SHIELD

We are building SHIELD — a universal, on-device system that can understand what type of sensor is connected and predict when it will fail, using only the raw signal.

Project SHIELD is a universal, on-device sensor health intelligence framework that predicts sensor failures before they occur using only the raw sensor signal.
It combines physics-informed residual modeling with ultra-light TinyML forecasting to detect drift, noise rise, bias, latency growth, and other early signs of degradation—without requiring system models, redundancy, or cloud computing. SHIELD automatically identifies the sensor’s modality, extracts real-time residual fingerprints, and runs a compact IRNN/LSTM-style predictor directly on MCU-class hardware to estimate Remaining Useful Life (RUL) with calibrated uncertainty. Designed for aerospace, UAVs, industrial systems, and medical devices, SHIELD delivers proactive early-warning, failover triggers, and health diagnostics in a fully embeddable firmware pipeline, enabling safer, more reliable, and self-diagnosing sensors across mission-critical platforms.

Sensor Modality Detection

The system standardizes sampling (1 kHz, fixed window sizes) and extracts physics-informed fingerprints: RMS, variance, skewness, kurtosis, zero-crossing rate, PSD band ratios, spectral centroid, spectral flatness, and an ADEV surrogate for stability. An optional AR/Burg model fits low-order dynamics to capture resonant frequencies or AFE bandwidth. Two encoders run in parallel: (1) AutoModality-Lite (SVM/RF/GMM) for interpretable baseline inference, and (2) a physics-guided SSL 1D-CNN encoder trained with contrastive objectives and auxiliary physics heads. An abstention mechanism outputs “Unknown/Ambiguous” and triggers micro-excitation. The entire pipeline is MCU-optimized with INT8 quantization and CMSIS-DSP fixed-point features.

SHIELD-Prognosis: Sensor Degradation Detection & Early Fault Detection

Residual analysis captures early deviation from expected behavior, while tiny IRNN/LSTM/TCN models predict degradation trajectories and issue early warnings. The system detects faults seconds to hours ahead of failure, enabling proactive failover or maintenance. This addresses catastrophic cases such as IMU drift in UAVs or pressure-sensor failures in aerospace and medical devices.

The pipeline consists of:

Residual modeling: ARX/KF/UKF grey-box observers estimate expected sensor behavior; residuals surface drift, instability, or latency changes.
Temporal forecasting: Lightweight TinyRNN (IRNN), quantized LSTM, or micro-TCN consumes residuals and learned features to estimate degradation score or time-to-failure (TTF).
Decision layer: Thresholding, uncertainty scoring, and degradation-mode flags enable safe failover.
Embedded constraints: INT8 quantization, pruning, and fixed-point DSP ensure MCU compatibility (<64kB flash, <5kB RAM, <10ms latency).
Benchmark validation: NASA C-MAPSS and bearing datasets train teacher models; SHIELD HIL data validates embedded students.

Sensor Benchmark Dataset

SHIELD-Dataset standardizes sensor acquisition, stimuli, and metadata across many sensor modalities. It combines (a) public run-to-failure datasets for RUL learning, (b) controlled SHIELD DAQ recordings for modality detection, and (c) HIL-generated degradation trajectories for embedded forecasting. Together, they form the ground-truth backbone for reproducible evaluations and publications.

Publications

In progress

Reference

Add a Title

Reference

Ackhnowledgements

Text

Gallery