JPH09305224A - Method and device for preventive maintenance - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a preventive maintenance device which decides abnormality with high precision. SOLUTION: The preventive maintenance device 4 which detects abnormality of equipment from measurement data of equipment facilities is equipped with a relative relation learning part 30 which previously learns the relation between decision object data and an influence factor affecting the decision object data, a normal value estimation part 40 which estimates the normal value of the decision object data according to the learning result of the relative relation learning part 30, a threshold value determination part 50 which determines a threshold value for abnormality decision making according to the value estimated by the normal value estimation part 40, and a decision part 60 which compares the decision object data with the threshold value found by the threshold value determination part 50 with the decision object data to decide whether or not the equipment is abnormal; and the influence factor and decision object data are learnt to decide abnormality with high precision.

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 detecting an abnormality of a target equipment based on the result.

[0002]

2. Description of the Related Art In a preventive maintenance method that detects an abnormal condition of a device or predicts an abnormal condition in advance, data to be judged such as voltage, current, and temperature of a device are measured by sensors, and The value is compared with a predetermined value (this reference value is referred to as a threshold value), and it is determined whether the value is “normal” or “abnormal” based on the magnitude relationship.

[0003]

However, in most cases, the characteristics of the equipment differ depending on the manufacturer and its model. Further, even when the same device is targeted, the judgment target data such as the voltage, current, and temperature of the equipment may change due to operating conditions and aging, etc., even in a normal state. (Here, these factors that affect the determination target data are referred to as “influence factors”.) In the conventional technology, a method of determining a threshold value according to operating conditions, aging, etc. is not taken. Therefore, the determination accuracy may decrease.

In the present invention, a threshold value for making an abnormality determination based on the correlation between the measurement data to be determined as an abnormality and the surrounding environment (influence factor) when the measurement data is normal. Therefore, it is an object of the present invention to provide a preventive maintenance device and method for highly accurately determining an abnormality.

[0005]

In the preventive maintenance method for detecting an abnormality of equipment from measurement data of equipment, the above-mentioned object has a relationship between pre-measured judgment target data and influence factor data affecting the judgment target data. In advance, the step of estimating the determination target data based on the measured influential factor data and the learned result, and the threshold value for abnormality determination is determined based on the estimated value of the estimated determination target data. This can be achieved by including a step and a step of judging an abnormality of the equipment according to the determination target data measured from the equipment and the determined threshold value.

Further, in the preventive maintenance method for detecting an abnormality of equipment from measurement data of equipment, the above-mentioned object is a step of learning a relation between preliminarily measured decision data and influence factor data affecting the decision data. And a step of extracting influential factor data in the vicinity of the influential factor data to be measured and a first determination target data corresponding to the influential factor data in the vicinity, and influential factor data measured based on the learned result. From the extracted first determination target data and the estimated third determination target data, and a step of estimating the second determination target data corresponding to A step of determining a deviation, and a step of obtaining an upper limit value and a lower limit value of a threshold value from the determined deviation and the estimated second determination target data,
It can be achieved by including a step of judging an abnormality in the judgment target data measured from the equipment according to the obtained upper limit value and lower limit value.

Further, in the preventive maintenance device for detecting the abnormality of the equipment from the measurement data of the equipment, the above-mentioned object is a storage means for storing a plurality of kinds of measurement data measured from the equipment and the storage means. Relationship between data to be determined for determining abnormality of the device from a plurality of types of measurement data and a data selection unit that selects an influencing factor that affects the determination data, and the selected determination target data and the influencing factor Based on the determination target data estimated from the equipment and the estimation means for estimating the determination target data for the influential factors measured from the equipment, based on the determination target data estimated by the estimation means This can be achieved by including a determining unit that determines an abnormality of the equipment.

