JP2009086896A - Computer failure prediction system and failure prediction method - Google Patents

Abstract

Disclosed is a computer failure prediction system and method capable of capturing various operation history data from a computer, determining the cause of the failure, and minimizing the provision of a replacement computer.
A center monitoring unit (21) stores a system log storage DB (23) for storing logs of sample data (a) of a plurality of hardware devices (22a-22n), and inputs the sample data (a) from the system log storage DB (23). The learning data calculation means 24 for generating learning data c based on the teacher data b obtained by calculating the weight of failure occurrence prediction from the actual measured value of the time until the occurrence of the failure 22n, and the learning data c are input. The failure model (desired mathematical formula) is defined, and the learned failure occurrence model defining means 25 for generating the failure occurrence time prediction data k of any hardware device 22 and the failure occurrence time prediction data k of any hardware device 22 And an output unit 27 that outputs the display data e.
[Selection] Figure 1

Description

  The present invention relates to a computer failure prediction system for monitoring the status of a network computer by connecting a maintenance device via a communication line to a maintenance target network computer that collects data to be measured from an object, and the failure It relates to the prediction method.

  Conventionally, in this type of computer failure prediction system and failure prediction method system, as disclosed in, for example, Japanese Patent Application Laid-Open No. 2004-247019 (Patent Document 1), when predicting a computer failure occurrence, A method of statistically and quantitatively evaluating the stability of the computer based on the statistics calculated by the calculation means and the past statistics accumulated by the data storage means has been used.

As shown in FIG. 7, the quantitative evaluation method of this computer failure prediction system is as follows.
A maintenance device is communicatively connected to the measuring device to predict a failure of the measuring device. This maintenance device uses the statistics calculated from the distribution data for each identification function of the measurement device obtained from the database to statistically evaluate the stability of the measurement device and predicts that a failure will occur in the measurement device If possible, a warning is issued to the operator terminal device and the maintenance terminal device.

  FIG. 7 is a block diagram showing an embodiment of a conventional computer failure prediction system.

  The computer failure prediction system 1 shown in FIG. 7 connects the maintenance device 3 to the measurement device 2 by communication, and performs the failure prediction of the measurement device 2.

  The measuring device 2 has a function of identifying an object. For example, the measuring apparatus 2 identifies the object or makes a determination in a plurality of stages. The identification function is an important function related to the performance of the measuring device 2 and is a function having an absolute judgment standard. The judgment function is a relative judgment standard based on human subjective / sensual judgment. It is assumed that the function of the machine is not as important as the identification function. Further, the identification function is determined by a plurality of inspection items, and if any item is identified as NG, the measuring device 2 outputs NG.

  The maintenance device 3 uses the statistics for each identification function of the measurement device 2 obtained from the database 4 and the statistics calculated from the distribution data for each identification function obtained from the measurement device 2 to stabilize the measurement device 2. When it is predicted that a failure will occur in the measuring device 2, the operator terminal device 5 is warned.

  The maintenance device 3 operates autonomously and periodically monitors the state of the measurement device 2. Periodic monitoring is based on the statistics calculated from the distribution data obtained from the measurement device 2 and the statistics obtained from the normal distribution data stored in the database 4 for a plurality of days. This is done by quantitative evaluation of sex.

  The maintenance device 3 includes a distribution data receiving unit 7 that receives distribution data from the measuring device 2 via the communication line 6, a statistic calculation unit 8 that calculates a statistic from the received distribution data, and the statistic. Statistics data recording unit 9 that records (stores) statistics of measuring device 2 obtained from calculation means 8 in database 4, and distribution data that records (stores) distribution data received by distribution data receiving unit 7 in database 4 Measurement using a recording unit 10, a statistic reading unit 11 that reads a plurality of days of statistics of the measuring device 2 obtained from the database 4, and a statistic obtained from the read statistic and the statistic calculation means 8. The quantitative evaluation means 12 that predicts whether a failure will occur in the device 2 and the operator terminal device 5 and the maintenance terminal device when the quantitative evaluation means 13 can predict that a failure will occur in the measurement device 2 Composed alert from the alert notification unit 14 as informing means for displaying a 13.

