Tweet
Tweet

Slide 1

Slide 1 text

Oracle Machine Learning Office Hours Usage Highlight Machine Learning Recommendations for Maintenance and Repair with Lee Sacco supported by Marcos Arancibia, Sherry LaMonica & Mark Hornick Product Management, Oracle Machine Learning June 2021

Slide 2

Slide 2 text

The following is intended to outline our general product direction. It is intended for information purposes only, and may not be incorporated into any contract. It is not a commitment to deliver any material, code, or functionality, and should not be relied upon in making purchasing decisions. The development, release, timing, and pricing of any features or functionality described for Oracle’s products may change and remains at the sole discretion of Oracle Corporation. Safe harbor statement 2

Slide 3

Slide 3 text

“Start with the Data” 3

Slide 4

Slide 4 text

4 wikipedia.org medium.com “Start with the Data”

Slide 5

Slide 5 text

Depot Repair Business Models • The repair, maintenance, or recycling of products returned by customers to a processing facility. • Often conjoined or contrasted with “field service”, which involves repair, maintenance, or recovery at a customer or off-premise location. • Requires tracking each item from customer to repair center to final disposition (back to customer, to warehouse, to supplier, or recycler). • Manufacturers / OEMs • Contract Repairers • In-store • Authorized Service Providers What is it?

Slide 6

Slide 6 text

User Schema Repair Technician • Wide spectrum of education and training – from cleaner to nuclear engineer. • Tactile, concrete, physical. Makers, builders, fixers. Good problem solvers. • Generally in a workshop or on a repair line. • Love fixing things but HATE data entry. • Don’t like being told what to do, especially by a computer. https://www.plexus.com/en-us/solutions#Aftermarket-Services • Track items from dock door to final disposition. • Diagnose problem and cause. • Determine best fix. • Track repair activities, parts, labor, analytic data, other documentation. Tasks • Very broad product sets. • Each item has a unique history. • Each item type has a different fault tree and diagnostic pathway. • “Best” fix is relative. Multi-level constraints. Problems Apply ML

Slide 7

Slide 7 text

VERY Structured Data ERP: Canonical Tasks, Data, KPIs ERP Database Canonical Data Model Task 1 Task 2 Task 3 ERP Functions Canonical Use Cases (decisions and actions)

Slide 8

Slide 8 text

VERY Structured Data “Structured” <> “Predictable” ERP Database Canonical Data Model • ~1000 different customers in many different industries • Shape of data: • Every business is different • Every item is different • Snapshots vs. Streams

Slide 9

Slide 9 text

EBS Depot Repair Technician Portal • Clipboard-and-checkbox metaphor • Tabs for process steps: • Evaluation: diagnose and prescribe • Execution: tasks, parts, labor • Debrief: root cause, claims, activities • Decision Support: • Recommend Services • Frequency • Bulletins

Slide 10

Slide 10 text

EBS Depot Repair Technician Portal

Slide 11

Slide 11 text

EBS Depot Repair Technician Portal: ML Recommendations • Released July 2018: Patch #28263445 • Codeline 12.2.7+ RDBMS: 12.1.0.2.0 • Demo: https://youtu.be/gWTXvvERfag • Targets: Service Code, Defect Code (RCC) • Predictors: • Item • Item Revision • Age • Diagnostic Codes • Technician Notes (TM) • Service Agent Notes (TM) • Problem Summary (TM) • Customer Problem Description (TM) • Specific, Actionable Recommendations • Confidence Thresholds • Insight as a Requirement • Automatic Data Update • Log Decision and Reasons

Slide 12

Slide 12 text

EBS Depot Repair Technician Portal: Automatic Updates

Slide 13

Slide 13 text

EBS Depot Repair Technician Portal: Automatic Updates

Slide 14

Slide 14 text

Architecture Technician Portal: ML Recommendations • View Creation • Data Preparation • Encode categoricals • Scale numericals • Impute missing values • Manage sparse / imbalanced data* • Feature selection* • Manage outliers* • Generating Models (SVM, NB, DT) • Choosing the Best Model • Recommendation Thresholds • Oldest Transaction Date • Plumbing for Future Predictions

Slide 15

Slide 15 text

Roadmap and Wishlist • Measure and report recommendation outcomes. • Provide clearer insights. • Feedback loop based on rejection log and reasons. • Additional predictors: • Quality / Inspection data • Meters • Installed Location • Number of failures • Time between failures • Component age • Supplier data • Next-up ML solutions: • Predict remaining useful life of parts • Predict next failure • Predict cost, time to repair • Recommend repair vs. replace • Recommend additional actions • Detect warranty fraud • More competing algorithms, hyper-parameters. • Log training time and memory usage for models.

Slide 16

Slide 16 text

What we did Wrong 1. Confused data mining and app building. 2. Put too much faith in the initial data set. 3. Put too much faith in outside experts. 4. Chose a complex case first (multi-value targets). 5. Locked onto specific problems and solutions early.

Slide 17

Slide 17 text

Take-aways 1. Start by knowing the business and the user. 2. Structured data is still cool. 3. No perfect model or algorithm for all customers. 4. Test and iterate before locking onto a problem or solution. 5. Don’t put blind faith in experts.

Slide 18

Slide 18 text

Take-aways http://www.intelligent-maintenance.ch/index.html

Slide 19

Slide 19 text

Q & A 19

Slide 20

Slide 20 text

No content

Slide 21

Slide 21 text

Lee Sacco Senior Director Applications Development 2021-06-15 E-Business Suite Depot Repair Machine Learning for Maintenance and Repair