JPH11212637A - Method and device for preventive maintenance - Google Patents

Abstract

PROBLEM TO BE SOLVED: To easily specify the cause of an abnormality by providing a normal data estimating part which estimates normal data, comparing the estimated data in a normal state with abnormality data and estimating an abnormality cause by specifying a measurement item that becomes abnormal. SOLUTION: An abnormality detecting part 40 receives measurement data obtained by a normal state leaning part 30 and calculates an outputs value. When the part 40 determines it as abnormal, the measurement data is sent to a normal data estimating part 50 and estimates measurement data (estimated normal data) when a diesel-engine generator 1 for emergency is in a normal state. An abnormal data actually measured and an estimated normal data estimated by the part 50 are sent to an abnormality cause estimating part 60, both of them are compared and a measurement item that becomes abnormal is specified. An abnormality cause is narrowed down from the correspondence relation between the measurement item that becomes abnormal and an abnormality cause. The abnormality cause that is narrowed is sent to an operation terminal 70 and is displayed by it.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a preventive maintenance method and apparatus for measuring or observing the state of equipment such as a plant, a power generation facility, and a communication facility, and estimating a cause of abnormality of a target equipment based on the result.

[0002]

2. Description of the Related Art An anomaly detection method using a neural network includes "anomaly detection based on a functional relationship between process variables by a self-associative neural network" (Chemical Engineering Transactions, Vol. 22, No. 4, p846-p853). ,
1996). The outline of the above-mentioned known example will be described by taking as an example a case where time series data X1 to Xn of n items are obtained as process data. First, the measured value Xi measured when the process is normal is input to the i-th input of the three-layer or five-layer self-associative neural network, and the weight coefficient inside the neural network is adjusted to satisfy the expression (1) (learning). ).

[0003]

(Equation 1)

[0004] By such learning, the measured values X1 to X in the normal state are added to the weighting coefficients of the self-associative neural network.
The correlation between n is stored. When the measurement data in the normal state is input to the self-associative neural network storing the correlation, the expression (1) is satisfied. When the measurement value in the abnormal state where the correlation is lost is input, the expression (1) is not satisfied. Therefore, it is possible to determine whether the state of the process is normal or not using the self-associative neural network in which the correlation between the measurement values when the process is normal is learned in advance.

[0005] Regarding abnormality diagnosis using a neural network, Japanese Patent Laid-Open No. 4-363705 and Japanese Patent Laid-Open No. 6-113459 are disclosed.
No. and so on.

[0006]

As described above, if the conventional technique is used, it is possible to determine whether or not the process state is normal from the value on the left side of the equation (1), and it is possible to detect an abnormality. is there. However, in general, when anomalous data in which some of the measured values X1 to Xn are abnormal are input to the neural network, their effects affect all outputs. It is difficult to identify an abnormal measurement value only by using the above. Therefore, it is also difficult to identify the cause of the abnormality.

[0007]

Means for Solving the Problems To solve the above-mentioned problems, at least a normal data estimating section for estimating normal data when an abnormal state occurs is provided. It is desirable to have an abnormality cause estimating unit that compares and identifies the measurement item that has become abnormal and estimates the cause of the abnormality.

The present invention relates to a preventive maintenance method for detecting an abnormality of equipment from measurement data of equipment and estimating a cause of the abnormality. The preventive maintenance method has a hierarchical neural network model having the same number of input layers and output layers, and A normal state learning unit that learns at least one of the measured value and the feature amount of the measured value as input data of the neural network model, and learns that the input data is equal to the output data; An abnormality detection unit that inputs at least one of the measured value and the feature amount of the measured value measured in the network model, and detects an abnormality from a deviation between the obtained output data of the neural network model and the input data; Normal data estimation to estimate the data when the equipment is in the normal state when the Has a section, compares the abnormality data measured with the normal data estimated in positive normal data estimating unit, in preventive maintenance method characterized by having an abnormal cause estimation part that estimates the cause of the abnormality.