Further, in the preventive maintenance device for detecting an abnormality of the equipment from the measurement data of the equipment, the above-mentioned object is a storage means for storing a plurality of types of measurement data measured from the equipment and the storage means. Data selection means for selecting determination target data for determining an abnormality of the device from a plurality of types of measurement data and influencing factor data that influences the determination target data, influencing factor data in the vicinity of the influencing factor data, and the influencing factor data The neighborhood data extraction means for extracting the first determination target data corresponding to the influential factor data in the vicinity, and the relation between the determination target data and the influential factor data selected by the data selecting means are learned in advance, and the device The second determination target data for the influential factor data measured from the equipment and the influential factor data extracted by the neighborhood data extracting means Estimating means for estimating the third determination target data, means for determining a deviation from the first determination target data extracted by the neighborhood data extracting means, and the estimated third determination target data, and the determination Means for determining the upper limit value and the lower limit value from the determined deviation and the estimated second determination target data, and the equipment device based on the determined upper limit value and the lower limit value and the determination target data measured from the equipment. This can be achieved by including a determination unit that determines the abnormality.

[0009]

BEST MODE FOR CARRYING OUT THE INVENTION In the present invention, the correlation learning unit learns the normal correlation between the judgment target data and the influencing factors, and based on the learning result, the normal value estimation unit calculates the judgment target data Estimate normal value. The threshold value for determining the state is determined based on this estimated value. Therefore, even if the normal value of the judgment target data changes depending on the operating condition, if the correlation learning unit learns the relationship between the normal value and the operating condition as an influencing factor, the normal value according to the operating condition at the time of judgment Can be estimated. Further, even when the normal value changes due to the secular change of the device characteristics, it can be dealt with by re-learning using the latest data.

That is, according to the present invention, even if the normal value of the judgment target data changes due to the secular change of the operating conditions or the device characteristics, the threshold value according to the case can be appropriately determined, Accuracy is improved.

Further, the threshold width determination unit evaluates the reliability of the estimated value based on the number of data used for learning, the learned correlation, and the deviation (learning error) from the learning data. If this reliability is high, that is, if the estimated value has a high probability, set the threshold interval small, and if the reliability is low, increase the threshold interval.
Determine the threshold.

[0012] With this, the threshold interval is set according to the quality of the learning state, so that erroneous determination can be reduced.

(First Embodiment) A description will be given below with reference to the drawings.

FIG. 1 shows an application of the present invention to an operation management system for managing safe operation of an aircraft.

The operation management system is a central site for controlling the operation of the aircraft or controlling the radars of the radar sites 1a-1n based on the data detected by the radar sites 1a-1n and the radar sites 1a-1n for detecting the position of the aircraft by the radar. It is composed of a monitoring center 3. Since the radar sites 1a to 1n are usually installed in remote locations such as a mountaintop and a remote island, there are many unmanned facilities. Therefore, the data representing the operating state of the equipment used for maintenance / inspection of the equipment and the data representing the measurement conditions such as the temperature are transferred to the central monitoring center 3 via the telephone line 2. The central monitoring center 3 monitors the operating state of each of the distributed radar sites 1a to 1n to control the communication device, and at the same time, the central monitoring center 3
The preventive maintenance device 4 installed in the device centrally manages the data of the equipment on the site and judges the equipment status. The radar sites 1a to 1n are indispensable facilities for ensuring safe operation of the aircraft. Therefore, a decrease or stoppage of the function due to an abnormality in the equipment means a serious accident. In particular, stopping power supply is very damaging, so safety measures against power outages are important. Therefore, the radar sites 1a to 1n are provided with an emergency generator and an uninterruptible power supply for securing a power source in the event of a power failure, in addition to communication equipment that transmits or receives radio waves. In the present invention, the central monitoring center 3 is provided with the preventive maintenance device 4 for detecting the abnormality of the emergency generator. The preventive maintenance device 4 collects the data of the emergency power generator in advance, and compares the data with the data when the emergency power generator is actually started to determine whether or not there is an abnormality. Therefore, the preventive maintenance device 4 has a start-up mode for collecting data, a learning mode for generating a reference for making a judgment from the collected data, and a judgment mode for judging whether or not there is an abnormality by collecting actually activated data. There are three driving modes. Specifically, the startup mode is a mode for collecting data used when learning is performed in the learning mode. Radar site 1a-1n
The measurement data sent from the data pre-processing unit 10
At, it is determined whether the data is abnormal due to a sensor failure or an error during communication. If it is determined that the data is abnormal, it is stored in the abnormal data storage file in the operation record database 20. On the other hand, if it is not judged as abnormal data, that is, normal data is stored in the normal data storage file.