  The statistic reading unit 11 reads the statistics of the measuring device 2 obtained from the database 4 for a plurality of days for every identification function / inspection item and transmits it to the quantitative evaluation means 12.

  Further, the quantitative evaluation means 12 predicts the occurrence of a failure in the measuring device 2 using a statistical quantity such as a histogram for each identification function / inspection item and a fractional-day statistical quantity obtained from the database 4.

According to the conventional computer failure prediction system 1, by performing the above quantitative evaluation unit 12 for each measurement device 2, it is possible to predict in advance the failure of the measurement device 2 to be maintained, Even if a failure occurs, it can respond quickly.
Japanese Patent Application Laid-Open No. 2004-247019

However, according to a conventional computer failure prediction system, it is possible to predict in advance the occurrence of a failure in a measuring device to be maintained, and to respond quickly even if a failure occurs. ,
In the accuracy of predicting the occurrence of a failure, logical and reasonable predictions are not necessarily made from uniform data from the computer in operation, so a replacement computer should be prepared in case of failure. It was necessary to always and fully prepare, and the preparation costs for that were also appropriate.

  The present invention has been made in view of the above points, and by capturing various operation history data from an operating computer to determine the cause and cause of the failure, and utilizing this, further failure as a result It is an object of the present invention to provide a computer failure prediction system and failure prediction method capable of improving the prior prediction accuracy of occurrence and minimizing the preparation of a replacement computer to reduce the preparation cost.

  In order to achieve the above object, according to the present invention, a center monitoring unit that performs operation management of a network computer, and a plurality of hardware devices that constitute a computer network, the center monitoring unit includes the plurality of hardware devices. A system log storage DB that stores a log of sample data of a hardware device, and a weight for predicting a failure occurrence from an actual measurement value until the failure of the hardware device is input by inputting the sample data from the system log storage DB Learning data calculation means for generating learning data based on the teacher data obtained by the calculation processing, a failure model (desired mathematical formula) is defined by inputting the learning data, and the failure occurrence time of any hardware device To learned fault occurrence model defining means for generating prediction data, and to learned fault occurrence model defining means An input unit for inputting sample data of an arbitrary hardware device from the plurality of hardware devices, and input of failure occurrence time prediction data of the arbitrary hardware device from the learned failure occurrence model defining means There is provided a computer failure prediction system comprising an output unit for outputting display data.

  In order to achieve the above object, according to the present invention, the system log storage DB stores sample data from a plurality of hardware devices <st1>, and the learning data storage DB of the learning data calculation means Step <st2> in which sample data is input and an actual measurement value of the time until the failure of the device is stored, and the learning function means of the learning data calculation means causes the learning function of the neural network to learn the sample data and the teacher data. Generating learning data by step <st3>, generating a learned model by the learning function of the neural network by means of the learned failure occurrence model defining means and storing the data <st4>, and learning failure occurrence model from the input unit Input the sample data of any hardware device to the prescribing means. A computer comprising: a step <st5> of generating failure occurrence time prediction data by collating with a model; and a step <st6> in which an output unit outputs display data based on the failure occurrence time prediction data. Provide a failure prediction method.

  According to the present invention, various operation history data is captured from an operating computer to find out the cause and factor of the failure, and by using this, the prior prediction accuracy of further failure occurrence is improved as a result, It is possible to provide a computer failure prediction system and a failure prediction method capable of minimizing the provision of an alternative computer and reducing the preparation cost.

  An embodiment according to a computer failure prediction system of the present invention will be described with reference to the accompanying drawings.

  FIG. 1 is a block diagram showing the overall configuration of a computer failure prediction system according to the present invention.

  A computer failure prediction system 20 shown in FIG. 1 includes a center monitoring unit 21 that performs operation management of a network computer, and a plurality of hardware devices (hereinafter referred to as HW devices) 22a to 22n that form a computer network. .