Further, in a preventive maintenance apparatus for detecting an abnormality of equipment from measurement data of equipment and estimating a cause of the abnormality, a preventive maintenance apparatus having a hierarchical neural network model in which the number of input layers and the number of output layers are equal to each other, and a measurement value in a normal state And normal state learning means for learning at least one of the feature values of the measured values as input data of the neural network model and learning the input data to be equal to output data; and a neural network model for learning a normal state by the normal state learning means. Abnormality detection means for inputting at least one of the measured value measured and the characteristic amount of the measurement value, and detecting an abnormality from a deviation between the obtained output data of the neural network model and the input data; and the abnormality detection means. Normal data for estimating data when the device is in a normal state when it is determined to be abnormal And estimating means compares the abnormal data measured with the normal data estimated in positive atmospheric data estimation means, in preventive maintenance apparatus characterized by comprising an abnormality cause estimating means for estimating a cause of the abnormality.

Preferably, the normal data estimating means is for estimating, using the genetic algorithm, data when the device is normal when the abnormality detecting means determines that the device is abnormal.

[0011]

FIG. 1 shows an embodiment of the present invention.

This embodiment is for detecting an abnormality of the emergency diesel engine generator from the measured value of the emergency diesel engine generator, and estimating the cause of the abnormality when the abnormality occurs.

The emergency diesel engine generator starts when a power failure occurs and the supply of commercial power is stopped. That is, since the emergency diesel engine generator is not normally operated, a test operation is periodically performed to perform maintenance and inspection. In this embodiment, a state quantity such as a generator voltage and a field current is measured from a diesel engine generator during a test operation, and abnormality detection and abnormality cause estimation are performed from the measured data.

This embodiment comprises an emergency diesel engine generator 1 and a monitoring device 2, and the emergency diesel engine generator 1 and the monitoring device 2 are connected by a communication line. The monitoring device 2 includes a preventive maintenance device 3 and an operation terminal 70. The preventive maintenance device 3 includes a data preprocessing unit 10, a data storage unit 20, a normal state learning unit 30, an abnormality detection unit 40, a normal data estimation unit 50, An abnormality cause estimating unit 60 is provided.

Next, an outline of the operation of the present embodiment will be described. The operation of this embodiment includes a learning mode and a determination mode.

The learning mode is a mode for preparing for operating the preventive maintenance device, and stores measurement data when the emergency diesel engine generator 1 is operating in a normal state, and learns the correlation therebetween. I do. The determination mode is a mode for detecting an abnormality from actually measured data and estimating the cause of the abnormality.

First, the operation in the learning mode will be described. The measurement data of the emergency diesel engine generator 1 is sent to the data pre-processing unit 10. During the learning mode, basically, the abnormality of the emergency diesel engine generator cannot be detected. However, the data pre-processing unit 10 checks data that has been lost due to a failure of the measurement sensor or a communication error, or data that has an error in the data format (data format error data). Therefore, if there is no abnormality in the data format, the data is stored in the data storage unit 20 as normal data. The normal state learning unit 30 receives the normal data stored in the data storage unit 20, and learns the correlation between the normal data using a neural network. The correlation obtained by the learning is reflected on the weight coefficient inside the neural network, and the weight coefficient is sent to the abnormality detection unit 40.

Next, the determination mode will be described. In the determination mode, similarly to the learning mode, the measurement data is sent to the data pre-processing unit 10 to check whether or not the data format is abnormal. If the data format is abnormal, the data is stored in the data storage unit 20 as data format abnormal data, and is sent to the operation terminal 70 at the same time, and a "data format abnormality" is displayed on the monitor screen of the operation terminal 70. If there is no abnormality in the data format, the measurement data is sent to the abnormality detection unit 40. The abnormality detection unit 40 inputs the measurement data to the neural network having the weight coefficient obtained by the normal state learning unit 30 in the learning mode, and calculates the output value of the neural network. Then, the sum of squares of the error between the input value and the output value is compared with the normal state, and it is determined as normal if the expression (1) is satisfied and abnormal if not. The judgment result is sent to the operation terminal 70, and the judgment result is displayed on the monitor screen of the operation terminal 70. The measurement data is stored in the data storage unit 20 as normal data when determined to be normal, and as abnormal data when determined to be abnormal. When the abnormality is determined to be abnormal by the abnormality detection unit 40, the measurement data is transmitted to the normal data estimation unit 50.
To estimate measurement data (estimated normal data) when the emergency diesel engine generator 1 is in a normal state. The abnormality cause estimating unit 60 is sent the abnormal data actually measured and the estimated normal data estimated by the normal data estimating unit 50, and the two are compared to identify an abnormal measurement item. Then, the cause of the abnormality is narrowed down based on the correspondence between the measurement item having the abnormality and the cause of the abnormality. The narrowed down cause of the abnormality is sent to the operation terminal 70 and displayed on the monitor of the operation terminal 70.

