from the obtained standard deviation σ _r and the principal component score t ′ to be estimated. T ² statistics Hotelling is an indicator indicating whether the off much compared to the data used for model creation amplitude (variation from the average value). Subsequently, the statistic calculation unit 25 calculates a Q statistic that is a square prediction error from the measured value _Xd of the estimation target recorded in the diagnosis information DB 31 and the estimated value X ^ _d converted inside the prediction model (Expression 14). Ask for. The Q statistic is an index representing how much the data of the oil-immersed transformer to be estimated for the remaining life deviates from this correlation with respect to the correlation between the variables of the data used for creating the model. The estimated reliability calculation unit 26 calculates the remainder of the oil-filled transformer calculated by the estimated value calculation unit 24 using a predetermined function based on the Hotelling T ² statistics and Q statistics obtained by the statistics calculation unit 25. An estimated reliability ER for evaluating the reliability of the lifetime is calculated. Given function by using the T ² statistic of Hotelling representing the degree of deviation from the data used in the model creation and Q statistic is given by equation 16. Since the remaining life of oil-filled transformers is estimated using a number of factors, the reliability of oil-filled transformer remaining life estimation results can be expressed in ^two statistics: Hotel ² 's T ² statistics and Q statistics. Judgment can be easily made based on the estimated reliability ER using only the target data. From the above, even if data outside the range of data used for modeling of the neural network unit 41 (for example, point B shown in FIG. 7) is input, the estimated value based on the output factor of the neural network unit 41 That is, the reliability of the accuracy of the estimated remaining life of the oil-filled transformer can be calculated immediately.

  The reliability evaluation apparatus 10 according to the first embodiment of the present invention accurately estimates the average degree of polymerization by using the neural network unit 41 that uses the operation history, the maintenance history, and the design specifications in addition to the analysis result of the insulating oil. In this case, the estimated reliability calculation unit 26 calculates the estimated reliability ER, so that the extrapolation problem of the neural network can be determined in advance. For this reason, it becomes possible to improve the reliability of the estimation result of the remaining life, and a sense of security can be given to the user with respect to the estimation result. In addition, in order to increase the accuracy of the above-described model, it is possible to present what data is necessary for learning.

In Example 1 mentioned above, the reliability evaluation apparatus 10 grade | etc., Of this invention was applied with respect to the evaluation of the reliability in the estimation result of the remaining life of an oil-filled transformer. As another application object, for example, the following objects can be considered.
(1) Evaluation of the reliability of the predicted flow rate in a flow prediction device that predicts downstream flow rate using rainfall data at multiple locations (2) Inflow of sewage into the sewage pump station or pump facility at the final treatment plant Evaluation of the reliability of the predicted inflow in the sewage inflow prediction device for predicting the amount (3) Using a prediction model with the air conditioning load performance data, weather data, calendar data and air conditioner operation schedule data as input factors Evaluation of reliability of predicted air conditioning load in air conditioning load prediction method for predicting air conditioning load (4) Evaluation of reliability of predicted energy demand in energy demand prediction method for prediction of energy demand ( 5) Various predicted loads in the plant load prediction method for predicting various loads such as power load, heat load, air load, etc. of the target plant Evaluation of reliability (6) Prediction of hydroelectric power generation prediction method for self-current dam, Evaluation of reliability of predicted hydroelectric power generation (7) Predicted power demand in power demand prediction method of predicting power demand In addition, the reliability evaluation apparatus 10 of the present invention can be applied to the evaluation of reliability for various predicted values.

  FIG. 9 is a block diagram showing an internal circuit 50 of the computer 11 that executes the computer program (reliability evaluation program) of the present invention for realizing the above-described embodiments. As shown in FIG. 9, the CPU 51, ROM 52, RAM 53, image control unit 56, controller 57, input control unit 59, and input / output I / F (interface) unit 61 are connected to a bus 62. In FIG. 9, the above-described computer program of the present invention is recorded on a recording device 58a such as ROM 52, HD, or a recording medium (including a removable recording medium) such as a DVD or CD-ROM 58n. The recording device 58a records the learning information DB 30, the diagnostic information DB 31, the neural network unit 41, and the like described above. The computer program of the present invention is loaded into the RAM 53 from the ROM 52 via the bus 62 or from the recording device 58a or a recording medium such as a DVD or CD-ROM 58n via the controller 57 via the bus 62. The input control unit 59 is connected to an input operation unit 60 such as a mouse or a numeric keypad and performs input control and the like. The VRAM 55 as an image memory has a capacity corresponding to the data capacity of at least one screen of the display unit 12 of the computer 11, and the image control unit 56 converts the data in the VRAM 55 into image data and sends it to the display unit 12. It has a function to do. The input / output I / F unit 61 has an input / output interface function.

