Trigger condition determination method for time series signal, abnormality diagnosis method for equipment to be monitored, and trigger condition determination device for time series signal
A trigger condition determination method for a time series signal determines a trigger condition for cutting out a monitored section being a target for abnormality diagnosis, from a monitored signal being a time series signal indicating a condition of a monitored facility in the abnormality diagnosis for the monitored facility, and includes: collecting signal groups including one or more monitored signals and a trigger candidate signal; cutting out the monitored section of the monitored signal; generating a learning model specifying a start time point of the cut-out monitored section, generating label data, and using one or more trigger candidate signals at each time point as an input and using the label data at each time point as an output, by using machine learning; and determining the trigger condition by using the learning model, for the monitored signal for which the abnormality diagnosis is performed.
1 . A trigger condition determination method for a time series signal, the method determining a trigger condition that is a condition for cutting out a monitored section being a target for abnormality diagnosis, from a monitored signal being a time series signal indicating a condition of a monitored facility in the abnormality diagnosis for the monitored facility, the method comprising:
a collection step of collecting signal groups including one or more monitored signals and a trigger candidate signal, the one or more monitored signals relating to the monitored facility, the trigger candidate signal being a time series signal relating to the monitored facility and detected at a same time as the monitored signal, and representing a time series signal that is able to be the trigger condition;
a cut-out step of cutting out the monitored section of the monitored signal, based on a predetermined criterion, from the signal groups;
a model generation step of generating, for the signal groups, a learning model specifying a start time point of the cut-out monitored section, generating label data in which a label for the start time point is turned on and a label for any other time point is turned off, and using one or more trigger candidate signals at each time point as an input and using the label data at each time point as an output, by using machine learning; and
a trigger condition determination step of determining the trigger condition by using the learning model, for the monitored signal for which the abnormality diagnosis is performed.
2 . The trigger condition determination method for a time series signal according to claim 1 , wherein
in the cut-out step,
a monitored section is cut out based on a facility characteristic of the monitored facility, from a first monitored signal, selected from a plurality of the monitored signals collected in the collection step; and
a section having a largest correlation coefficient with a waveform included in the monitored section of the first monitored signal is searched for to cut out the monitored section, from each of the monitored signals other than the first monitored signal, of the plurality of the monitored signals.
3 . The trigger condition determination method for a time series signal according to claim 1 , wherein the learning model is a decision tree.
4 . The trigger condition determination method for a time series signal according to claim 1 , wherein in the model generation step, when the trigger candidate signal is one pulse signal, the machine learning is performed after the trigger candidate signal is converted to a sawtooth wave.
5 . The trigger condition determination method for a time series signal according to claim 1 , wherein in the model generation step, when accuracy in determination provides no predetermined value in the machine learning, the process returns to the cut-out step, the monitored section that has been cut out last time is shifted back and forth, a new monitored section of the monitored signal is cut out, and the model generation step is performed again.
6 . An abnormality diagnosis method for a monitored facility, comprising:
cutting out and accumulating signals in a monitored section being a target for abnormality diagnosis, from a monitored signal being a time series signal indicating a condition of a monitored facility, according to a trigger condition determined by the trigger condition determination method for a time series signal according to claim 1 ; and
performing the abnormality diagnosis for the monitored facility based on the accumulated signals.
7 . A trigger condition determination device for a time series signal, the device determining a trigger condition that is a condition for cutting out a monitored section being a target for abnormality diagnosis, from a monitored signal being a time series signal indicating a condition of a monitored facility in abnormality diagnosis for the monitored facility, the device comprising:
a collection unit configured to collect signal groups including one or more monitored signals and a trigger candidate signal, the one or more monitored signals relating to the monitored facility, the trigger candidate signal being a time series signal relating to the monitored facility and detected at a same time as the monitored signal, and representing a time series signal that is able to be the trigger condition;
a cut-out unit configured to cut out the monitored section of the monitored signal, based on a predetermined criterion, from the signal group;
a model generation unit configured to, for the signal group, generate a learning model specifying a start time point of the cut-out monitored section, generate label data in which a label for the start time point is turned on and a label for any other time point is turned off, and use one or more trigger candidate signals at each time point as an input and using the label data at each time point as an output, by using machine learning; and
a trigger condition determination unit configured to determine the trigger condition by using the learning model, for the monitored signal for which the abnormality diagnosis is performed.