LAB TO LOGIS TICS Phys ics Informed Machine Learning (PIML) for Advanced Diagnos tics & Prognos tics of Ground Combat Vehicles J uan F. Betts & Aras h Alizadeh PrediciveIQ MOTIVATION: WHY USE PIML FOR ENGINE HEALTH? LAB TO LOGIS TICS AOAP OR BIG DATA ML WILL NOT LEAD TO A PREDICTIVE OIL HEALTH SOLUTION, THEREFORE A PHYSICS INFORMED ML APPROACH IS REQUIRED • The number of AOAP sample is too small over the life of a program to properly train a data driven machine learning solution. • PIML PPMX becomes predictive prior to production and can handle nonstationary Army mission specific plans PROGNOSTICS & PREDICTIVE MAINTENANCE (PPMX) WORKFLOW 1. Problem Assessment 2. Experiments on Engine 3. Physics Model Development 4. PIML Model Development 5. PIML V&V UQ LAB TO LOGIS TICS 6. PIML Integration into PPMX Dashboard 7. PIML PPMX Demonstration ENGINE HEALTH DRIVERS • Engine health is driven by the oil degradation & contam ination, and the depos its created through engine wear. • Category 1 s ens ors are optional phys ical s ens ors that can m eas ure s om e of the des ired drivers of engine health (e.g. oil vis cos ity) and alert operators of potential is s ues • Category 2 s ens ors are s ens ors that are difficult to be im plem ented in realworld conditions outs ide of a tes t s etting. Therefore, thes e s ens ors are derived from Category 1 s ens ors . Category 2 s ens ors are s om etim es called “virtual s ens ors .” • A m odular PIML PPMx PDT fram ework was propos ed to us es phys ics -bas ed m odels and ML fram eworks to predict Category 1 and Category 2 s ens ors . Thes e m odels would be Scalable to other GVS engines with s om e m odel trainings . LAB TO LOGIS TICS ENGINE WEAR LAB TO LOGIS TICS EXPERIMENTS ON ENGINE • Cum m ins perform ed NATO50 and durability tes ts on ACE in accordance with the m ethods and s tandard conditions of AEP-5 Part-II • A data driven m odel was built to identify the m os t im portant features influencing TAW in the engine and reduce the num ber of features us ed in the final PIML m odel. • Features m ay depend on each other, s o we perform ed a correlation analys is to rem ove the correlated features from the m odel. • The m odel was built us ing the durability tes t data-s et for the ACE provided by Cum m ins . • Wear volum e was defined as ∑(𝑁𝑁_{𝑝𝑝𝑎𝑎𝑟𝑟𝑡𝑡𝑖𝑖𝑐𝑐𝑙𝑙𝑒𝑒𝑠𝑠} ×𝑆𝑆𝑖𝑖𝑧𝑧𝑒𝑒_{𝑝𝑝𝑎𝑎𝑟𝑟𝑡𝑡𝑖𝑖𝑐𝑐𝑙𝑙𝑒𝑒𝑠𝑠} ) • We reduced the num ber of features to three by em ploying m ethods like Factor Analys is (FA), Variance Inflation Factor (VIF) and Engineering J udgem ent LAB TO LOGIS TICS Coolant Outlet Temp PHYSICS INFORMED TAW MODEL • Engine TAW Model –Sensor Data: Oil Temperature, Oil Pressure, Engine Speed, and Oil Viscosity –Data Driven Models: Fuel Dilution, Oil Acidity –Physics Based Models: Oil Film Thickness, Friction –Hybrid Models: Wear, Oil Status • A PINNs based virtual sensor (VISCO module) was developed that predicted the engine oil viscosity for every given temperature • The physics model was validated with literature examples. LAB TO LOGIS TICS PIML MODEL DEVELOPMENT • Deep Operator Network (DeepONet) is based on classical universal approximation theorem of continuous functions. The operator G is a nonlinear continuous operator. • Transfer learning is the idea of overcoming the isolated learning paradigm and use the knowledge acquired for one task to solve related ones. • In transfer learning, we can leverage knowledge (features, weights etc) from previously trained models for training newer models with significantly less data for the newer task. • Transfer learning enables building a predictive model for a specific mission prior to having any field data. Physics Model Train • The architecture is based on a design of two subnetworks, the branch net for the input function and the trunk net for the location to evaluate the output function. LAB TO LOGIS TICS Transfer Learning DeepONet BACKEND PREDICTIVE METHODOLOGY Pre-trained Model LAB TO LOGIS TICS Predictive Model New NATO 50 Data DeepONet Physics Model PIML VALIDATION WITH NATO50 DATA LAB TO LOGIS TICS PIML TAW V&V UQ LAB TO LOGIS TICS ENGINE HEALTH PPMX ARCHITECTURE PredictiveIQ Models LAB TO LOGIS TICS OIL HEALTH AND MODULAR EXTENSION LAB TO LOGIS TICS PPMX DEMONSTRATION LAB TO LOGIS TICS LAB TO LOGIS TICS THANK YOU J uan F. Betts & Aras h Alizadeh PrediciveIQ