Fault predicting system and fault prediction method

1. A fault predicting system comprising:

a storage apparatus that stores data in such a manner that the data can be overwritten and read out; and

a processor that executes software program processing by using the data recorded in the storage apparatus, wherein

the storage apparatus stores

operation log information including a past operational state of a prediction target apparatus, and

fault record information including a fault content and a fault-occurrence date/time related to a fault having occurred in the prediction target apparatus in a past, and

the processor

sequentially changes a reference time point which is equivalent to a fault-prediction-execution time point

by setting, as a symptom sensing period, a period from a time point which is a predetermined first period-length before the reference time point until the reference time point, and

by setting, as a prediction target period,

a period from the reference time point until a time point which is a predetermined second period-length after the reference time point if a fault does not occur in the period from the reference time point until the time point which is the second period-length after the reference time point, and

a period from the reference time point until a fault-occurrence time point if a fault occurs in the period from the reference time point until the time point which is the second period-length after the reference time point, and

generates a fault prediction model for predicting a fault that occurs in the prediction target apparatus in a future by performing machine learning by

using, as an explanatory variable, a feature quantity based on the operation log information of the symptom sensing period, and

using, as a response variable, a fault index value based on

whether or not there is a particular-event occurrence in the operation log information corresponding to a fault occurrence in the fault record information of the prediction target period, and

a period-length from the reference time point until the fault-occurrence time point, wherein

the machine learning is performed repetitively over a plurality of staggered and/or overlapping periods, each comprising the symptom sensing period, the prediction target period, and the sequentially changed reference time points.

2. The fault predicting system according to claim 1 , wherein

the fault index value includes a coefficient which is a value according to the period-length from the reference time point until the fault-occurrence time point.

3. The fault predicting system according to claim 2 , wherein

the fault index value is computed by multiplying a particular-event occurrence probability by the coefficient.

4. The fault predicting system according to claim 1 , wherein

the processor divides the symptom sensing period into a plurality of symptom-sensing sub-periods, and performs the machine learning by using operation log information of each of the symptom-sensing sub-periods as a separate piece of information.

5. The fault predicting system according to claim 4 , wherein

the processor divides the symptom sensing period into the symptom-sensing sub-periods at a statistical change point which is a predetermined change point of the feature quantity obtained by statistically processing the operation log information.

6. The fault predicting system according to claim 4 , wherein

the processor divides the symptom sensing period into the symptom-sensing sub-periods at the fault-occurrence time point in the fault record information.

7. The fault predicting system according to claim 4 , wherein

a maintenance-implementation time point at which maintenance has been implemented in response to the fault is recorded in the fault record information, and

the processor divides the symptom sensing period into the symptom-sensing sub-periods at the maintenance-implementation time point in the fault record information.

8. The fault predicting system according to claim 4 , wherein

a maintenance-implementation time point at which maintenance has been implemented in response to the fault is recorded in the fault record information, and

the processor excludes, in the machine learning, operation log information of a symptom-sensing sub-period including the maintenance-implementation time point in the fault record information.

9. The fault predicting system according to claim 1 , wherein

the processor divides the prediction target period into a plurality of prediction-target sub-periods, and performs the machine learning by using the operation log information of each of the prediction-target sub-periods as a separate piece of information.

10. The fault predicting system according to claim 9 , wherein

the processor divides the prediction target period into the prediction-target sub-periods at a statistical change point which is a predetermined change point of the feature quantity obtained by statistically processing the operation log information.

11. The fault predicting system according to claim 9 , wherein

the processor divides the prediction target period into the prediction-target sub-period at the fault-occurrence time point in the fault record information.

12. The fault predicting system according to claim 9 , wherein

a maintenance-implementation time point at which maintenance has been implemented in response to the fault is recorded in the fault record information, and

the processor divides the prediction target period into the prediction-target sub-periods at the maintenance-implementation time point in the fault record information.

13. The fault predicting system according to claim 9 , wherein

a maintenance-implementation time point at which maintenance has been implemented in response to the fault is recorded in the fault record information, and

the processor excludes, in the machine learning, operation log information of a prediction-target sub-period including the maintenance-implementation time point in the fault record information.

14. A fault prediction method performed by a computer having:

a storage apparatus that stores data in such a manner that the data can be overwritten and read out; and

a processor that executes software program processing by using the data recorded in the storage apparatus, the method comprising:

by the storage apparatus, storing operation log information including a past operational state of a prediction target apparatus, and fault record information including a fault content and a fault-occurrence date/time related to a fault having occurred in the prediction target apparatus in a past, and

by the processor,

sequentially changing a reference time point which is equivalent to a fault-prediction-execution time point

by setting, as a symptom sensing period, a period from a time point which is a predetermined first period-length before the reference time point until the reference time point, and

by setting, as a prediction target period,

a period from the reference time point until a time point which is a predetermined second period-length after the reference time point if a fault does not occur in the period from the reference time point until the time point which is the second period-length after the reference time point, and

a period from the reference time point until a fault-occurrence time point if a fault occurs in the period from the reference time point until the time point which is the second period-length after the reference time point, and

generating a fault prediction model for predicting a fault that occurs in the prediction target apparatus in a future by performing machine learning by

using, as an explanatory variable, a feature quantity based on the operation log information of the symptom sensing period, and

using, as a response variable, a fault index value based on whether or not there is a particular-event occurrence in the operation log information corresponding to a fault occurrence in the fault record information of the prediction target period, and a period-length from the reference time point until the fault-occurrence time point, wherein

the machine learning is performed repetitively over a plurality of staggered and/or overlapping periods, each comprising the symptom sensing period, the prediction target period, and the sequentially changed reference time points.

15. The fault prediction method according to claim 14 , wherein

the fault index value includes a coefficient which is a value according to the period-length from the reference time point until the fault-occurrence time point.

16. The fault prediction method according to claim 14 , wherein

the processor divides the symptom sensing period into a plurality of symptom-sensing sub-periods, and performs the machine learning by using operation log information of each of the symptom-sensing sub-periods as a separate piece of information.

17. The fault prediction method according to claim 14 , wherein

the processor divides the prediction target period into a plurality of prediction-target sub-periods, and performs the machine learning by using the operation log information of each of the prediction-target sub-periods as a separate piece of information.