Base Neural Network
What you'll see here
The evaluation criterion, and how a single-hidden-layer SGD network performs as the reference point for every improvement that follows.
Model evaluation criterion — recall on the failure class
- A false negative is a missed failure and becomes a full generator replacement — the costliest outcome by far.
- A false positive only triggers an inspection, the cheapest outcome.
- Accuracy alone is misleading: predicting “no failure” every time already scores 94.5%.
- F1 is tracked alongside recall so the model does not buy recall with an unusable false-alarm rate.
Model 1 — baseline network with SGD
32-unit hidden layer · ReLU · sigmoid output · SGD (lr 0.01) · binary cross-entropy · 30 epochs · batch 128 · seed 42
Validation accuracy
97.6%
looks strong — but misleading
Validation recall
58.56%
failures caught
Validation precision
97.01%
alarm reliability
Validation F1
73.03%
balance
Reading the baseline
- Accuracy looks strong at 97.6%, but recall is only 58.56% — the network misses over 40% of real failures.
- Training recall (59.12%) and validation recall (58.56%) are almost identical, so this is underfitting, not overfitting.
- SGD with a single narrow hidden layer converges slowly and the imbalanced loss pushes the network toward the majority class.
- This sets the improvement agenda: more capacity, a faster optimizer, regularisation, and explicit imbalance handling.
AI: Interpreting the baseline
Click Generate to produce an evidence-based observation, insight, recommendation, and business benefit for this section.