Predictive Maintenance — Nonlinear Fault Detection

An MLP (Tanh-activation) autoencoder reconstructs the healthy operating manifold; reconstruction error above a learned threshold flags an anomaly SPC's linear PCA-style residual would miss on nonlinear process relationships.

mlautoencoderanomaly-detection

Concept

Upgrades the linear PCA detector to a nonlinear autoencoder: an MLP with a tanh bottleneck learns to reconstruct the healthy operating envelope; reconstruction error above a learned percentile threshold is the anomaly score. Because the bottleneck is nonlinear it flags a fault off a curved manifold that a linear PCA subspace can't separate.

The math

Show the governing equations

d \xrightarrow{\tanh} h \xrightarrow{\tanh} \text{latent} \xrightarrow{\tanh} h \xrightarrow{\tanh} d, \quad \mathcal{L} = \text{MSE}(x, \hat{x})

Forward pass, backprop, and the Adam optimizer are hand-written in numpy — no deep-learning dependency. Xavier/Glorot init keeps tanh training stable; a fixed seed makes it deterministic.

Execution

Same interface as the PCA detector (ml.anomaly) — an endpoint can pick either method without changing its calling code.