Instrument Health Monitoring
We developed the theoretical basis for runtime metrics logging, filament and emission behavior tracking, anomaly-driven fault classification, health score models, and remaining useful life estimation with confidence.
Atonarp is a Japan-based company working at the intersection of molecular insights and digital technology, with a strong focus on semiconductor and industrial process control. Its ASTON platform is positioned as an in-situ, real-time, high-sensitivity mass spectrometry solution for semiconductor manufacturing, designed to deliver molecular-level process visibility that can improve throughput, yield, and operational efficiency. This page focuses on the Aston-related work around health monitoring, endpoint intelligence, drift awareness, interlock logic, and bounded autonomous process control.
The Aston work spans a multi-layer technical program rather than a single model. It covers runtime metrics, anomaly detection, signal intelligence, safety-aware actions, and AI feedback loops, all grounded in the operational realities of a real-time instrument.
We developed the theoretical basis for runtime metrics logging, filament and emission behavior tracking, anomaly-driven fault classification, health score models, and remaining useful life estimation with confidence.
We mapped out endpoint detection and spectral intelligence through threshold logic, plateau analysis, Dynamic Time Warping, half-etch timing, spectral pattern analysis, and fingerprint-style comparison against known-good behavior.
We covered training-to-inference architecture, confidence-aware recommendations, guardrail evaluation, bounded control proposals, run-to-run control logic, and closed-loop feedback for continual model improvement.
The documentation points to a coordinated architecture where analytics, supervision, and control all work together. These are the workstreams that stand out most clearly from the Aston material.
Work included runtime metrics logging, PWM trend analysis, emission noise quantification, SHA-256 data integrity checks, anomaly-based fault classification, alert thresholds, maintenance triggers, and remaining useful life estimation.
Work included threshold crossing, derivative logic, plateau detection, Dynamic Time Warping, half-etch endpoint detection, spectral pattern analysis, cosine and correlation-based matching, and fingerprinting for live process recognition.
Work included SPC, z-score normalization, EWMA, CUSUM, trend regression, drift direction classification, severity scoring, fault taxonomies, and interlock trigger integration so abnormal process behavior could be recognized and acted on safely.
Signal-level analytics handled health monitoring, endpoint recognition, fault detection, drift detection, and process-state interpretation from live mass spectrometry data.
Decision logic connected anomaly scores, severity levels, drift states, and confidence outputs to recommendations, interlocks, and action-blocking rules before unsafe behavior could propagate.
Run-to-run control proposals, guardrails, bounds enforcement, trigger recommendations, and outcome-based feedback formed the basis for closed-loop improvement rather than one-shot inference.
The Aston package points to concrete statistical and machine-learning methods such as Isolation Forest, autoencoders, Mahalanobis distance, EWMA, CUSUM, DTW, and rule-based severity logic rather than vague AI claims.
The work includes object-dictionary mappings, latency targets, integrity verification, model metadata, trigger schemas, reason codes, and explicit runtime action pathways that fit an operating device.
The architecture captures outcomes, failure states, low-confidence cases, and control results so retraining priority can be based on real operational value and not only offline benchmark metrics.
We can turn that into a scoped conversation around deployment architecture, AI supervision, and the fastest path to something credible in production.