What is an autoencoder and how does it detect anomalies?

An autoencoder is a neural network trained to reconstruct its own input through a narrow bottleneck layer. Trained only on normal sensor readings, it reconstructs normal patterns with low error and unfamiliar patterns with high error, which is used as the anomaly score.

How is this network trained in the browser?

Every frame the simulation generates fresh synthetic normal sensor data, runs a forward pass, computes mean-squared reconstruction error, and runs a genuine backward pass using the chain rule to update every weight and bias via gradient descent — no ML library involved.

Why does the network only train on normal data?

Because real anomalies are rare and varied, autoencoder anomaly detection only needs abundant normal examples; anomaly detection then falls out as an unsupervised by-product of learning to compress normal operation well.

Home▸AI & Machine Learning▸Sensor Anomaly Detector — Autoencoder Live

🔧 Sensor Anomaly Detector — Autoencoder Live

Watch a real neural autoencoder train live via backpropagation to reconstruct normal simulated sensor readings, flagging anomalies by genuine reconstruction-error spikes it was never trained to compress well.

AI & Machine Learning3DAdvanced60 FPS

⚙ Under the hood

Three.jsWebGLAIMachine Learning

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install