Algorithm 1: MTPL Model Training and Prototype Learning
	Require: Training dataset ${\mathcal{D}}_{\mathrm{train}}={\left\{\left(x_i,y_i\right)\right\}}_{i=1}^N$; Number of known classes K; Loss weights λp, λr; Batch size B; Number of training epochs, Ep. Ensure: Optimized model parameters $(\varphi ,\theta ,\psi )$; Prototype set $C={\left\{c_k\right\}}_{k=1}^K$.
1	Initialize parameters for the encoder $E_{\varphi }$$,\mathrm{decoder}{D}_{\theta }$, and classifier $F_{\psi }$.
2	Initialize the learnable prototype matrix $C\in {ℝ}^{K\times D_{\mathrm{feat}}}$ using Gaussian random initialization.
3	For epoch = 1 to Ep do:
4	for mini-batch $\left\{\left(x_b,y_b\right)\right\}$ from ${\mathcal{D}}_{\mathrm{train}}$ do:
5	//Forward propagation
6	Latent features: $z_b\leftarrow E_{\varphi }\left(x_b\right)$;
7	Reconstructed signal: ${\widehat{x}}_b\leftarrow D_{\theta }\left(z_b\right)$;
8	Classification logits: ${\mathrm{logits}}_b\leftarrow F_{\psi }\left(z_b\right)$;
9	Classification probabilities: $p_b\leftarrow \mathrm{Softmax}({\mathrm{logits}}_b)$.
10	//Multi-Task Loss Calculation
11	Classification loss: $L_c\leftarrow CrossEntropyLoss\left(p_b,y_b\right)$;
12	Prototype loss: $L_p\leftarrow {\displaystyle \frac{1}{B}}\sum _{i=1}^B{‖\begin{array}{c}{z}_{b,i}-c_{y_{b,i}}\end{array}‖}_2^2$;
13	Reconstruction loss: $L_r\leftarrow {\displaystyle \frac{1}{B}}\sum _{i=1}^B{‖\begin{array}{c}{x}_{b,i}-{\widehat{x}}_{b,i}\end{array}‖}_2^2$;
14	Total loss: $L_{total}\leftarrow L_c+{\lambda }_p{L}_p+{\lambda }_r{L}_r$.
15	//Backward Propagation and Parameter Update
16	Compute gradients of $L_{total}$ with respect to all learnable parameters $\left(\varphi ,\theta ,\psi ,C\right)$;
17	Update parameters $\left(\varphi ,\theta ,\psi ,C\right)$ using an optimizer.
18	end for
19	End for
20	Return Trained parameters $(\varphi ,\theta ,\psi )$ and prototype set $C$