Modelling

Once the historical data is gathered, a system modelling is carried out. There is a large variety of modelling methods available to encode the historical data in a compact form, which then can be used for signal monitoring. One grouping of methods divides models into: (i) physical; (ii) and data-driven. The physical models rely on equations describing the physics in order to represent historical data. The data-driven models use historical data and general-purpose equations to encode the system behaviour. 

For the data-driven models, the left-to-right order can be interpreted in several ways:

• from a clear yes/no answer to vague probabilistic distributions;
• from zero predictive power to a certain predictive potential;
• from fully deterministic algorithms and clear answer, to deterministic algorithms and probabilistic answer, to non-deterministic algorithms and probabilistic answer;
• from low, to moderate, to high computational cost in order to develop and train the model;
• from full trust to result, to high-degree of trust, to low-level of trust;

Monitoring

Automatic execution of monitoring application depends on the operation state:

• triggered by Post Mortem events, e.g., Fast Power Abort, quench heater discharge
• triggered by change in the LHC beam mode
• in regular intervals, e.g., every hour

The monitoring applications perform the following types of comparisons in order to detect anomalies:

1. With historical data we develop digital-twin models and derive trends
2. With on-line data we compare behaviour with redundant copies of a system (intra-component) and across circuits of similar topology (cross-population)