  The center monitoring unit 21 causes a failure or failure of each of the HW devices 22a to 22n and other HW devices without any notice, and suddenly a warning or notification such as “critical”, “warning”, or “Notice” is issued. Any HW device can be monitored before it occurs.

  The center monitoring unit 21 is provided at one location in the center of a specific area, and a plurality of HW devices 22a to 22n as sites of several hundreds are provided as peripheral devices.

  Since the HW devices 22a to 22n have no log data that cannot be obtained by inspection at the shipping level, the HW devices 22a to 22n cannot be predicted and may be suddenly suspended due to “warning”. For this reason, measures can be taken from the HW devices 22a to 22n by collecting and using the sample data a (a1 to an) as log information of the HW devices at all times or at any time.

  For this reason, the center monitoring unit 21 of the computer failure prediction system 20 is provided with a system log storage DB 23 for storing sample data a (a1 to an) of the HW devices 22a to 22n.

  Further, the center monitoring unit 21 is obtained by inputting the sample data a from the system log storage DB 23 and calculating the weight of failure prediction from the actual measured time until the failure of the HW devices 22a to 22n. A learning data calculation means 24 for generating learning data c based on the teacher data b, a failure model (desired mathematical formula) is defined by inputting the learning data c, and failure occurrence time prediction data k of an arbitrary HW device 22 is obtained. A learning failure occurrence model defining means 25 to be generated, an input unit 26 for inputting sample data a of any HW device 22 from among the plurality of HW devices 22a to 22n to the learned failure occurrence model defining means 25, and learning An output unit 27 is provided for inputting failure occurrence time prediction data k of an arbitrary HW device 22 from the completed failure occurrence model defining means 25 and outputting display data e. That.

  On the other hand, the HW devices 22a to 22n provided as sites generate sample data (input data) a that is system log information data of specific hardware devices generated in the respective HW devices 22a to 22n.

  As shown in FIG. 1, for example, the HW device (1) 22a includes a sample output device 30a, and creates sample data a1 obtained during operation and sequentially outputs it to the system log storage DB 23. Configured to do. The sample output device 30a includes a hard disk drive (HDD) 30a1, a central processing unit (CPU) 30a2, and a memory (MEM) 30a3. The operation data of the sample output device 30a is collected and accumulated by the functions of the HDD 30a1, the CPU 30a2, and the MEM 30a3, and as shown in FIG. 2, the sample data a1 can be output to the system log storage DB 23.

  Similarly to the sample output device 30a, the sample output devices 30b to 30n also have a function of outputting to the system log storage DB 23 as sample data a2 to an.

  FIG. 2 is a diagram illustrating an example of sample data (input data) used in the computer failure prediction system 20.

  For example, the HW device (1) of the sample data a shown in FIG. 2 has a variable (1) and a variable (1 day before the day), data 2 days before the day, and data N days before the day, respectively. 2) The data of the variable (n) is obtained by numbers or symbols (not shown).

  Since the HW device (2) to the HW device (n) are the same as the HW device (1), the description thereof is omitted.

  The system log storage DB 23 receives sample data a from the HW devices (1) to (n) as shown in FIG. 2, and outputs the sample data a to the learning data calculation means 24 side.

  The learning data calculation means 24 includes a learning data storage DB 24a and a learning function means 24b.

  As shown in FIGS. 1 and 3, the learning data storage DB 24a receives the variables of the sample data a and the teacher data b output from the system log storage DB 23 and causes the learning function means 24b to learn. .

  FIG. 3 is an explanatory diagram showing a state before calculation of the learned data calculation means 24 of the failure prediction system 20 of the computer.

  The learning data storage DB 24a shown in FIG. 3 generates teacher data b obtained by extracting data related to a failure based on the sample data a. The teacher data b is to obtain a variable (1) as shown in FIG. FIG. 4 shows an example of sample data (teacher data) used in the computer failure prediction system 20.

  The learning data storage DB 24a receives the variable of the sample data a and the teacher data b, and selects the variable (1) to be incorporated in the arithmetic expression stored in advance so that the time until the failure occurs can be predicted. It is a database. Although the calculation formula is not shown, the time (seconds) from the variable (1) to the variable (n) is calculated, and the predicted time until the occurrence of the failure is comprehensively calculated. It is output.

