Autoencoder Anomaly Detection: Per-Channel Reconstruction Error

Network

Bottleneck width 3

Anomaly Stream

Anomaly injection rate 8%

Corruption magnitude 1.6×

Live Detection

Reconstruction MSE

0.000

Adaptive threshold

0.000

Top-error channel

Status

NORMAL

Detected / injected anomalies

0 / 0

How it works

An autoencoder is trained only on normal data to compress each 8-channel sensor reading x through a narrow bottleneck and reconstruct it as x̂. The training loss is the mean squared reconstruction error:

L(x) = (1/n) Σᵢ (xᵢ − x̂ᵢ)²
Encoder: h = tanh(W₁x + b₁)   (n → bottleneck)
Decoder: x̂ = W₂h + b₂         (bottleneck → n)
Trained by gradient descent: minimize E[L(x)] over normal x only

Because the bottleneck is narrower than the 8 input channels, the network can only learn the correlation structure shared across channels (here, two hidden sinusoidal factors that drive all 8 sensors together) — it cannot memorize every channel independently. When a real anomaly hits, it typically breaks that correlation by moving one or a few channels independently of the rest. The reconstruction snaps back toward the learned normal manifold, so the corrupted channels show a large per-channel error (xᵢ − x̂ᵢ)² while the untouched channels stay near zero.

Bottleneck width — a wider bottleneck reconstructs everything well (including anomalies, hurting detection); a narrower one forces stronger compression and sharper anomaly contrast, but can under-fit normal data too.
Per-channel bars — the front (actual) and back (reconstructed) column pair for each sensor; a bar's color runs green→red with that channel's own squared error, so the reddest bar is the diagnosed root cause, not just a single flagged/not-flagged bit.
Adaptive threshold — mean + 3σ of the MSE measured on training-time (assumed normal) samples, recomputed as the stream evolves, the same rule production anomaly detectors use.
Training toggle — freezes the weights so you can compare a still-learning network against a converged one on the same anomaly.

Real-world relevance: this per-feature attribution is what turns a bare "anomaly score" into an actionable alert in manufacturing (which sensor drifted?), IT operations (which metric spiked?) and fraud systems (which field is inconsistent?).

Actual sensor value Reconstructed value High per-channel error

frame 0

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.