Hereinafter, each block in FIG. 1 will be described in more detail.

FIG. 2 shows an emergency diesel engine generator 1
Is shown. This emergency diesel engine generator 1
Is a three-phase AC generator, and the diesel engine 100,
It comprises a power generation unit 110, a storage battery 120, a rectifier 130, and a starting device 140. Then, when the commercial power is stopped due to a power failure and when a start command is issued from the operation terminal 70 of the monitoring device 2, the diesel engine 100 is started by the start device 140. The power supply process by this emergency diesel engine generator 1 is as follows. When a start command is issued, the switch 141 in the start device 140 is turned on, power is supplied from the start storage battery 143 to the cell motor 142, and the cell motor 142 is started. When the initial rotation is given by the cell motor 142, the diesel engine 100 can ignite the fuel. When the rotation speed reaches 20% of the rated rotation speed, rotation by itself becomes possible, so the starting device 140 is disconnected. Thereafter, the speed increases by itself and is adjusted to the rated speed by the governor. Armature 1
11 rotates at the same rotation speed as the diesel engine 100,
A field current is supplied from the field storage battery 120 to the field winding 112, power generation is started, and the diesel engine 100
The generator voltage rises with an increase in the number of revolutions. There is a system in which the generated power is sent to the load 160 via the switch 150, and a system in which the power is converted into DC by the rectifier 130 and sent to the storage battery 120 and the starting storage battery 143. However,
Until the generated voltage reaches the rating, the switch 150 is off, and no power is supplied to the load 160. This emergency diesel engine generator 1 has a generator speed,
Sensors for measuring a generator voltage, a generator current, a generator field current, a generator field voltage, a starting storage battery total voltage, and a cooling water pressure are provided. These sensors can set the measurement range, and generate a "measurement range over" signal when the measured value does not fall within the measurement range. Next, the data pre-processing unit 10 will be described. The data preprocessing unit 10 checks data in which data is lost or data format is abnormal (data format abnormal data). For the measurement data measured by the emergency diesel engine generator 1, a start time can be applied from the operation terminal 70 of the monitoring device 2 to specify a collection time and a collection interval. Therefore, for example, if 20 seconds are measured at 1 second intervals at the time of starting and stopping, 20 points are taken for each measurement data, and 300 points are taken at 30 second intervals during normal operation, and 10 points are taken for each measurement item.
The number of data to be collected, such as points, is determined. Therefore, by comparing the number of data with the number of data actually sent, it is possible to determine the presence or absence of data loss. Further, if a signal of “measurement range over” is added to the measurement data due to the abnormality of the measurement sensor, it is known that the data format is different from the normal state, and the data format is abnormal. When using a sensor that cannot transmit a signal of “measurement range over”, a range to be used as a measurement value is set in advance by the data preprocessing unit 10, and the measurement value from the sensor exceeds the set range. In this case, it is regarded as a data format error. In this way, it is determined whether or not the data is abnormal data format, and the measurement data having the abnormal data format is stored in the data format abnormal data storage file of the data storage unit 20. Otherwise, it is stored in the normal data storage file of the data storage unit 20.