  As shown in FIG. 5, the learning function means 24b is in a state before obtaining the learning data c, and is a stage where the learning data c is input to the neural network (NN) and learning is completed.

  FIG. 5 shows a state after calculation of the learned data calculation means 24 that has completed learning.

  The neural network (NN) shown in FIG. 5 is a learning obtained as a result of weighting data by removing defective values, normalizing processing, and learning using a neural network (NN), as shown in the learning function means 24b of FIG. Data c is output to the learned failure occurrence model defining means 25 side.

  The learned failure occurrence model defining unit 25 has a function of inputting the learning data c from the learning function unit 24b and collating the learned data c with the learned model to generate failure occurrence time prediction data k.

  That is, the failure occurrence time prediction data k generation function is the failure occurrence obtained as a result of weighting data after learning by the neural network (NN) as shown in the learned failure occurrence model defining means 25 in FIG. Temporal prediction data k is generated, subjected to prediction processing and denormalization processing, and then output to the learned failure occurrence model defining means 25 side.

  In determining the failure occurrence time prediction data k, the failure occurrence model stored in the learned failure occurrence model defining means 25 is compared with the sample data a of the corresponding HW device input from the input unit 26. It is the obtained result.

  For this collation (comparison) method, an existing Scheffe method or Tukey method can be used.

  The failure occurrence time prediction data k obtained by these methods is not only the time prediction data, but also outputs both analysis results such as different time prediction data depending on the model unit or the model number of the specific device. be able to.

  The input unit 26 receives the sample data a of the HW device specified from the HW device (1) to the HW device (n), and can perform a failure prediction of the HW device. For example, when the failure occurrence prediction of the HW device (1) is performed, the failure occurrence time prediction data k including the use environment data d of the HW device (1) can be obtained in addition to the sample data a1. Furthermore, for the HW device (not shown) other than the HW device (1) to HW device (n) from which the sample data a is collected, the sample data a ′ of this HW device is input to predict failure occurrence. It can be carried out. In this case, the learned fault occurrence model defining means 25 does not reflect the sample data a ′ in the defined model, but is approximate if the model is the same or similar to the sampled HW device. The failure occurrence time prediction data k thus obtained can be obtained.

  The output unit 27 receives the failure occurrence time prediction data k output from the learned failure occurrence model defining means 25 and, for example, the failure occurrence time of the specific HW device (1) based on the failure occurrence time prediction data k. The predicted value is output as display data, for example, image display data e.

  The display data e can also be displayed on the screen as a failure occurrence model by a timetable or graph until the failure occurrence of each of the target HW devices (1) to (n).

  Next, the operation of the computer failure prediction system 20 will be described with reference to FIG.

  FIG. 6 is a flowchart showing the operation of the failure prediction system of the computer failure prediction system 20.

  The operation will be described according to the flow shown in FIG.

  Step <st1> in which the system log storage DB 23 stores sample data a from a plurality of HW devices 22, and the learning data storage DB 24a of the learning data calculation means 24 inputs the sample data a until the failure of the device occurs. The step <st2> for storing the actual measurement value of the time and the learning function means 24b of the learning data calculation means 24 receive the sample input data component a or ((a + d) or a ') and the teacher data b of the neural network. A step <st3> for generating learning data c by learning using a learning function, a step <st4> for generating a learned model by the learning function of the neural network and storing data by the learned fault occurrence model defining means 25, and input To the learned failure occurrence model defining means 25 from any HW device 22 Sample data a is input and collated with the learned model to generate failure occurrence time prediction data k <st5>, and the output unit inputs the failure occurrence time prediction data k and outputs display data e. By taking the steps <st6>, the monitoring operator H of the center monitoring unit 21 can recognize that the HW device (1) 22a is the HW device that has received the failure prediction warning k. .

