Dissertation — Time-series Analysis for Sensor Fusion

Thesis

Reliable sensor fusion in always-on systems is a governance problem over uncertainty through time, not merely a filtering problem. Systems that explicitly model temporal alignment, propagate uncertainty into policy, and formalize degradation modes outperform systems optimized only for average-case estimation accuracy.

Method

This study progressed through six units:
1. System framing and failure-aware foundations
2. Formal state-space and probabilistic tools
3. Implementation architectures for asynchronous sensors
4. Evaluation/observability under realistic faults
5. Optimization and scaling tradeoffs
6. Capstone synthesis into an edge-central operating model

Artifacts were produced per unit (notes, applied design output, and checks), then integrated into a deployment-oriented blueprint.

Evidence

1) Temporal correctness dominates raw model sophistication

Even strong filters fail under timestamp skew, irregular sampling, and unmodeled delays. Robust synchronization and bounded-lag interpolation reduced false innovation spikes and improved estimator stability.

2) Uncertainty must be a first-class interface

Treating covariance/confidence as internal metadata leads to brittle control. Routing policies by uncertainty thresholds produced safer behavior under dropout and drift, with predictable mode transitions.

3) Edge/cloud separation is required for practical reliability

Low-latency safety decisions must remain local; cloud should focus on diagnostics, retraining, and staged rollout. This split supports both responsiveness and continuous improvement.

4) Degradation design decides real-world performance

Production resilience depends on explicit fallback ladders (healthy → noisy → partial outage → fail-safe), not ad hoc exception handling.

Limitations

• Ground-truth instrumentation quality can cap achievable gains.
• Learned residual models may overfit seasonal/device-specific quirks.
• Rare compound failures (multi-sensor correlated faults) remain difficult to model comprehensively.

Implementation Roadmap

1. Enforce time integrity contracts at ingestion (clock health + drift alarms).
2. Standardize uncertainty schema across all downstream consumers.
3. Deploy policy routing with conservative/fail-safe bands and formal acceptance tests.
4. Introduce fault-injection CI (dropout, lag, bias, burst noise) as release blocker.
5. Run weekly drift review with shadow retraining and rollback-ready deployment.

Final Claim

The mature architecture for sensor fusion is a temporally disciplined, uncertainty-governed control system with explicit degradation guarantees. Accuracy matters, but trustworthy behavior under imperfect time and imperfect sensors is the real benchmark of intelligence in always-on environments.

Grade

93/100
Strong synthesis with operational realism, clear safety posture, and actionable deployment pattern.