Maintenance Planning WorkflowTrend-Based PredictionScheduling Support

Preventive Maintenance Workflow

A predictive workflow designed around maintenance decisions.

AI Engineer / Project Lead2022 - 2024AUO Corporation

Preventive Maintenance · Forecasting Flow

Trend-Aware LSTMCustom Loss FunctionSchedule-Oriented PM
  1. 01

    Sensor Stream

    Equipment Time-Series

  2. 02

    Sequence Prep

    Imbalanced Pattern Prep

  3. Custom Loss03

    Trend Forecast

    PyTorch LSTM + Custom Loss

  4. 04

    Risk Scoring

    Actionable Alert Signals

  5. 05

    PM Schedule

    Downtime & Timing Tradeoff

  6. 06

    Preventive Act

    −20% RMSE · ~90 hrs/mo

Replaces fixed-cycle maintenance with trend-aligned LSTM forecasting tied directly to scheduling tradeoffs.

RMSE −20% · ~90 hrs/mo Saved
Workflow overview. Some names and data are anonymized or recreated for confidentiality.

Before

Traditional maintenance scheduling followed fixed cycles and often reacted too slowly to emerging equipment issues.

After

A prediction workflow emphasized anomaly trends and fed scheduling decisions with more usable maintenance signals.

Context & Background

  • Operations teams needed better timing signals for maintenance planning.
  • The goal was to support maintenance decisions, not just to improve a benchmark score.

Constraints

  • Failure patterns were imbalanced and hard to model.
  • Outputs needed to map back to scheduling tradeoffs and operational timing.

Problem Solved

  • Fixed-cycle maintenance was not responsive enough to evolving equipment conditions.
  • Raw prediction accuracy did not automatically translate into useful maintenance action.

My Role

  • Model design was aligned with the actual maintenance decision problem.
  • The predictive approach and output design were both shaped around scheduling workflows.

System Design

  • Used time-series prediction with a custom objective shaped around trend consistency.
  • Integrated results into maintenance planning logic so outputs were more actionable.

Implementation Highlights

  • Designed a custom loss function around trend behavior rather than point-by-point fit.
  • Focused the workflow on planning support and timing decisions.
  • Connected predictive outputs to scheduling use cases.

Outcome & Impact

  • Improved defect prediction error and reduced wasted maintenance effort.
  • Made predictive output more useful for planning rather than just monitoring.

Tech Stack

PythonPyTorchLSTMCustom Loss FunctionFastAPI