Machine Learning Event Prediction for Condition-Based Maintenance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current predictive maintenance techniques face challenges in accurately predicting equipment failures in complex systems like aerospace and power systems, often resulting in unnecessary maintenance costs and unexpected failures.
Innovation Solution
A method involving machine learning models, specifically training Machine Learning Anomaly Detection Models and Event Prediction Models using sensor data to generate labels for unlabeled data, enabling the prediction of event probabilities and facilitating condition-based maintenance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If predictive maintenance techniques are implemented, then maintenance costs are reduced by performing tasks only when warranted, but measurement precision of equipment condition is insufficient leading to inaccurate failure predictions
Solution Approach 1:
The patent introduces multiple intermediaries to bridge the gap between raw sensor data and accurate failure predictions: (1) Anomaly detection models serve as intermediaries to identify unusual patterns in sensor data that precede failures; (2) Event prediction models act as intermediaries to forecast future failures based on historical data and current anomalies; (3) The integrated system combines these intermediaries with engineering analysis tools to provide precise, actionable predictions that reduce unnecessary maintenance while improving prediction accuracy
2Reliability
If routine preventive maintenance is performed, then unexpected failures are prevented, but unnecessary maintenance costs increase due to maintenance performed before actually needed
Solution Approach 1:
The patent applies preliminary action by performing anomaly detection and event prediction before actual failures occur. The system continuously monitors sensor data, detects anomalies that precede failures, and predicts future failures with timing estimates. This allows maintenance to be scheduled just before predicted failures without the need for routine preventive maintenance, preventing unexpected failures while avoiding unnecessary maintenance costs
Solution Approach 2:
The system implements feedback loops where prediction results and actual failure outcomes are continuously fed back into the training data. This feedback mechanism allows the anomaly detection and event prediction models to learn from real-world outcomes, continuously improving their accuracy in predicting when maintenance is actually needed, thereby reducing unnecessary maintenance while maintaining high reliability
3Measurement precision
If machine learning models are trained using only labeled data, then prediction accuracy improves, but loss of time increases due to difficulty in obtaining sufficient labeled failure data
Solution Approach 1:
The patent applies preliminary action by using anomaly detection models to pre-process and label unlabeled sensor data before training the event prediction models. The anomaly detection models identify unusual patterns in historical sensor data and automatically generate labels indicating potential failure conditions. This preliminary labeling action creates a large training dataset without manual annotation, reducing data preparation time while providing sufficient labeled data for accurate prediction model training
Solution Approach 2:
The system implements self-service by enabling the anomaly detection models to automatically generate training labels from unlabeled sensor data. The models serve themselves by identifying anomalies in the data and creating their own training datasets without external intervention. This self-labeling capability eliminates the time-consuming manual data annotation process while providing sufficient labeled data for training accurate prediction models
Data Source
AI summary
A computerized system performs training of machine learning models to enable prediction of occurrence of event(s), which are associated with a system to be analyzed. The system performs the following: (a) provide trained Anomaly Detection Model(s). (b) provide Analysis Tool(s), configured to provide quantitative indications of the event(s). The quantitative indications of the event(s) are based on input events(s). (c) receive first unlabeled data associated with the system, where this data comprise sensor data, (d) input the first unlabeled data to the Anomaly Detection Model(s). (d) generate, using the Anomaly Detection Models, indications of occurrence of the input event(s), based on the first unlabeled data, (e) input the indications of the occurrence into the Tool(s). (f) generate, using the Tools, quantitative indications of the events, based the indications of the occurrence, (g) generate, using the quantitative indications, labels for the first unlabeled data.