Figure 2.

Diagram of neural dynamic learning algorithm (NDLA). If training error $\epsilon (k)$ calculated by the class possibility matrix $\mathit{P}(k)$ and the label matrix L satisfies $\epsilon (k)\ge {\epsilon }^{\prime }$, where ${\epsilon }^{\prime }$ denotes a threshold, the weight matrix connecting between hidden and output layers $\mathit{W}(k)$ will be updated as $\mathit{W}(k+1)$ at the $(k+1)\text{th}$ training round under the situation of known NDLA design coefficient $\alpha$, mapping function $\Phi (·)$, hidden output matrix $\mathit{Q}$, the diagnosis deviation matrix $\mathit{E}$, predicted diagnosis output matrix $\mathit{Y}$, and the label matrix $\overline{\mathit{Y}}$.