The prognostic advantages associated with AI-MSET represent significant performance advances and therefore warrant rigorous empirical validation. To that end, we propose a proof-of-concept (POC) evaluation in which performance is assessed using the customer’s own operational data. These POCs are relatively easy to conduct where there are existing data historian archives. This approach enables independent verification and direct, quantitative comparison against existing commercial tools or internally developed algorithms.

The proposed POC will apply AI-MSET to large-scale, high-frequency time-series datasets, supporting assets from a few dozen signals to fleets of assets with millions of signals with sampling rates ranging from sub-second resolution to kilohertz (KHz) and megahertz (MHz) regimes. The evaluation will include diverse sensor modalities, including vibration and acoustic transducers, and will span multiple years of historical operational data, where available.

The POC will quantitatively evaluate the following performance metrics:

• Earliest anomaly detection latency, defined as the minimum time between the statistically detectable onset of anomalous behavior and actual failures that have occurred in data acquisition systems and data historians

• False-alarm probability, evaluated using a non-threshold-based probabilistic framework, with target performance on the order of 1:10⁵ over 10,000 operating hours; probability tolerances are specified prior to model training and are statistically independent of missed-alarm probability

• Missed-alarm probability, evaluated using the same non-threshold-based probabilistic framework, with target performance on the order of 1:10⁵ over 10,000 operating hours; probability tolerances are specified prior to model training and are statistically independent of false-alarm probability

• Computational resource requirements, demonstrating reductions of two to three orders of magnitude relative to conventional multivariate machine learning, such as neural network–based approaches

• Human resource requirements, demonstrating that model deployment, model scaling, and retraining can be performed by non-specialists with minimal effort

The results of this POC will allow stakeholders to independently assess performance, scalability, and operational practicality against all in-house approaches and/or commercially available systems such as LSTM or other neural network systems.