Diagnostics, prognostics, and health management for vehicles using kinematic clusters, behavioral sensor data, and maintenance impact data

Abstract

Machine learning based methods for vehicle maintenance are disclosed. In one aspect, there is a method that includes uploading a flight dataset. The flight dataset includes timestamped vectors of kinematic and behavioral sensor data elements acquired by on-board sensors. The method further includes assigning each timestamped vector of the flight dataset to a kinematic cluster. The method further includes identifying behavioral sensor data that is anomalous for the assigned kinematic cluster of the timestamped vector. The method further includes generating an alert criticality score for each combination of kinematic cluster and behavioral sensor with anomalous sensor data.

1 - 20 . (canceled) 
     
     
         21 . A system, comprising:
 at least one data processor; and   memory storing instructions which, when executed by the at least one data processor, result in operations comprising:
 uploading an activity dataset, the activity dataset comprising a plurality of time stamped vectors of kinematic sensor data elements and behavioral sensor data elements, each kinematic sensor data element and each behavioral sensor data element acquired by at least one sensor on board a vehicle; 
 assigning, to a kinematic cluster, each timestamped vector of the activity dataset by at least applying a machine learning model trained to differentiate between kinematic sensor data elements associated with different kinematic dimensions; 
 identifying behavioral sensor data that is anomalous for the assigned kinematic cluster of the timestamped vector; 
 generating an alert criticality score for each combination of kinematic cluster and behavioral sensor with anomalous sensor data; 
 aggregating the generated alert criticality scores into a first alert vector; 
 identifying previous activity with an alert vector corresponding to the first alert vector, the identified previous activity having subsequent maintenance actions and follow-on activity; 
 generating an impact score for each of the identified previous activity of the respective maintenance action on the respective follow-on activity; 
 generating a composite impact score for each maintenance action; and 
 selecting the maintenance action with the largest composite impact score. 
   
     
     
         22 . The system of  claim 21 , wherein each timestamped vector comprises a vehicle identifier, a timestamp, at least one kinematic sensor reading, and at least one behavioral sensor reading. 
     
     
         23 . The system of  claim 23 , wherein each timestamped vector further comprises a vehicle type, a version number, an engine type, and/or an identifier. 
     
     
         24 . The system of  claim 21 , further comprising labeling a dataset element as a kinematic sensor data element or a behavioral sensor data element. 
     
     
         25 . The system of  claim 21 , wherein assigning each timestamped vector of the dataset to a kinematic cluster comprises generating the kinematic clusters, wherein generating the kinematic clusters comprises training at least one machine learning model on a training set of timestamped kinematic sensor vectors from a plurality of datasets. 
     
     
         26 . The system of  claim 26 , wherein the training of the machine learning model comprises applying a clustering method. 
     
     
         27 . The system of  claim 26 , wherein the clustering method comprises a MAPPER algorithm. 
     
     
         28 . The system of  claim 28 , wherein the clustering method further comprises a k-means clustering algorithm or a hierarchical clustering algorithm. 
     
     
         29 . The system of  claim 21 , wherein the behavioral sensor value is anomalous for the corresponding kinematic cluster when the behavioral sensor value differs from the expected sensor value for the corresponding kinematic cluster by more than a normalized threshold. 
     
     
         30 . The system of  claim 30 , wherein a distribution of the behavioral sensor value for the corresponding kinematic cluster has a Gaussian distribution, and wherein the normalized threshold corresponds to a z-score of two standard deviations. 
     
     
         31 . The system of  claim 21 , wherein the behavioral sensor data elements for a kinematic cluster are filtered with a double exponential moving average filter. 
     
     
         32 . The system of  claim 21 , wherein generating the alert criticality score for each combination of kinematic cluster and behavioral sensor with anomalous sensor data comprises:
 ordering the timestamped vectors of kinematic sensor data and behavioral sensor data for each kinematic cluster by time;   computing, over a time window, each alert criticality score as the absolute value of the difference between a stationary average and a moving average of the corresponding behavioral sensor, divided by the moving average of the corresponding behavioral sensor; and   comparing each alert criticality score to a minimum threshold for an alert.   
     
     
         33 . The system of  claim 21 , wherein the first alert vector comprises at least one ordered pair of a behavioral sensor labels paired with an anomalous behavioral sensor value that exceeds an alert criticality threshold. 
     
     
         34 . The system of  claim 21 , wherein each impact score quantifies the impact of the maintenance action on the corresponding follow-on activity, by summing the differences between alert criticality scores for each behavioral sensor for the follow-on activity after the maintenance action from the corresponding alert criticality scores for each behavioral sensor from the previous activity before the maintenance action. 
     
     
         35 . The system of  claim 34 , wherein generating a composite impact score for each maintenance action comprises determining an average impact score for the follow-on activity following each respective maintenance action. 
     
     
         36 . A computer-implemented method comprising:
 uploading a activity dataset, the activity dataset comprising a plurality of time stamped vectors of kinematic sensor data elements and behavioral sensor data elements, each kinematic sensor data element and behavioral sensor data element acquired by at least one sensor on board a vehicle;   assigning each timestamped vector of the activity dataset to a kinematic cluster;   identifying behavioral sensor data that is anomalous for the assigned kinematic cluster of the timestamped vector;   generating an alert criticality score for each combination of kinematic cluster and behavioral sensor with anomalous sensor data;   aggregating the generated alert criticality scores into a first alert vector;   identifying previous activity with an alert vector corresponding to the first alert vector, the identified previous activity having subsequent maintenance actions and follow-on activity;   generating an impact score for each of the identified previous activity of the respective maintenance action on the respective follow-on activity;   generating a composite impact score for each maintenance action; and   selecting the maintenance action with the largest composite impact score.   
     
     
         37 . A non-transitory computer-readable medium storing instructions, which when executed by at least one data processor, result in operations comprising:
 uploading a activity dataset, the activity dataset comprising a plurality of time stamped vectors of kinematic sensor data elements and behavioral sensor data elements, each kinematic sensor data element and behavioral sensor data element acquired by at least one sensor on board a vehicle;   assigning each timestamped vector of the activity dataset to a kinematic cluster;   identifying behavioral sensor data that is anomalous for the assigned kinematic cluster of the timestamped vector;   generating an alert criticality score for each combination of kinematic cluster and behavioral sensor with anomalous sensor data;   aggregating the generated alert criticality scores into a first alert vector;   identifying previous activity with an alert vector corresponding to the first alert vector, the identified previous activity having subsequent maintenance actions and follow-on activity;   generating an impact score for each of the identified previous activity of the respective maintenance action on the respective follow-on activity;   generating a composite impact score for each maintenance action; and   selecting the maintenance action with the largest composite impact score.