In the learning mode, the correlation learning unit 30
Data indicating the state of the target device by taking the accumulated measurement data into the normal data storage file of the operation result database 20 and its influencing factors (data such as temperature and humidity affecting the target device) Then, the learned neural network is transferred to the normal value estimating unit 40.

In the determination mode, the normal value estimation unit 40 estimates the normal value of the determination target data from the measured data, and the threshold value determination unit 50 uses the estimated value obtained by the normal value estimation unit 40 as a reference. , Determine the upper and lower limits of the normal range, that is, the threshold value. Then, the determination unit 60 compares the threshold value determined by the threshold value determination unit 50 with the measured data, and if the threshold value is within the range, it is “normal”, and if it is outside the range, it is “abnormal”.
Is determined. The determination result is displayed on the monitor screen of the central monitoring center 3.

Next, details of this embodiment will be described.

When the power supply to the radar sites 1a to 1n is interrupted and commercial power supply is stopped, the emergency power generators secure the power source. However, from the power failure to the steady operation of the emergency generators. Power cannot be supplied. Therefore,
The communication device is always connected to the uninterruptible power supply, and immediately receives power from the uninterruptible power supply immediately after the power outage. In the present embodiment, a case will be described in which an abnormality is determined for an emergency generator among the devices in these sites.

First, the outline of the emergency generator will be described with reference to FIG.

This emergency generator is a diesel engine driven three-phase AC generator.
0, power generation unit 110, storage battery 120, rectifier 130, starter 140. Then, when the commercial power is stopped due to the power failure and when the activation signal from the central monitoring center 3 is received, the activation device 140 activates the diesel engine 100.

Next, the process of supplying voltage by this emergency generator will be described. When there is a start command, the start device 1
The switch 141 in 40 is turned on, and the starting storage battery 1
Electric power is supplied from 43 to the starter motor 142, and the starter motor is driven. When the initial rotation is given by the starter motor 142, the diesel engine 100 can ignite the fuel. When the rotation speed reaches 20% of the rated rotation speed, it becomes possible to rotate by itself, so the starter 140 is disconnected.
After that, the speed increases by itself and the speed governor adjusts to the rated speed. The armature 111 is rotating at the same speed as the diesel engine 100,
Simultaneously with the startup of the field storage battery 120 to the field winding 11
A field current is supplied to 2 to start power generation, and the power generation voltage increases as the rotation speed of the diesel engine 100 increases. The generated power is transmitted to the communication device 1 via the switch 150.
There is a system that is sent to the battery 60 and a system that is converted into direct current by the rectifier 130 and sent to the storage battery 120 and the startup storage battery 143. However, until the generated voltage reaches the rating,
The switch 150 is off, and the communication device 160 is not supplied with power.

In such an emergency generator, the most serious failure is that the diesel engine 100 does not start. Therefore, the voltage of the starting storage battery 143, which is the power supply unit of the starting device 140, is measured to determine whether or not the starting storage battery is normal. An example in which the storage battery voltage is used as the determination target data will be described below.

The emergency generators installed at the radar sites 1a to 1n include storage battery voltage, generator engine speed, generator voltage, generator current, generator field current, storage battery total voltage, and storage battery liquid. Sensors are installed to measure temperature and temperature and humidity inside the radar site. 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. Even if the sensor is not able to set the measurement range in this way, the area to be used as the measurement value is set in advance, and when the measurement value from the sensor exceeds this set area, the "measurement range over" signal is generated. The device may be provided with a sensor.
Then, each of the radar sites 1a to 1n sends the measurement data from these various sensors to the preventive maintenance device 4 of the central monitoring center 3 via a telephone line. If the measurement data is judged to generate a “measurement range over” signal, this signal is added to the measurement data and the central monitoring center 3
Sent to the preventive maintenance device 4.

