Mesh Failure Prediction
A deep-learning model that reads a mesh node’s recent telemetry and predicts whether it is healthy, at risk or failed, early enough to reroute around it.
Overview
In a disaster-response mesh, one weakening node can become a bottleneck and then break a whole path. This project asks whether telemetry alone gives enough warning to reroute before that happens.
Each node reports six signals: signal strength, packet loss, latency, hop count, battery level and uptime. The model looks at sliding windows of ten readings, so it learns how a node degrades over time rather than judging single measurements.
Worth noting
Three models, one dataset
A feedforward baseline, an LSTM and a 1D CNN were compared on the same telemetry. The LSTM was kept because it captures gradual degradation best.
A small network
One LSTM layer of 64 units, dropout, a dense layer of 32, and a softmax over the three states.
Evaluation without leakage
A 70/15/15 split, a scaler fitted on training data only, and early stopping on validation loss. The script reports accuracy, macro F1 and a confusion matrix.
Results
About 97.5% test accuracy with a macro F1 of roughly 0.975.
Limits
The telemetry is synthetic, generated to mimic realistic degradation. The same pipeline is meant to take real ESP32 telemetry next, and the numbers will need measuring again when it does.