Next, the normal state learning section 30 will be described. The normal state learning unit 30 operates when the preventive maintenance device is in the learning mode. In the normal data learning unit, the data storage unit 2
The correlation of the normal data stored in 0 is learned by a neural network. FIG. 3 shows the neural network used in this embodiment. The feature of this neural network is that it has a bottleneck structure in which the number n of the input layer and the output layer is left-right symmetric and the number m of the intermediate layers is smaller than the number n of the input layers. In such a neural network, the time series measurement data X1 to
Xn, and outputs the i-th output data Xih (i = 1 to 1).
n) is learned by the back-propagation method so as to be equal to the i-th (i = 1 to n) input Xi, the measured data X1 is added to the weighting coefficient inside the neural network.
It is known that the correlation between Xn and Xn can be learned (see the above-mentioned known example).

The reason why the correlation can be learned will be described. However, for the sake of simplicity, a case where the number of input layers and the number of output layers: 2, the number of intermediate layers: 1 will be described. The two input data X1 and X2 input to the input layer become one in the intermediate layer, and the two data X1h and X1 again in the output layer.
h. If it is assumed that there is no correlation between the input X1 and X2 and that the input X1 and X2 are independent, part of the information held in the input layer when the number of data becomes 1 in the intermediate layer is lost. Data cannot be restored. However, when there is a causal relationship between the input data, the input data can be restored to the output layer from the data that has become one in the intermediate layer and the weight coefficient of the neural network that stores the causal relationship. This is exactly the same principle that the original X1 and X2 can be reproduced from the “value of X1” and “the relational expression: X2 = aX1 + b” for the data sets X1 and X2 in which the relation X2 = aX1 + b exists. .

In the present embodiment, during the test operation of the emergency diesel engine generator, four of the measured values of the seven items described above, namely, the generator rotation speed, the generator field voltage, the generator field current, and the generator voltage The items were designated as X1 to X4 in order, and used as input items. The measurement point in one trial operation was measured at intervals of one second from the start to the time when the generator voltage reached a peak. Normally, the time required to reach the peak is 5 to 6 seconds, and if four sets of measurement data are used as one set, the data measured in one test operation is five or six sets of data. The learning data to be learned in the learning mode was measured values measured in ten test runs measured in the past. Therefore, the total number of learning data is 50 to 60 sets of data. Further, since the absolute value of the measurement data differs depending on the type, all the measurement data are normalized with a lower limit value of 0.05 and an upper limit value of 0.95.

Next, the abnormality detecting section 40 will be described.
The abnormality detection unit 40 operates when the preventive maintenance device 3 is in the determination mode, like the normal data estimation unit 50 and the abnormality cause estimation unit 60. The abnormality detection unit 40 inputs the measurement data to the neural network that has learned the correlation between the measurement values in the normal state in the learning mode. In the present embodiment, similarly to the learning mode, a period from the start to the time when the generator voltage reaches the peak is measured at intervals of 1 second, and is input to the neural network. The data is normalized on the same basis as the normal state learning unit 30.

FIG. 4 shows the operation when the measurement data is actually input. The case shown in (A) is a case where the data is normal. The deviation between the data input to the neural network and the output data is small, and satisfies the expression (1). The case shown in (B) is a case where the data is abnormal, the deviation between the data input to the neural network and the output data is large, and does not satisfy the expression (1). The abnormality detection unit 40 detects an abnormality based on whether or not Expression (1) is satisfied. The measurement data determined to be normal or abnormal by the abnormality detection unit 40 is stored in the normal data file and the abnormal data file of the data storage unit. Further, the determination result is displayed on the monitor screen of the operation terminal 70 and transmitted to the operator.

Next, the normal data estimating section 50 will be described. The normal data estimation unit 50 estimates data (estimated normal data) when the device is normal when the abnormality detection unit 40 determines that the device is abnormal. Since normal data is stored in the data storage unit, past normal data may be used as estimated normal data. However, the measured values that are normal and are measured at the same time after startup are not always the same. Values may vary depending on operating conditions such as temperature and humidity. However, even if there is variation, the causal relationship between the measured values is considered to be the same.
Therefore, in this embodiment, normal data is estimated according to the following concept. The concept will be described with reference to FIG. Here, the normal data is all points on the straight line that satisfies the correlation, and the abnormal data is a point that deviates from the straight line as indicated by point A, for example. When the abnormal data A is measured and the estimated normal data is estimated, it is difficult to specify which point on the straight line is the estimated normal data because all points on the straight line are normal data. However, it is unlikely that all the measured values are completely different between the normal time and the abnormal time, and it is considered that some of the measured values show normal values. Therefore, in the present embodiment, the estimated normal data is obtained on the assumption that some of the measured abnormal data show the same value as in the normal state, such as the point B or the point C. Here, the normal data is
When the signal is input to the neural network, the sum of the squares of the error between the output and the input becomes equal to or smaller than the reference value, and satisfies the expression (1). Therefore, when searching for normal data, the input value may be changed to find a value that makes the sum of the squared error between the output and the input smaller. In this example, estimated normal data was obtained using a genetic algorithm.