  As described above, the CPU 51 executes the computer program of the present invention, whereby the object of the present invention can be achieved. The computer program can be supplied to the computer CPU 51 in the form of a recording medium such as a DVD or a CD-ROM 58n as described above, and a recording medium such as a DVD or a CD-ROM 58n on which the computer program is recorded is also recorded in this manner. It constitutes the invention. As a recording medium on which the computer program is recorded, for example, a memory card, a memory stick, an optical disk, an FD, or the like can be used in addition to the recording medium described above.

  As an application example of the present invention, the present invention can be applied to the reliability evaluation in the estimation result of the remaining life of the oil-filled transformer. In addition, (1) Evaluation of the reliability of the predicted flow rate in the flow rate prediction device that predicts downstream flow rate using rainfall data at multiple locations, (2) Inflow to the pump facility at the sewage pump station or terminal treatment plant Evaluation of reliability of predicted inflow in sewage inflow prediction device for predicting inflow of sewage, (3) Prediction using air load data, weather data, calendar data and air conditioner operation schedule data as input factors Air conditioning load prediction method for predicting air conditioning load using a model, evaluation of reliability of predicted air conditioning load in (4) Energy demand prediction method for predicting energy demand, Prediction in reliability evaluation, (5) Plant load prediction method for predicting various loads such as power load, heat load, air load, etc. of the target plant Evaluation of reliability of various loads, (6) hydropower generation prediction method of self-flow dam, evaluation of reliability of predicted hydropower generation, (7) power demand prediction method of predicting power demand It can be applied to the evaluation of the reliability of various predicted values such as the evaluation of the reliability of the predicted electric power demand amount.

DESCRIPTION OF SYMBOLS 10 reliability evaluation apparatus, 11 computer, 12 display part, 20 function block, 21 learning part, 22 loading matrix calculation part, 23 diagnostic part, 24 estimated value calculation part, 25 statistic calculation part, 26 estimated reliability calculation part, 30 learning information DB, 31 diagnostic information DB, 40 recording device, 41 neural network unit,
50 internal block, 51 CPU, 52 ROM, 53 RAM, 54 display unit, 55 VRAM, 56 image control unit, 57 controller, 58a recording device, 58n recording medium, 59 input control unit, 60 input operation unit, 61 input / output I / F section, 62 buses.

“Neural Network Computational Intelligence, written by Katsura Watanabe, published by Morikita Publishing Co., Ltd., September 5, 2006. “New world of independent component analysis and signal analysis”, written by Aapo Hyvarinen, Juha Karhunnenn, Erkki Oja, translated by Iku Nemoto, Maki Kawakatsu, published by Tokyo Denki University Press, February 10, 2005. "Multivariate analysis for chemists-Introduction to chemometrics-", Yukihiro Ozaki, Akifumi Uda, Toshio Akai, published by Kodansha Scientific Co., Ltd., July 10, 2007. “Principal component analysis”, [online], Manabu Kano, Department of Chemical Engineering, Kyoto University Graduate School of Engineering, Process System Engineering Laboratory (searched on February 15, 2010), Internet, <URL: http: //www.pse .cheme.kyoto-u.ac.jp / ~ kano / document / text-PCA.pdf>.

Claims (7)

