An Efficient Machine Learning Algorithm for Valve Fault Detection
- Delivery
- Provide download link
- Format
- Price
- Non-members (tax incl.)：￥6,600 Members (tax incl.)：￥5,280
- Paper/Info type
- SAE Paper
No.2022-01-0163
- Pages
- 1-10(Total 10 p)
- Date of publication
- Mar 2022
- Publisher
- SAE International
- Language
- English
- Event
- WCX SAE World Congress Experience 2022
Detailed Information
| Author(E) | 1) Joe Serrano, 2) Elliott Ortiz-Soto, 3) S Kevin Chen, 4) Li-Chun Chien, 5) Abhishek Joshi |
|---|---|
| Affiliation(E) | 1) Tula Technology Inc., 2) Tula Technology Inc., 3) Tula Technology Inc., 4) Tula Technology Inc., 5) Tula Technology Inc. |
| Abstract(E) | Multi-level Miller-cycle Dynamic Skip Fire (mDSF) is a combustion engine technology that improves fuel efficiency by deciding on each cylinder-event whether to skip (deactivate) the cylinder, fire with low (Miller) charge, or fire with a high (Power) charge. In an engine with two intake and two exhaust valves per cylinder, skipping can be accomplished by deactivating all valves, while firing with a reduced charge is accomplished by deactivating one of the intake valves. This new ability to modulate the charge level introduces new failure modes. The first is a failure to reactivate the single, high-charge intake valve, which results in a desired High Fire having the air intake of a Low Fire. The second is a failure to deactivate the single intake valve, which results in a Low Fire having the air intake of a High Fire. Reliably detecting these two faults has proven challenging for classical techniques that se measured MAP (Manifold Absolute Pressure) and/or crank angle acceleration to identify characteristic features of the failures. However, the fault detection problem proves to be very tractable using machine learning techniques like artificial neural networks and logistic regression. This paper presents a computationally efficient machine learning model for fault detection in an mDSF engine using a three-class Logistic Regression solution based on commonly available engine controller signals. Training and testing accuracy exceeded 98% based on steady-state engine dyno data with valve faults induced at a 1% rate. The model requires about 100 multiply and accumulate operations each cylinder-event. |
