2nd April 2026
© TECHiFAB
Predictive Maintenance Edge AI Memristor Computing

CLA-RiSTOR® Edge Analytics on the MAFAULDA Machinery Fault Dataset

This whitepaper shows how CLA‑RiSTOR converts high-rate acoustic and vibration signals into compact, stable features for on‑device fault classification. The method reduces data volume, power, and latency while keeping inference reliable under real industrial noise.

Preview

Industrial monitoring often relies on cloud‑based neural networks that need high compute, bandwidth, and power. CLA‑RiSTOR takes a different path. It processes high‑sample‑rate raw sensor streams through analog temporal integration, condensing the signal history into compact signal representations. Feature computation and inference are then performed digitally downstream. Using the MAFAULDA dataset, this paper demonstrates how the physics‑inspired pipeline delivers meaningful classification performance while staying compatible with strict edge‑deployment constraints.

Executive Summary

Predictive maintenance is often demonstrated with neural networks (NNs) operating on raw high-rate sensor streams. While these approaches can achieve strong benchmark accuracy, they are frequently difficult to deploy on edge devices due to compute-budget, power, latency, and data-transfer constraints.

This white paper presents a physics-inspired, analog signal-processing-based feature pipeline based on the CLA-RiSTOR® methodology and evaluates it on the MAFAULDA machinery fault dataset [1]. The approach compresses each 5-second measurement into a small set of engineered, physically interpretable features derived from event-triggered integration across multiple parameterized “views” of the same signal.

In our current MAFAULDA benchmark, our approach reaches macro-F1 ≈ 0.75, compared to macro-F1 ≈ 0.94 reported by a neural network as reference [2]. While the neural network remains stronger on pure benchmark accuracy, CLA-RiSTOR® is designed to win on what matters in production edge deployments: reduced data volume, deterministic inference, low energy consumption, and reliable on-device operation.

1. Why Edge Analytics Needs a Different Kind of Model

Industrial customers typically care about more than accuracy alone:

Low latency: detect issues in real time without cloud round-trips.
Low power: continuous monitoring must fit power and thermal budgets.
Low bandwidth: streaming raw high-rate signals is expensive and often impractical.
Noise robustness: industrial signals are often noisy; reliable monitoring must be stable under interference, mounting differences, and changing operating conditions.
Robust deployment: deterministic behavior, stable updates, and simple explainability.

Neural networks can be excellent when compute and data are abundant. On-device processing reduces the need to send raw sensor signals to the cloud for centralized processing, which helps lower bandwidth costs, latency, and operational complexity—especially in high-rate vibration monitoring.

Want to see the full methodology, results, and benchmarks?
Get the full whitepaper below.