The genetic algorithm is an optimization algorithm that simulates the process of evolution of an organism (for details, see "Genetic Algorithm", Industrial Book: Kitano). In the genetic algorithm, an object to be optimized is expressed as one “chromosome”. Those with "chromosomes" are "individuals" (effectively, "individuals" ≒ "chromosomes"),
What determines the evaluation of an “individual” is a “gene” on a “chromosome”. Genetic algorithm is the best by repeating the process of rearranging the "gene" on the "chromosome" to create a new "individual" and the process of duplicating the individual with a higher evaluated trait from the "individual" population This is an algorithm that creates individuals with evaluations. In the present embodiment, in order to obtain a normal value of the four items of measurement data, a “chromosome” having the four items of measurement data as a “gene” may be created. However, in the present example, the chromosome was not the data in which the four items of measurement data were directly set as "genes", but the "chromosomes" were those in which the deviation from the measured abnormal data was set as "genes". An example is shown in FIG. Here, the abnormal data [0.88, 0.7] shown in FIG.
5, 0.70, 0.90] as an example. The gene of a certain "individual" is [-0.06,0,0.10, -0.05]
Then, the estimated normal data assumed at this time is the sum of the two values [0.82, 0.75, 0.80, 0.8.
5]. (However, the values of the abnormal data and the estimated normal data (candidate) shown in FIG. 6 are values normalized by [0.05, 0.95]).

The "chromosome" is defined in this way, and optimization is performed in accordance with each step of FIG. 7 so that the sum of squares of the input / output error of the neural network is reduced. Hereinafter, each step of FIG. 7 will be described in order. In the initial value creation process, the "individual" having the "chromosome" defined above using random numbers
It was created. Here, 30 different “individuals” were created. Next, in the evaluation value calculation step, the evaluation value of each “individual” was calculated. The evaluation value is the sum of squares of the input / output error when the estimated normal data created from each individual is input to the learned neural network, and is the value on the left side of Expression (1). Here, the candidate for the estimated normal data is the “gene” on the “chromosome”.
From the measured abnormal data, a candidate for estimated normal data is created. In the example of FIG. 6, the estimated normal data created is
[0.82, 0.75, 0.80, 0.85]. In the next end determination step, the process ends if the end condition is satisfied, and if not, the process proceeds to the next step. In this embodiment, the termination condition is that three or more “individuals” whose sum of squares of input / output errors obtained in the evaluation value calculation step satisfies the expression (1) are generated.
Therefore, if the number of individuals satisfying the expression (1) is less than two, the process proceeds to the next step, and the process is terminated if three or more individuals satisfy the expression (1) are generated. In the present embodiment, such an end condition is used. However, based on the number of repetitions from the evaluation value calculation process to the reassembly process shown in FIG. 7, the process is terminated when the number of repetitions exceeds the reference value. Conditions may be used. Next, in the selection step, the “individual” having a high evaluation value calculated, that is, the “individual” having a small error is replicated (proliferated) M times, and m chromosomes with a low evaluation value are deleted (selection). = 2). Therefore, in the present embodiment, the number of “individuals” is always 30, but the ratio of “individuals” having a high evaluation value increases by the selection process. Next, in the rearrangement step, the “gene” is rearranged by the operations of “crossover” and “mutation”. Here, “crossover” is an operation of exchanging a part of two “genes”, and “mutation”
Is an operation of changing some values of the “gene” to another value. In "Crossover", select two "Individuals" and select "Chromosome"
The j-th to k-th of the above seven “genes” were exchanged. Here, j and k are 1 ≦ j ≦ 4, 1 ≦ k ≦ 4,
This value satisfies the condition of j <k. FIG. 8 shows an example of this operation. In the example of FIG. 8, the "individual" shown in FIG.
0.06, 0, 0.10, -0.05] and another "individual"
The second to third "genes" of [0.21, -0.05, 0.05, 0.12] were rearranged. In the present embodiment, crossover was performed in this manner. However, instead of exchanging “genes” at the same site, for example, one “individual”
J-th to (j + 1) -th "gene" and another "individual"
An operation of replacing the k-th to (k + 1) -th “gene” may be performed (where k ≠ j). In “mutation”, the value of the m-th “gene” was changed to another value using random numbers (where 1 ≦ m ≦ 4).

