Advanced and emerging methods

Explainable AI: why did it say that?

A technician will not open a gearbox because a black box said 0.83. Explainable AI (XAI) methods attribute each prediction to its inputs. SHAP values, based on Shapley values from game theory, are the most widely used: each feature gets a signed contribution that pushes the prediction up or down from the average.

Other useful tools: permutation feature importance for global behaviour, Grad-CAM and saliency maps for CNNs on spectrograms, and counterfactuals ("the alert would clear if kurtosis were below 4.1"). Explanations also catch bad models: if the top feature is the sensor ID or the date, the model has learned a shortcut.

Physics-informed machine learning

Pure data-driven models ignore what engineers already know. Physics-informed approaches put that knowledge back in:

• Physics-based features: bearing defect frequencies, gear mesh orders and slip-dependent sidebands as inputs.
• Hybrid models: a physical model predicts the expected behaviour and ML learns the residual it cannot explain.
• Physics-informed neural networks (PINNs): the loss function penalises predictions that violate known equations, such as heat balance or a wear law.
• Constraints: health indicators forced to be monotonic, RUL forced to decrease with time.

The payoff is better behaviour outside the training data, which is where plant models spend much of their life.

Digital twins

A digital twin is a model of a specific asset kept in step with it by live data. For maintenance, the useful twin is usually modest: a calibrated model of the machine's normal behaviour plus its degradation state, updated as measurements arrive. It answers "what if" questions: what happens to remaining life if we raise throughput by 10% until the shutdown? Twins combine the physics-informed and Bayesian ideas above with the data pipeline from page 2.

Edge AI and TinyML

Models increasingly run on the sensor node or gateway. A quantised autoencoder or small CNN fits in a microcontroller's memory and scores every window locally, sending only scores, features and alerts upstream. Benefits: lower bandwidth, alerts that work without the network, and raw data that never leaves the site. Costs: limited model size, harder updates, and the need to monitor many distributed models.

MLOps and drift

A deployed model degrades silently as conditions change: a new product, a replaced sensor, a seasonal temperature swing. MLOps is the practice of versioning data and models, monitoring them in production and retraining on a schedule or a trigger. The first thing to monitor is input drift: whether today's features still look like the training data.

Beyond input drift, track the outcomes that matter: alert precision from technician feedback, missed failures, and the time from alert to work order. A model is only as good as its last quarter's results.

Federated learning

Federated learning trains one model across many sites without moving their data. Each site trains locally and shares only model updates, which are combined centrally. It suits equipment makers and multi-site operators who want fleet-level models but face data-sharing restrictions. The difficulty is that sites differ (domain shift again), so personalisation layers per site are common.

LLM copilots and agents in maintenance

Large language models (LLMs) do not replace signal models, but they are changing the work around them:

• Work-order mining: turning decades of free-text maintenance notes into structured failure codes, which become training labels.
• Alert explanation: drafting a plain-language summary of an alert from its features, SHAP values and the asset's history.
• Procedure assistants: answering technician questions from manuals and procedures, with citations.
• Building tools: AI coding agents let engineers build analyzers and dashboards themselves. That is the subject of the Vibe Coding section.

Quick quiz