Project Overview
This project tackled the challenge of transitioning urban mobility infrastructure from reactive to predictive maintenance by processing a massive dataset of 7.5 million telemetry logs. Operating in a high-noise industrial environment without dedicated IoT sensors, the study focused on identifying the precise transition points between normal operational states and system failures.
A core component of the work was a rigorous Root-Cause Analysis (RCA) designed to decode machine behavior. By isolating pervasive maintenance-related noise (such as system reboots) from genuine hardware anomalies, I significantly improved the reliability of failure detection and optimized the Signal-to-Noise Ratio (SNR) of the diagnostic profile.
Modeling Strategy and Iterative Refinement
The modeling approach followed an iterative strategy, transitioning from classical machine learning (Classification and RUL estimation) to complex time-series architectures and Deep Learning. This rigorous benchmarking allowed for a critical evaluation of the trade-offs between interpretability, performance, and robustness.
Given the high-noise nature of the telemetry data, the final architecture prioritized models that provided reliable diagnostic insights over pure predictive accuracy. This ensured that the resulting system remained grounded in physical reality, avoiding the pitfalls of overfitting to stochastic industrial artifacts.
Key Milestones and Strategic Impact
The analysis led to a pivotal discovery: failures manifested as sudden stochastic events rather than gradual degradations. This insight shifted the project's focus from standard time-series forecasting to advanced knowledge discovery and anomaly detection, providing a precise map of unpredictable system vulnerabilities.
I engineered an interactive Streamlit dashboard to serve as a diagnostic hub. This platform transforms raw telemetry into historical insights, allowing stakeholders to explore error patterns and perform real-time root-cause investigations. To ensure architectural scalability, I implemented a high-performance pipeline using DuckDB and vectorized temporal aggregations.
Beyond modeling, I delivered a comprehensive data governance roadmap. This included identifying data quality gaps and outlining standards for future sensor integration (vibration and temperature) to enable more granular, multi-modal predictive capabilities in future deployments.
For a deep dive into the experimental pipeline and the architectural workflow of this feasibility study, the GitHub repository offers full access to the analytical process.