By repeating the steps (alternative generations) shown in FIG. 7 in this manner, an "individual" having a small sum of squares of input / output errors, that is, estimated normal data can be obtained.

However, as shown in FIG. 5, since there are an infinite number of normal data for reducing the sum of squares of the input / output error, the present algorithm cannot always find a solution like points B and C in FIG. There is a possibility. In such a case, for example,

[0031]

(Equation 2)

[0032] Here, αi = 0 (when Gi = 0) = β (when Gi ≠ 0) Gi: gene of the individual β: constant (β> 0). The second term of the evaluation value increases in proportion to the number of genes having a value other than “0”. Therefore, by optimizing the evaluation value to be small, an “individual” including many “0” can be obtained. In other words, it is possible to obtain estimated normal data in which the values of some measurement items among the abnormal data measured as points B and C in FIG. 5 are the same as the values of the estimated normal data.

In this embodiment, the genetic algorithm is used as the normal data search means.
An optimization method such as “simulated annealing” may be used.

Finally, the abnormality cause estimating section 60 will be described. The abnormality cause estimating unit 60 compares both the abnormal data actually measured and the normal data estimated by the normal data estimating unit 50, and specifies the measurement item that has become abnormal.
In the present embodiment, the deviation between the abnormal data actually measured for each measurement item and the normal data estimated by the normal data estimating unit is determined, and the measurement item whose deviation is larger than the reference value is identified as abnormal. For example, abnormal data [0.88, 0.75,
0.77, 0.90], the estimated normal data obtained is [0.87, 0.75, 0.71, 0.80], and the reference value is 0.0.
If it is 2, only the data of the fourth item is abnormal. Further, when there are a plurality of measurement items whose deviations are larger than the reference value, it is determined that the plurality of measurement items are abnormal. When there are a plurality of normal data candidates, an abnormal measurement item is specified based on each normal data. For example, if the estimated normal data is [0.87, 0.75, 0.71,
0.80] and [0.81, 0.85, 0.71, 0.90]
If there are two cases, each abnormal item is obtained separately.

After specifying the abnormality measurement item, the cause of the abnormality is estimated. In the present embodiment, the determination is made using the correspondence table between the measurement item and the cause of the abnormality. This is a summary of the relationship between measurement items indicating abnormal values and causes of abnormalities. An example is shown in FIG. In the example shown in FIG. 9, in addition to the relationship between the measurement item indicating the abnormal value and the cause of the abnormality, information on the time when the abnormality was measured is also described. For example, it is assumed that the abnormal data [0.88, 0.75, 0.71, 0.90] shown in FIG. 4 was measured at the time of the voltage peak, and the abnormal item was identified to be the generator voltage. Then, it is presumed from this table that "AVR is abnormal". If there are a plurality of normal data candidates and a plurality of abnormal items are considered correspondingly, the cause of the abnormality is estimated for each abnormal item. In this embodiment, the cause of the abnormality is specified by the above-described correspondence table.
The knowledge for specifying the cause may be described in advance by “If Then Rule”, and the cause of the abnormality may be estimated by the knowledge processing.

The cause of the abnormality estimated in this way is displayed on the monitor screen of the operation terminal 70. If the process of the estimation is necessary, it can be similarly displayed on the monitor screen of the operation terminal 70.

