For Path-Independent Insertion-Loss (PILOSS) optical switching networks, the traditional XY routing algorithm fails to ensure high-quality end-to-end communication, highlighting the need for a more efficient routing algorithm. This paper introduces a novel Convolutional Neural Networks (CNN) routing algorithm for PILOSS optical switching technologies to address common routing selection. The routing algorithm is redefined as a classification task, with the CNN monitoring multipath optical power and providing three classification references. The results show that the accuracy of the CNN training model is around 60.96%. Compared with the routing table generated by the XY algorithm, the proposed CNN algorithm reduces the first-order crosstalk by about 10 dBm.