🔧 Rilevatore di Anomalie dei Sensori — Autoencoder dal Vivo

An autoencoder is a neural network trained to reconstruct its own input through a narrow bottleneck layer that forces it to learn a compressed representation of the data. When trained only on normal sensor readings, the network becomes very good at compressing and reconstructing the patterns it has seen, but poor at reconstructing patterns it has never encountered. Anomaly detection exploits this: feed the trained network any new reading and measure the reconstruction error, the mean squared difference between the input and the network's output. Normal readings reconstruct with low error; readings that fall outside the learned manifold reconstruct with high error, because the compressed bottleneck simply has no efficient code for them.

If the hidden layer were as wide as the input, the network could learn a trivial identity mapping that copies every value straight through, and every input, normal or anomalous, would reconstruct with near zero error, making anomaly detection meaningless. This simulation's encoder compresses five sensor channels down to a 3-unit bottleneck before a mirrored decoder expands them back to five outputs. That narrow bottleneck forces the network to discover the low-dimensional structure that actually governs a healthy machine, its shared load cycle, and discard anything outside that structure.

Every animation frame, the simulation draws fresh synthetic normal readings, never real data files, runs them through the current network in a forward pass, computes the mean-squared reconstruction error against that same input as the training target, and then runs a full backward pass: the chain rule is applied layer by layer to compute gradients for every weight and bias, using tanh derivatives for the hidden layers and a plain linear derivative for the output layer, then every parameter is nudged by minus the learning rate times its gradient. This is genuine backpropagation and gradient descent, implemented from scratch in plain JavaScript with no machine-learning library, and you can watch the loss curve fall as training proceeds.

Five failure modes are available: a Sensor Spike where one channel jumps abruptly by several normal standard deviations, a Sensor Drift where one channel slowly biases away from its healthy range, a Stuck Sensor where one channel freezes at a fixed value while the real load keeps changing, a Correlation Break where all sensors keep moving but stop moving together even though real machines have their sensors coupled through a shared physical load, and a Noise Burst where all channels become far noisier than anything seen in training. Because the network was never trained on any of these five patterns, every one of them typically produces a reconstruction error tens to hundreds of times higher than normal operation.

The threshold slider sets the reconstruction-error value above which the simulation flags ANOMALY DETECTED. Set it too low and ordinary noisy-but-healthy readings will trigger false alarms; set it too high and real faults will slip through undetected. A common practical approach, mirrored here, is to set the threshold a few standard deviations above the mean reconstruction error observed on validated normal data, then adjust it while watching the live error chart to balance sensitivity against false-alarm rate for the specific machine being monitored.

This is the defining trick of autoencoder-based anomaly detection: because true anomalies are rare, varied and often unseen in advance, it is usually impractical to collect a labelled dataset of every possible failure mode. Instead the network only ever needs examples of what normal looks like, which are abundant, and it learns anomaly detection as an unsupervised by-product of compressing that constantly generated normal data well. This simulation strictly separates its training generator, which only ever produces the healthy correlated sensor pattern, from its live anomaly-injection code path, so any high reconstruction error you see is a genuine, unrehearsed generalisation failure, not a memorised example.

Yes. The five glowing sensor nodes correspond to the network's five input channels, temperature, vibration, pressure, current and RPM, and each node's colour and glow intensity is driven directly by that channel's squared reconstruction error at the current instant, the actual reading minus the network's reconstruction, squared. The central housing's colour reflects the combined score against the current threshold. Nothing is scripted or pre-baked: every value comes from the live forward pass of the same network shown training in the loss chart.