  According to the failure prediction system 10, the learning data calculation means 24 is provided. The learning data calculation means 24 captures various operation history data and is a learned failure model based on the cause and cause of the failure. Since it was possible to predict the occurrence of a failure of a computer that is operating within the scope of this criterion, it was possible to improve the prior prediction accuracy of further failure occurrence, It is possible to provide a computer failure prediction system and failure prediction method capable of minimizing computer preparation and reducing preparation costs.

The block diagram which shows the whole structure of the failure prediction system of the computer of this invention. The figure which shows the example of the sample data (input data) used for the failure prediction system of the computer of this invention. Explanatory drawing which shows the state before the calculation of the learned data calculating means of the failure prediction system of the computer of this invention. The figure which shows the example of the sample data (teacher data) used for the failure prediction system of the computer of this invention. Explanatory drawing which shows the state after the calculation of the learned data calculating means of the failure prediction system of the computer of this invention. The flowchart which shows the effect | action of the failure prediction system of the computer of this invention. The block diagram which shows embodiment of the failure prediction system of the conventional computer.

Explanation of symbols

20 Failure prediction system 21 Center monitoring unit 22 (22a to 22n) Hardware (HW) device 23 System log storage DB
24 learning data calculation means 24a learning data storage DB
24b Learning function means 25 Learned failure occurrence model defining means 26 Input section 27 Output sections 30a to 30n Sample output device a (a1 to an), a 'sample data (input data)
b Teacher data c Learning data (output data)
d Usage environment data e Display data k Failure occurrence time prediction data

Claims (7)

A center monitoring unit for managing the operation of a network computer, and a plurality of hardware devices constituting a computer network,
The center monitoring unit includes a system log storage DB that stores a log of sample data of the plurality of hardware devices,
Learning data is generated based on teacher data obtained by calculating the weight of failure prediction from the measured value of time until the failure of the hardware device by inputting the sample data from the system log storage DB Learning data calculation means;
A learning failure occurrence model defining means for inputting the learning data to define a failure model (desired mathematical formula) and generating failure occurrence time prediction data of an arbitrary hardware device;
An input unit for inputting sample data of any hardware device from the plurality of hardware devices to the learned failure occurrence model defining means;
A computer failure prediction system comprising: an output unit that inputs failure occurrence time prediction data of any hardware device from the learned failure occurrence model defining means and outputs display data. The learned failure occurrence model defining means inputs sample data of an arbitrary hardware device and usage environment data of an arbitrary hardware device from the plurality of hardware devices from an input unit, and generates failure occurrence time prediction data. The computer failure prediction system according to claim 1, wherein the computer failure prediction system is generated. The learned failure occurrence model defining means generates sample failure occurrence time by inputting sample data of any hardware device other than the plurality of hardware devices from an input unit. Item 5. A computer failure prediction system according to Item 1. The learning data calculation means is a learning data storage DB for storing teacher data in which sample data is inputted from the system log storage DB and weights of failure occurrence prediction are weighted from actual measured values until the failure of the hardware device. ,
The computer failure prediction system according to claim 1, wherein learning data is generated by inputting the teacher data from the learning data storage DB. The computer failure prediction system according to claim 1, wherein the display data output from the output unit is image display data. A step <st1> in which the system log storage DB stores sample data from a plurality of hardware devices;
Step <st2> in which the learning data storage DB of the learning data calculation means inputs the sample data and stores an actual measurement value of the time until the failure of the device;
A step <st3> in which the learning function means of the learning data calculation means generates the learning data by learning the sample data and the teacher data by the learning function of the neural network;
A step <st4> of generating a learned model and storing data by a learning function of the neural network by means of a learned failure occurrence model defining means;
Step <st5> of inputting sample data of an arbitrary hardware device from the input unit to the learned failure occurrence model defining means and generating failure occurrence time prediction data by collating with the learned model;
A step <st6> in which an output unit outputs display data based on the failure occurrence time prediction data;
A computer failure prediction method comprising: 7. The method according to claim 6, further comprising the step of inputting usage environment data of the hardware device in addition to sample data of the arbitrary hardware device from the input unit to the learned failure occurrence model defining means. Computer failure prediction method.

Images