A reliability evaluation apparatus that evaluates the reliability of a predetermined estimated value for a predetermined object obtained using a neural network,
A learning information recording unit that records input factors used for learning of the neural network and predetermined measurement values measured from the predetermined object and predetermined output factors used for learning of the neural network;
Diagnostic information to be diagnosed by the neural network, and a diagnostic information recording unit that records a predetermined measurement value measured from the predetermined object;
Learning means for inputting a predetermined measurement value and a predetermined output factor recorded in the learning information recording unit, and learning the neural network;
Loading matrix calculation means for obtaining a loading matrix composed of coupling coefficients of principal components using principal component analysis based on predetermined measurement values recorded in the learning information recording unit;
Diagnostic means for inputting a predetermined measurement value recorded in the diagnostic information recording unit to a neural network learned by the learning means and outputting a predetermined output factor as a diagnostic result;
An estimated value calculating means for calculating a predetermined estimated value based on the predetermined output factor output by the diagnostic means;
Based on the loading matrix calculated by the loading matrix calculation means and a predetermined measurement value recorded in the learning information recording unit, a score matrix composed of principal component scores is obtained, and the predetermined measurement recorded in the diagnostic information recording unit The Hotelling T ² statistic is calculated from the value, the loading matrix, and the score matrix, and a Q statistic that is a square prediction error is obtained from the predetermined measurement value recorded in the diagnostic information recording unit and the loading matrix. A statistic calculation means;
By a predetermined function based on the T ² statistic and Q statistic of Hotelling obtained by the statistical quantity computing means, the estimation reliability evaluating the reliability of a predetermined estimation value calculated by the estimated value calculation means A reliability evaluation apparatus comprising an estimated reliability calculation means for calculating. The reliability evaluation apparatus according to claim 1,
The predetermined object is an oil-filled electrical device, and the predetermined estimated value is a remaining life of the oil-filled electrical device,
The predetermined measurement value includes a deterioration index included in the insulating oil of the oil-filled electrical device, an operating state of the oil-filled electrical device, and design specifications of the oil-filled electrical device,
The reliability evaluation apparatus, wherein the predetermined output factor is an average degree of polymerization of insulating paper of an oil-filled electrical device. The reliability evaluation apparatus according to claim 2, wherein the oil-filled electrical device is an oil-filled transformer, the deterioration index includes a furfural amount and CO ₂ + CO, and the operation state includes a replacement history of insulating oil, The design specification includes an amount of insulating paper and an amount of insulating oil. A reliability evaluation program executed by a reliability evaluation apparatus that evaluates the reliability of a predetermined estimated value for a predetermined object obtained using a neural network, and is an input factor used for learning of the neural network, A learning information recording unit that records a predetermined measurement value measured from the predetermined object and a predetermined output factor used for learning of the neural network, and diagnostic information to be diagnosed by the neural network, from the predetermined object Using the diagnostic information recording unit that records the measured values measured, the computer of the reliability evaluation device,
Learning means for inputting a predetermined measurement value and a predetermined output factor recorded in the learning information recording unit, and causing the neural network to learn,
A loading matrix calculating means for obtaining a loading matrix composed of coupling coefficients of principal components using principal component analysis based on predetermined measurement values recorded in the learning information recording unit;
Diagnostic means for inputting a predetermined measured value recorded in the diagnostic information recording unit to a neural network learned by the learning means and outputting a predetermined output factor as a diagnostic result;
An estimated value calculating means for calculating a predetermined estimated value based on the predetermined output factor output by the diagnostic means;
Based on the loading matrix calculated by the loading matrix calculation means and a predetermined measurement value recorded in the learning information recording unit, a score matrix composed of principal component scores is obtained, and the predetermined measurement recorded in the diagnostic information recording unit The Hotelling T ² statistic is calculated from the value, the loading matrix, and the score matrix, and a Q statistic that is a square prediction error is obtained from the predetermined measurement value recorded in the diagnostic information recording unit and the loading matrix. Statistic calculation means,
By a predetermined function based on the T ² statistic and Q statistic of Hotelling obtained by the statistical quantity computing means, the estimation reliability evaluating the reliability of a predetermined estimation value calculated by the estimated value calculation means A reliability evaluation program for functioning as an estimated reliability calculation means for calculating. In the reliability evaluation program according to claim 4,
The predetermined object is an oil-filled electrical device, and the predetermined estimated value is a remaining life of the oil-filled electrical device,
The predetermined measurement value includes a deterioration index included in the insulating oil of the oil-filled electrical device, an operating state of the oil-filled electrical device, and design specifications of the oil-filled electrical device,
The reliability evaluation program characterized in that the predetermined output factor is an average degree of polymerization of insulating paper of oil-filled electrical equipment. A reliability evaluation method for causing a reliability evaluation apparatus to evaluate reliability of a predetermined estimated value for a predetermined object obtained using a neural network, the input factor used for learning of the neural network being the predetermined factor A learning information recording unit that records a predetermined measurement value measured from a target of the target and a predetermined output factor used for learning of the neural network, and diagnostic information to be diagnosed by the neural network, measured from the predetermined target Using the diagnostic information recording unit that records the predetermined measurement value, the computer of the reliability evaluation device,
A learning step of inputting a predetermined measurement value and a predetermined output factor recorded in the learning information recording unit, and causing the neural network to learn,
A loading matrix calculation step for obtaining a loading matrix composed of coupling coefficients of principal components using principal component analysis based on predetermined measurement values recorded in the learning information recording unit;
A diagnostic step of inputting a predetermined measurement value recorded in the diagnostic information recording unit to the neural network learned in the learning step and outputting a predetermined output factor as a diagnostic result;
An estimated value calculating step for calculating a predetermined estimated value based on the predetermined output factor output in the diagnosis step;
Based on the loading matrix calculated in the loading matrix calculation step and the predetermined measurement value recorded in the learning information recording unit, a score matrix composed of principal component scores is obtained, and the predetermined measurement recorded in the diagnostic information recording unit The Hotelling T ² statistic is calculated from the value, the loading matrix, and the score matrix, and a Q statistic that is a square prediction error is obtained from the predetermined measurement value recorded in the diagnostic information recording unit and the loading matrix. A statistic calculation step;
By a predetermined function based on the T ² statistic and Q statistic of Hotelling obtained by the statistic calculation step, the estimated reliability of evaluating the reliability of a predetermined estimation value calculated by the estimated value calculation step A reliability evaluation method comprising: an estimated reliability calculation step for calculating. In the reliability evaluation method of Claim 6,
The predetermined object is an oil-filled electrical device, and the predetermined estimated value is a remaining life of the oil-filled electrical device,
The predetermined measurement value includes a deterioration index included in the insulating oil of the oil-filled electrical device, an operating state of the oil-filled electrical device, and design specifications of the oil-filled electrical device,
The reliability evaluation method according to claim 1, wherein the predetermined output factor is an average degree of polymerization of insulating paper of an oil-filled electrical device.

Images