Next, each processing unit of the preventive maintenance device 4 will be described.

As described above, the preventive maintenance device 4 has three operation modes, that is, the startup mode, the learning mode, and the determination mode.

First, the start-up mode will be described.
In the startup mode, processing for collecting and storing normal data used in the learning mode from the radar sites 1a to 1n is performed. Normally, the emergency generator is stopped, so for these measurements, the central monitoring center 3 issues a start command,
Specify the collection time and collection interval. Therefore,
For example, when the emergency power generator is started at intervals of 0.1 second for 60 seconds, 600 points, when stopped for 0.5 seconds at 60 seconds, 120 points, and during steady operation at 30 seconds for 300 seconds, 10 points. The number of measurement data to be collected is decided. Then, the measurement data transferred from the radar sites 1a to 1n is checked by the data preprocessing unit 10 for abnormality and loss of the measurement data. The abnormality of the measurement data is determined by whether or not the “measurement range over” signal is added to the measurement data sent from the radar sites 1a to 1n, and when the measurement data is lost, it is collected in advance. Since the number of data to be collected is determined because the interval is specified, it is performed by comparing with the number of actually measured measurement data. In this way, it is determined whether or not there is an abnormality, and the measurement data that is determined to be abnormal is stored in the abnormal data storage file of the operation record database 20. On the other hand, if it is not judged to be abnormal, it is stored in the normal data storage file of the operation record database 20. Then, next, a feature amount for determining an abnormality is extracted from the measurement data collected in time series.

In this embodiment, the voltage drop rate of the start-up storage battery of the emergency generator is the object of determination, and in this case, the storage battery stored in the normal data storage file of the operation result database 20 as shown in FIG. The "voltage drop rate" is extracted as a feature amount from the time series data of the voltage.

Here, the "voltage drop rate" means the initial voltage V
s, the minimum voltage is V _min ,

[0030]

## EQU1 ## The voltage drop rate = (Vs- _Vmin ) / Vr (1) Then, the extracted feature amount is stored in the normal data storage file of the operation record database 20. The operator can set the collection time of measurement data, the setting of the interval, and the determination of the feature amount from the input unit.

Next, the learning mode will be described. In the learning mode, the correlation learning unit 30 learns the correlation between measurement data. The correlation learning unit 30 will be described with reference to FIG. In the correlation learning unit 30, first,
The learning item selection unit 310 selects an influential factor that is considered to be correlated with the measurement data to be determined, and sets it as a learning item. Next, the learning data selecting unit 320 selects data to be used for learning from the learning data accumulated. Finally, the learning unit 330 learns the correlation of the selected data.

When the learning item selection unit 310 is activated, FIG.
As shown in, the data items used for normal and abnormal judgment are displayed on the monitor. The items shown in FIG. 5 are examples, and other measurable data include “generator engine speed”, “generator voltage”, “generator current”, “generator field current”, and “storage battery total”. There are "voltage", "battery battery temperature", "room temperature" and "humidity" in the radar site. In addition, items such as the “voltage drop rate of the storage battery” obtained from the measurement data can be displayed on the monitor by registering the item name together with the calculation formula.

While looking at the monitor, the operator determines the judgment target data for judging an abnormality and the influence factor data which is considered to affect the judgment target data, that is, the influence factor data having a correlation. In the case of the present embodiment, the operator sets the “voltage drop rate of the storage battery” as the determination target data,
Then, “battery battery liquid temperature” is selected as an item having a correlation with “battery voltage drop rate”.

If the operator is left to select the items in this way, there is no problem if the operator is a skilled operator, but if there is no experience, it may not be possible to select an appropriate item. Therefore, from all data items, (System / Information / Control, vol.39, No4, pp.
185 to 190, 1995), a method may be adopted in which an item having a correlation is selected by the data mining technique, the operator checks the selected item, and selects a necessary item. The data mining technique is a method of finding a hidden correlation between data, and is generally a method of finding a correlation between data as a simple numerical value without considering the physical meaning of data items. Therefore, it is possible to find a correlation between data that the operator does not notice. However, on the other hand, if there is a correlation between data items that are considered to have no physical relationship, there is a risk that the item may be erroneously selected. So, from the correlation found by data mining,
The operator selects items that are considered to be appropriate.

