Why my training model(K-Fold) return Nan Value?

So I've trained a model using 10 K-fold in Spyder, at K-fold(iterations 1-3) training model. Everything was fine with at least 99% accuracy, 98-99% val_accuracy.

Then with four folds the accuracy and val_accuracy suddenly drop and stagnate at +- 70%. and when I check the score there's a nan value, which I don't understand.

So what I actually do is classify heart signal using PTB-Database for 8 class signal. and I'm using 1 second segmentation+DWT Sym8+Normalization(up=1, lb=0).

example photos using colab google.

This is an example at Specificity score/fold This is an example every segmentation and labeling that used

Can any one can drop a solution for this problem ? because I need at least 17-20 hours to train a model on my PC.

NB : I'm not good at speaking English if someone wants to correct my English it's okay