In the embodiment described above, the abnormality was detected as the input data of the neural network using the four measurement data every one second after activation, but the format of the input data is not limited by such a form. . Therefore, the measurement items are not limited to the four items, and the voltage of the starting storage battery and the temperature at that time may be added. Further, the measurement data at t seconds after the activation and the measurement data at (tk-t) seconds after the activation may be used as one set of input data. (Where t is a positive real number,
(k: positive integer, Δt: sampling interval) Also, instead of directly inputting measurement data, for example, feature amounts such as “ratio between maximum value and rated value of generator voltage” and “voltage establishment time” are extracted. Then, the value of the feature amount may be used as input data of the neural network.

[0038]

As described above, when a device or process is in an abnormal state, normal data is estimated by the normal data estimating unit, and the abnormal measurement item is specified by comparing the data with the abnormal data. It is possible to estimate the cause of the abnormality.

[Brief description of the drawings]

FIG. 1 is a diagram showing an outline of an embodiment of the present invention.

FIG. 2 is a diagram showing an outline of a diesel engine.

FIG. 3 is a diagram schematically illustrating a neural network.

FIG. 4 is a diagram illustrating an example of abnormality detection.

FIG. 5 is a diagram showing an outline of a normal data estimation method.

FIG. 6 is a diagram showing a chromosome expression method.

FIG. 7 is a diagram illustrating an algorithm of a normal data estimation unit.

FIG. 8 is a diagram illustrating an example of crossover.

FIG. 9 is a diagram illustrating an example of a relation table between an abnormality cause and a measured value.

[Explanation of symbols]

1. Emergency diesel engine generator, 2. Monitoring device,
3 preventive maintenance device, 10 data pre-processing unit, 20 data storage unit, 30 normal state learning unit, 40 abnormality detection unit
50: normal data estimation unit, 60: abnormality cause estimation unit, 70
... operation terminal.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Satoshi Goto 5-2-1 Omikacho, Hitachi City, Ibaraki Prefecture Inside the Hitachi, Ltd. Omika Plant

Claims (3)

[Claims] In a preventive maintenance method for detecting an abnormality of equipment from measurement data of equipment and estimating a cause of the abnormality, a preventive maintenance method having a hierarchical neural network model in which the number of input layers is equal to the number of output layers and measuring values in a normal state And a normal state learning unit for learning at least one of the feature values of the measured values as input data of the neural network model and learning the input data to be equal to the output data, and a neural network model for learning a normal state by the normal state learning unit. At least one of the measured value measured and the feature value of the measured value is input, and an abnormality detection unit that detects an abnormality from a deviation between the obtained output data of the neural network model and the input data, and an abnormality detection unit. It has a normal data estimation unit that estimates data when the device is in a normal state when it is determined to be abnormal. Preventive maintenance method characterized by comparing the abnormal data measured with the normal data estimated in positive normal data estimating unit, with abnormal cause estimation part that estimates the cause of the abnormality. 2. The preventive maintenance method according to claim 1, wherein
The preventive maintenance method according to claim 1, wherein the normal data estimating unit estimates, using a genetic algorithm, data when the device is in a normal state when the abnormality is determined to be abnormal by the abnormality detecting unit. 3. A preventive maintenance device for detecting an abnormality of an apparatus from measurement data of the equipment and estimating a cause of the abnormality, having a hierarchical neural network model having the same number of input layers and the number of output layers, and measuring measured values in a normal state. And normal state learning means for learning at least one of the feature values of the measured values as input data of the neural network model and learning the input data to be equal to output data; and a neural network model for learning a normal state by the normal state learning means. Abnormality detection means for inputting at least one of the measured value measured and the characteristic amount of the measurement value, and detecting an abnormality from a deviation between the obtained output data of the neural network model and the input data; and the abnormality detection means. Normal data estimation that estimates data when the device is in a normal state when it is determined to be abnormal Stage and compares the abnormality data measured with the normal data estimated in positive atmospheric data estimation means, prevention is characterized in that a abnormality cause estimating means for estimating a cause of the abnormality security device.