Next, the learning data selection unit 320 selects the data item selection unit 310 from among the sufficiently accumulated learning data.
The data used for learning is selected based on.

In the above, learning is performed by using the judgment target data and the influencing factors corresponding thereto, but the learning data selection unit 320 extracts data having an influencing factor similar to the influencing factor at the time of judgment and uses it as the learning data. You may take the method of doing. When the condition similar to the influencing factor at the time of determination is “storage battery liquid temperature”, the similar data is selected as ± 5 degrees. Therefore, the “battery battery temperature”, which is an influencing factor at the time of determination, is 25
If it is degrees, only the data satisfying the condition of “storage battery liquid temperature” of 20 to 30 degrees is extracted as learning data.

In the learning section 330, the learning item selection section 310,
The correlation between the determination target data and the influencing factors is learned from the learning data selected by the learning data selection unit 320. For learning, we used a Lamer heart-type neural network consisting of three layers: input layer, intermediate layer, and output layer. The number of nodes in each layer is one input layer, three intermediate layers, and one output layer. The learning method was the pack propagation method.
This is because the value of "battery battery temperature", which is an influencing factor, is given to the node of the input layer, and the output of the output layer is equal to the value of "voltage drop rate of the storage battery". Is a way to adjust. Regarding learning by the backpropagation method, refer to the document "Learni
ng international representations by error propagat
ion ”, Parallel Distributed Processing: Exploration
s in the Microstructures of Condition, Vol.1, DERu
melhart and JLMcClelland (Eds.), Cambridge, MA ,:
MIT Press, pp 318-362. When a plurality of sets of data are given as the learning data, with the influencing factor and the determination target data as one set, and the learning is repeated, the correlation is stored in the “weighting coefficient” inside the neuro. When the learning is completed, the weighting coefficient is transferred to the normal value estimating unit having a neural network of the same type as the learning unit 330.

As a result, the learning unit 330 and the normal value estimation unit 4
0 is the same neural network. Further, although the neuron is used for learning the correlation in the present embodiment, a method of approximating the determination target data by linear combination of influencing factors and learning the correlation in the coefficient may be used.

Next, the judgment mode will be described. In the determination mode, the measurement data sent from the radar sites 1a to 1n are collected to determine the abnormality of the emergency generator. The measurement data sent from the radar sites 1a-1n are
Feature extraction is performed by the data preprocessing unit 10. Then, the “voltage drop rate of the storage battery” is sent to the determination unit 60, and the influence factor data “storage battery liquid temperature” is sent to the normal value estimation unit 40.
In the normal value estimation unit 40, the influence factor data “storage battery liquid temperature”
Data to be judged according to "Battery drop rate of storage battery"
The estimated value of is determined by a neural network. The estimated value of the determined “voltage drop rate of the storage battery” is determined by the threshold value determining unit 50 from the value C (constant) that determines the range of the normal state according to the following equation.

[0040]

[Equation 2] Threshold = Estimated value ± C (2) Then, the determination unit 60 compares the threshold determined by the threshold determination unit 50 with the “voltage drop rate of the storage battery” which is the measurement data. If the “voltage drop rate of the storage battery” is within the threshold value, it is determined as “normal”, and if it is out of the range, it is determined as “abnormal”. When the determination is “abnormal”, “battery abnormality” is displayed on the monitor screen of the central monitoring center 3.

(Embodiment 2) Next, FIG. 6 shows an arrangement in which a neighborhood data extraction unit 70 and a threshold width determination unit 80 are added in addition to the first embodiment. The neighborhood data extraction unit 70 and the threshold value width determination unit 80 operate in the determination mode.

