Agents

polars_ts.iiot_agents.agents

Predictive-maintenance agents: spectral features, health, RUL, scheduling.

The agents form a pipeline: :class:SpectralFeatureAgent extracts vibration band energies, :class:HealthIndexAgent fuses multi-sensor signals into a health index, :class:RULEstimator projects Remaining Useful Life, and the reinforcement-learning :class:MaintenanceSchedulerAgent decides when to operate, inspect, or maintain.

SpectralFeatureAgent

Extract vibration band-energy features from a sensor window via FFT.

A dependency-free complement to the imaging module's to_spectrogram / to_scalogram: computes RMS amplitude and the fraction of spectral energy in low/mid/high frequency bands, which shift as bearings and gears degrade.

Parameters

n_bands Number of equal-width frequency bands to summarise energy over.

extract(window)

Return [rms, band_0_frac, …, band_{n_bands-1}_frac] for a 1D window.

HealthIndexAgent

Fuse multi-sensor readings into a health index in [0, 1].

Degradation is modelled as growth in each sensor's RMS amplitude relative to a healthy baseline; per-sensor degradation scores are fused by weighted mean and mapped to a health index (1 = healthy, 0 = failed).

Parameters

baseline Per-sensor healthy RMS baseline. Inferred from the first warmup steps when omitted. warmup Number of initial steps used to infer baseline when not supplied. fail_ratio RMS ratio (current / baseline) at which health reaches 0. weights Optional per-sensor fusion weights (defaults to equal weighting).

fit_baseline(sensors)

Infer the healthy per-sensor RMS baseline from initial observations.

score(window)

Return a fused health index in [0, 1] for a multi-sensor window.

window is (w, n_sensors) or a single (n_sensors,) row.

RULEstimator

Estimate Remaining Useful Life by extrapolating the health trend.

Fits a linear trend to the recent health-index history and projects the number of steps until it reaches failure_threshold.

Parameters

failure_threshold Health level defining failure. min_history Minimum number of points before a finite RUL is returned.

estimate(health_history)

Return estimated steps until failure (inf if not yet declining).

MaintenanceSchedulerAgent

Tabular Q-learning agent that schedules maintenance from health state.

The health index is discretised into n_states buckets; the agent learns a Q-value for each (health_bucket, action) pair from environment reward, balancing uptime against maintenance cost and failure risk.

Parameters

n_states Number of health buckets (finer = more granular timing). n_actions Size of the maintenance action set. alpha, gamma, epsilon Learning rate, discount factor, and epsilon-greedy exploration rate. seed Seed for the exploration RNG.

bucket(health)

Map a health index in [0, 1] to a discrete state bucket.

act(state, explore=False)

Return an action for state (epsilon-greedy when explore).

update(state, action, reward, next_state)

Apply a Q-learning temporal-difference update.