Machine Defect Detection Using ML-Predicted Condition Indicators

The patent segments the condition indicator analysis into two distinct components: a baseline component representing normal operating condition fluctuations, and a defect component representing actual machine defects. By training a machine learning model to learn the baseline behavior from historical data during normal operation, the system can subtract this expected behavior from current readings, isolating only the defect-related deviations. This segmentation allows lower alarm thresholds without increasing false alarms, as the baseline variability is already accounted for.

The system performs preliminary action by collecting and analyzing historical operating data during normal machine operation to establish a baseline behavior model before actual defect detection begins. The machine learning algorithm is trained in advance on data from multiple operating conditions to learn how condition indicators naturally fluctuate with changes in load, speed, and environmental factors. This preliminary learning phase enables the system to distinguish between normal variability and defect-induced changes, allowing for more sensitive defect detection without false alarms.

The system implements feedback by continuously comparing current condition indicator readings against the machine learning model's predictions of normal behavior. The difference between actual and predicted values (residuals) is analyzed to detect defects. When a defect occurs, it causes deviations from the learned baseline pattern, generating feedback signals that trigger alerts. This feedback mechanism enables low alarm thresholds because the system only triggers when there is a genuine deviation from the baseline, not during normal operating condition changes.

The machine learning model acts as an intermediary between the raw condition indicator data and the defect detection decision. Instead of directly comparing current readings to fixed alarm thresholds, the system uses the ML model to predict what the reading should be under current operating conditions, then compares the actual reading to this dynamic prediction. This intermediary layer filters out normal operating variability while preserving defect-related signals, enabling sensitive detection without false alarms.

The patent transitions from static fixed thresholds to dynamic adaptive thresholds by using a machine learning model that continuously adapts to changing operating conditions. The model learns the relationship between operating parameters (load, speed, temperature) and condition indicator behavior, enabling the detection threshold to dynamically adjust based on current operating conditions. This dynamic approach maintains high adaptability while the underlying ML model structure provides a systematic framework that manages complexity through automated learning rather than manual configuration.

The system changes the parameter used for defect detection from fixed absolute thresholds to relative deviations from predicted baseline behavior. Instead of using a static threshold value, the system calculates the difference between actual condition indicators and ML-predicted values under current operating conditions. This parameter transformation enables the system to adapt to different operating conditions automatically, as the threshold becomes a function of operating parameters rather than a fixed value, improving versatility without proportionally increasing complexity.