In the neighborhood data extraction unit 70, the neighborhood influence factor data is learned on the basis of the influence factor data sent from the data preprocessing unit 10, and in the correlation learning unit 30, the neighborhood influence factor data is learned. The determination target data used for (hereinafter, referred to as “neighboring determination target data”) is extracted. For example, when the determination target data is “storage battery voltage drop rate” and the influencing factor data is “storage battery liquid temperature”, the data preprocessing unit 10 sends “storage battery liquid temperature” 20 degrees as the influencing factor data, as in the previous case. Then, the storage battery liquid temperature in the vicinity of "storage battery liquid temperature" 20 degrees and the "storage battery voltage drop rate" used for learning corresponding to this are extracted. Then, the extracted “storage battery liquid temperature” is passed to the normal value estimation unit 40, and the extracted “voltage drop rate of the storage battery” is passed to the threshold width determination unit 80. Here, the determination of the neighborhood data will be described. In the determination of the neighborhood data, assuming that the influence factor data is x, the maximum and minimum values of x are x _max and x _min , and the range is {x−βΔx _L ≦ x _i ≦ x + βΔx _L }.

However, given by Δx _L = x _max −x _min , β is determined so that the number of neighboring data becomes a predetermined number.

For example, as shown in FIG. 7, the range of the storage battery liquid temperature of the learning data is (10 to 35), and the input data at the time of determination is the storage battery liquid temperature: 20 degrees as shown by the black dots in FIG. Suppose there is. At this time, β is 0.2 when the width of the neighborhood is determined so that the number of neighborhood data is 30.

In the normal value estimating unit 40, the data preprocessing unit 1
The determination target data is estimated based on the influence factor data received from 0, and the determination target data is estimated based on the neighborhood influence factor data received from the neighborhood data extraction unit 70 (hereinafter, “estimated neighborhood determination target”). "Data".). Then, the normal value estimation unit 40 uses the determination target data estimated based on the influence factor data as the threshold value determination unit 5
To 0, the determination target data of the neighborhood estimated based on the influence factor data of the neighborhood is passed to the threshold width determination unit 80.

The threshold value width determining unit 80 determines the threshold value width from the determination target data obtained from the neighborhood data extraction unit 70 and the estimated neighborhood determination target data obtained from the normal value estimation unit 40 by a method described later. Determine the deviation ± ασ. σ is a standard deviation obtained from the neighborhood data, and α is a threshold width determination coefficient.

Then, in the threshold value determining section 50, the judgment target data obtained from the normal value estimating section 40 and the threshold value width determining section 8 are determined.
Based on the deviation obtained from 0, the upper limit value and the lower limit value of the determination target data, which are the reference of determination, are determined and passed to the determination unit 60.

In the judging unit 60, when the judgment target data obtained from the data preprocessing unit 10 is within the range between the upper limit value and the lower limit value obtained from the threshold value determining unit 50, it is judged to be normal, and the upper limit value, If it is outside the range of the lower limit value, it is determined to be abnormal.

Here, the deviation in the threshold width determination unit 80 ±
The determination of ασ will be described. The number of neighborhood data extracted by the neighborhood data extraction unit 70 is m, and the neighborhood data extraction unit 70 is
In the determination target data z _i of neighboring extracted and the determination target data of the neighborhood estimated normal value estimating unit 40 and z _i ^*, the standard deviation σ is determined by the following equation.

[0050]

Equation 3] _{σ = Σ (z i -z i} *) 2 / m ... (3) Further, alpha is obtained from the proximate data number and the desired accuracy of determination (probability of misdiagnosis with the normal data error). A method of determining α so as to obtain a desired determination accuracy will be described below.

In this embodiment, it is assumed that the distribution of the error between the estimated judgment target data z _i ^* in the vicinity and the judgment target data z _{i in the} vicinity is a normal distribution, and the above misdiagnosis rate (misdiagnosis rate of normal data). And the threshold width determination coefficient α is derived.

In general, the probability of occurrence of the data x is the normal distribution N
If (m, σ ² ) is obeyed, the data x

[0053]

## EQU00004 ## The probability of deviating from the range of m-.alpha..sigma..ltoreq.x.ltoreq.m + .alpha..sigma. (4) is obtained as a function of α.
For example, when α = 3, the probability is 0.3% or less.

