Information processing apparatus, information processing method, and recording medium for predicting abnormality of wind turbine

1 . An information processing apparatus comprising processing circuitry, the processing circuitry configured to:

receive learning time-series data from at least one sensor attached to a wind turbine, the at least one sensor sensing at least one of: a main bearing temperature, wind speed, outside air temperature, and active electric power;

construct, using the learning time-series data and at least one of long short-term memory (LSTM), a linear regression, an autoencoder, or graph neural network (GNN), a prediction model for predicting time-series data related to an abnormality of the wind turbine;

calculate a prediction error that is a difference between a predicted value of the time-series data predicted by the prediction model and an actual value of the time-series data;

divide the prediction error into a plurality of first sections in a time axis direction;

perform fast Fourier transform (FFT) processing on the prediction error divided into the plurality of first sections, and calculate a variation amount including a plurality of vibration amplitudes selected in descending order of magnitude for each of a plurality of second sections divided in a frequency axis direction;

classify the variation amounts divided into the plurality of second sections into a plurality of clusters by comparing the plurality of vibration amplitudes calculated for each of the plurality of second sections;

calculate an anomaly score of the actual value based on a distance between a centroid calculated for each of the plurality of clusters and the variation amount of each of the plurality of second sections;

construct a state determination model for determining the abnormality of the wind turbine based on the anomaly score;

predict a state of the wind turbine by using the prediction model and the state determination model; and

output at least an alarm indicating a predicted failure to prompt performance of predictive maintenance based on the predicted failure.

2 . The information processing apparatus according to claim 1 , wherein the processing circuitry is further configured to:

identify k (k is an integer of 1 or more) first sections having shortest distances from first sections other than a target first section among the plurality of first sections; and

calculate an average of distances from the target first section to the k first sections as the anomaly score.

3 . The information processing apparatus according to claim 1 , wherein

time widths of the divided plurality of first sections are the same.

4 . The information processing apparatus according to claim 1 , wherein

the prediction error is divided into the plurality of first sections each having a time width corresponding to kernel density estimation.

5 . The information processing apparatus according to claim 1 , wherein the processing circuitry is further configured to:

calculate a prediction error that is a difference between the predicted value of the time-series data predicted by the prediction model and the actual value of the time-series data;

divide the calculated prediction error into the plurality of first sections in a frequency axis direction;

calculate an anomaly score of the actual value based on the prediction error divided into the plurality of first sections; and

determine the state based on the calculated anomaly score and the state determination model.

6 . The information processing apparatus according to claim 5 , wherein

when the calculated anomaly score exceeds an upper limit value or falls below a lower limit value of the state determination model, it is determined to be abnormal.

7 . The information processing apparatus according to claim 5 , wherein

the divided plurality of first sections are converted into variation amounts divided into a plurality of second sections in the frequency axis direction and the variation amounts are grouped to calculate the anomaly score, or an average of distances from a target first section to k (k is an integer of 1 or more) first sections having shortest distances from first sections other than the target first section is calculated as the anomaly score.

8 . The information processing apparatus according to claim 1 , the processing circuitry is further configured to:

calculate a correlation coefficient between each of the calculated plurality of prediction error and an explanatory variable based on a plurality of the constructed prediction model candidates based on a plurality of prediction methods; and

select the prediction model to be used to construct the prediction model based on an average value of a plurality of the correlation coefficients corresponding to each of the plurality of prediction model candidates.

9 . An information processing method comprising:

receiving learning time-series data from at least one sensor attached to a wind turbine, the at least one sensor sensing at least one of: a main bearing temperature, wind speed, outside air temperature, and active electric power;

constructing, using the learning time-series data and at least one of long short-term memory (LSTM), a linear regression, an autoencoder, or graph neural network (GNN), a prediction model for predicting time-series data related to an abnormality of the wind turbine;

calculating a prediction error that is a difference between a predicted value of the time-series data predicted by the prediction model and an actual value of the time-series data;

dividing the prediction error into a plurality of first sections in a time axis direction;

performing fast Fourier transform (FFT) processing on the prediction error divided into the plurality of first sections, and calculating a variation amount including a plurality of vibration amplitudes selected in descending order of magnitude for each of a plurality of second sections divided in a frequency axis direction;

classifying the variation amounts divided into the plurality of second sections into a plurality of clusters by comparing the plurality of vibration amplitudes calculated for each of the plurality of second sections;

calculating an anomaly score of the actual value based on a distance between a centroid calculated for each of the plurality of clusters and the variation amount of each of the plurality of second sections;

constructing a state determination model that determines the abnormality of the wind turbine based on the anomaly score;

predicting a state of the wind turbine by using the prediction model and the state determination model; and

outputting at least an alarm indicating a predicted failure to prompt performance of predictive maintenance based on the predicted failure.

10 . A non-transitory computer readable recording medium storing a program for causing a computer to execute:

receiving learning time-series data from at least one sensor attached to a wind turbine, the at least one sensor sensing at least one of: a main bearing temperature, wind speed, outside air temperature, and active electric power;

constructing, using the learning time-series data and at least one of long short-term memory (LSTM), a linear regression, an autoencoder, or graph neural network (GNN), a prediction model for predicting time-series data related to an abnormality of the wind turbine;

calculating a prediction error that is a difference between a predicted value of the time-series data predicted by the prediction model and an actual value of the time-series data;

dividing the prediction error into a plurality of first sections in a time axis direction;

performing fast Fourier transform (FFT) processing on the prediction error divided into the plurality of first sections, and calculating a variation amount including a plurality of vibration amplitudes selected in descending order of magnitude for each of a plurality of second sections divided in a frequency axis direction;

classifying the variation amounts divided into the plurality of second sections into a plurality of clusters by comparing the plurality of vibration amplitudes calculated for each of the plurality of second sections;

calculating an anomaly score of the actual value based on a distance between a centroid calculated for each of the plurality of clusters and the variation amount of each of the plurality of second sections;

constructing a state determination model for determining the abnormality of the wind turbine based on the anomaly score;

predicting a state of the wind turbine by using the prediction model and the state determination model; and

outputting at least an alarm indicating a predicted failure to prompt performance of predictive maintenance based on the predicted failure.