Abnormality determination apparatus, learning apparatus and abnormality determination method for power plant or factory

A learning apparatus comprising:

• processing circuitry that is configured to realize:
• an acquisition unit that acquires, via a network, a plurality of normal-state input time-series data from at least one of sensors, plant systems or other devices in a target facility, the plurality of normal-state input time-series data being output from the target facility and corresponding to a plurality of normal-state process amounts generated in the target facility, wherein the plurality of normal-state process amounts comprise at least one of measurement values output from the sensors used in the target facility, setting values of the plant systems or the other devices, or output values of the plant systems or the other devices,
• a classification unit that classifies the plurality of normal-state input time-series data corresponding to the plurality of normal-state process amounts generated in the target facility into a plurality of groups such that each group includes, from among the plurality of normal-state input time-series data, normal-state input time-series data corresponding to normal-state process amounts which are physically correlated and which originate from physically correlated subsystems or devices in the target facility, wherein, from among the plurality of normal-state input time-series data, normal-state input time-series data corresponding to process amounts which are not physically correlated are classified into different groups, and wherein physical correlation of the normal-state process amounts is determined based on a correlation coefficient threshold between the normal-state input time-series data corresponding to respective normal-state process amounts;
• a first training unit that trains, for each of the groups, a corresponding first neural network dimension reduction/restoration model, using, as training data, the plurality of normal-state input time-series data corresponding to the normal-state process amounts of the group, such that each trained first neural network dimension reduction/restoration model outputs a plurality of output time-series data representing restored normal-state main trend components of corresponding input time-series data, excluding fluctuation signal components of the corresponding input time-series data;
• a subtraction unit that generates, for each of the groups, a plurality of normal-state input difference data by calculating a difference between the plurality of normal-state input time-series data and the plurality of output time-series data output from the trained first neural network dimension reduction/restoration model of a corresponding group, the plurality of input difference data including normal state fluctuation signal components other than main trend components of the input time-series data of the corresponding group;
• a second training unit that trains a second neural network dimension reduction/restoration model, using the plurality of normal-state input difference data generated for each of the groups as training data, such that the second neural network dimension reduction/restoration model restores normal-state fluctuation signal components from a plurality of input difference data and outputs a plurality of output difference data, the restored normal-state fluctuation signal components being constituted by the plurality of output difference data;
• and
• an addition unit that generates a plurality of addition data by combining the plurality of output difference data and the plurality of output time-series data.

An abnormality determination method executed by processing circuitry, the method comprising:

• acquiring, via a network, a plurality of input time-series data from at least one of sensors, plant systems or other devices in a target facility, the plurality of input time-series data being output from the target facility and corresponding to a plurality of process amounts generated in the target facility, wherein the process amounts comprise at least one of physical measurement values output from the sensors used in the target facility, setting values of the plant systems or the other devices, or output values of the plant systems or the other devices,
• classifying the plurality of input time-series data corresponding to the plurality of process amounts generated in the target facility into a plurality of groups such that each group includes, from among the plurality of input time-series data, input time-series data corresponding to process amounts which are physically correlated and which originate from physically correlated subsystems or devices in the target facility, wherein, from among the plurality of input time-series data, input time-series data corresponding to process amounts which are not physically correlated are classified into different groups, and wherein physical correlation of the process amounts is determined based on a correlation coefficient threshold between the input time-series data corresponding to the respective process amounts;
• applying one or more input time-series data included in each of the groups to corresponding trained first neural network dimension reduction/restoration models which differ per group and are trained on training data including normal-state input time-series data corresponding to the process amounts of a corresponding group, thereby outputting one or more output time-series data, wherein each of the trained first neural network dimension reduction/restoration models restores main trend components of corresponding input time-series data excluding a fluctuation signal component of the corresponding input time-series data, the restored main trend components being constituted by the output time-series data;
• generating, for each of the groups, a plurality of input difference data between the plurality of input time-series data and the output time-series data output from the trained first neural network dimension reduction/restoration models, the plurality of input difference data including fluctuation signal components other than the main trend components of the input time-series data of the corresponding group;
• applying the plurality of input difference data to a single trained second neural network dimension reduction/restoration model, thereby outputting a plurality of output difference data, wherein the single trained second neural network dimension reduction/restoration model is trained on training data including normal-state fluctuation signals and restores fluctuation signal components included in the input difference data, restored fluctuation signal components being constituted by output difference data;
• generating a plurality of addition data by combining the plurality of output difference data and the output time-series data;
• and
• determining an abnormality or a sign of abnormality of the target facility, based on a comparison between the plurality of addition data and the plurality of input time-series data to evaluate a deviation between restored expected behavior and actual behavior of physical processes or operations indicated by the input time-series data to detect the abnormality or the sign of abnormality in real-world physical operation of the target facility, and
• controlling a display device to display a determination result of the determining indicating presence or absence of the abnormality or the sign of abnormality.