However, when σ is not known and is calculated from the given n pieces of data, the calculated σ _n varies depending on the selected data. That is, σ to be calculated
_n has a distribution. Further, this distribution varies depending on the number of data n, and the larger n is, the narrower the width of the distribution is. Therefore, it is determined by σn calculated from n pieces of data.

[0055]

From a range of Equation 5] _{m-ασ n ≦ x ≦ m} + ασ n ... (5), the probability that the data is out is a function of the number of data n and alpha.

The derivation of the probability P _n (α) will be described below.

When the population follows a normal distribution with a mean of 0 and a standard deviation of 1, for sample volumes x ₁ , x ₂ , x ₃ , ..., X _n taken from this population, those 2 The sum of multiplications follows a χ ² distribution with n degrees of freedom. However, the χ ² distribution is given by equation (5)
It is a distribution that follows the probability distribution shown in.

[0058]

(Equation 6)

However,

[0060]

(Equation 7)

The probability that the measured value x is determined to be abnormal is σ≤ (x / α); 1 (x / α) ≤σ; 0. Therefore, the probability of misdiagnosing the measured value x as abnormal is equal to the probability of σ ≦ (x / α).

If the probability is P _n (σ ≦ x / α),

[0063]

(Equation 8)

However, (a, 0, z) is the incomplete gamma function of the first kind,

[0065]

[Equation 9]

Further, t is the sum of squares of x divided by the number of data n (dispersion).

Since the measured value x follows the normal distribution N (1,0), the probability Q (x) that the value will be x is

[0068]

(Equation 10)

Is obtained. Therefore, for the misdiagnosis rate P (n) when the number of data is n, the measured value is x, and the probability of determining the measured value x as abnormal is integrated for all x. Ie

[0070]

[Equation 11]

It can be obtained by

The relation between the number of data n and the misdiagnosis rate P (n, α) can be obtained by numerically integrating the equation (11).

It can be seen that this misdiagnosis rate is expressed as a function of the number of neighboring data n and α (P (n, α)). As an example, FIG. 8 shows the relationship between P (n, α) and α of 30 data items. As shown in FIG. 8, the misdiagnosis rate decreases as the number of data increases, and the misdiagnosis rate decreases as α increases. Therefore, when a desired determination accuracy (misdiagnosis rate) is given, this relationship can be used to obtain the threshold width determination coefficient α. For example, the misdiagnosis rate is 0.
When α is calculated so as to be 30%, α = 3.32. When the threshold width is calculated in this way, the value of σ varies depending on the distribution state of the data near the judgment data.
It is possible to make a judgment according to the distribution state.

As described above, the first and second embodiments have been described, but the contents described in the first embodiment and the contents described in the second embodiment are not completely different. For example, the preventive maintenance device 4 is the same as the first embodiment. It has both the functions of Example 2 and may be used by switching each, or may be used in parallel.

[0075]

As described above, the correlation learning means learns the correlation between the judgment target data and the influencing factors,
In the normal value estimation means, based on the learning result, the normal value of the judgment target data is estimated from the influencing factors at the time of data measurement, and the threshold value is determined based on the estimated value.
The threshold value can be appropriately determined according to the secular change.

Further, even if the variation of the learning data varies depending on the driving condition, and as a result, the reliability of the estimated value varies, the threshold value determining unit determines the optimum threshold width, so that the determination accuracy is improved. To do.

[Brief description of drawings]

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

FIG. 2 is a schematic diagram of an emergency generator.

FIG. 3 is a diagram showing characteristics of time and a storage battery voltage.

FIG. 4 is a diagram showing a correlation learning unit.

FIG. 5 is a diagram showing display contents of a learning item selection unit.

FIG. 6 is a diagram showing a second embodiment of the present invention.

FIG. 7 is a diagram for explaining neighborhood data.

FIG. 8 is a diagram for explaining a misdiagnosis rate.

[Explanation of symbols]

1a, 1b, 1n ... radar site, 2 ... telephone line, 3 ...
Central monitoring center, 4 ... Preventive maintenance device, 10 ... Data preprocessing unit, 20 ... Operation record database, 30 ... Correlation learning unit, 40 ... Normal value estimation unit, 50 ... Threshold value determination unit, 6
0 ... Judgment unit, 70 ... Neighborhood data extraction unit, 80 ... Threshold width determination unit.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Shoji Nakahara 5-2-1 Omika-cho, Hitachi City, Ibaraki Prefecture Hitachi Ltd. Omika Plant, Ltd. (72) Inventor Masahiro Yoshioka 5-chome, Omika-cho, Hitachi City, Ibaraki Prefecture No. 1 Incorporated company Hitachi Ltd. Omika factory (72) Inventor Masanori Kaminaga 5-2-1 Omika-cho, Hitachi City, Ibaraki Prefecture Incorporated company Hitachi Ltd. Omika factory

Claims (5)

[Claims] 1. A preventive maintenance method for detecting an abnormality of a device from measurement data of a device / equipment, a step of previously learning a relation between pre-measured determination target data and influence factor data affecting the determination target data, The determination target data based on the influential factor data and the learned result, the step of determining a threshold value for abnormality determination based on the estimated value of the estimated determination target data, and the measurement from the equipment. A preventive maintenance method, comprising: a determination target data to be determined; and a step of determining an abnormality of the equipment according to the determined threshold value. 2. A preventive maintenance method for detecting an abnormality of a device from measurement data of a device and equipment, the step of learning the relation between pre-measured judgment target data and influence factor data affecting the judgment target data, and measurement. Of the influential factor data in the vicinity of the influential factor data and the first determination target data corresponding to the influential factor data in the vicinity, and the second to the influential factor data measured based on the learned result. Estimating the determination target data and the third determination target data corresponding to the influential factor data in the vicinity, and determining a deviation from the extracted first determination target data and the estimated third determination target data. And a step of obtaining an upper limit value and a lower limit value of a threshold value from the determined deviation and the estimated second determination target data, and a step of obtaining the upper limit value and the lower limit value. Preventive maintenance method characterized by comprising the step of determining abnormality determination target data measured from the facility. 3. A preventive maintenance device for detecting an abnormality of a device from measurement data of the device, a storage unit for storing a plurality of types of measurement data measured from the device, and a plurality of types stored in the storage unit. The relationship between the determination target data for determining the abnormality of the device and the influencing factors that affect the determination target data from the measurement data of the device, and the relation between the selected determination target data and the influencing factors in advance. An estimation unit that learns and estimates the determination target data for the influential factor measured from the device / equipment, the determination target data measured from the device / equipment, and the device based on the determination target data estimated by the estimation unit. A preventive maintenance device comprising: a determination unit that determines an abnormality in equipment. 4. The method according to claim 3, wherein the estimating means is composed of a neural network, and learns the relationship between the determination target data and the influencing factors by using the determination target data and the influencing factors as input data. Preventive maintenance device. 5. A preventive maintenance device for detecting an abnormality of a device from measurement data of the device, the storage device storing a plurality of types of measurement data measured from the device, and a plurality of types stored in the storage device. Data selecting means for selecting the determination target data for determining the abnormality of the device from the measurement data of and the influence factor data that influences the determination target data, and the influence factor data near the influence factor data and the vicinity of the influence factor data. The neighborhood data extraction means for extracting the first determination target data corresponding to the influencing factor data, and the relationship between the determination target data and the influencing factor data selected by the data selecting means are learned in advance, and Second determination target data for the measured influential factor data and third determination for the influential factor data extracted by the neighborhood data extracting means. Estimating means for estimating the target data, means for determining a deviation from the first determination target data extracted by the neighborhood data extracting means and the estimated third determination target data, and the determined deviation And a means for determining an upper limit value and a lower limit value from the estimated second determination target data, and an abnormality of the equipment according to the determined upper limit value and the lower limit value and the determination target data measured from the equipment. A preventive maintenance device comprising a determination unit for determining.