The massive growth of genomic data has made it necessary to have effective computational algorithms that are able to detect small, biologically significant sequence patterns which are called motifs. The Planted Motif Problem (PMP) that consists of locating (l, d) motif in a number of sequences with a maximum of d mutations is still computationally challenging because of its combinatorial complexity. PMM (Planted Motif Miner) is a hybrid approximate framework that we suggest to identify (l, d) planted motifs efficiently using pattern mining and deep learning. In this work, PMM has seven steps including data preprocessing, k-mer segmentation, frequent-pattern aggregation, motif extension, consensus motif mining, d-neighbor generation, and CNN-based classification module. The model uses multiprocessing and random sampling to deal with the exponential growth of d-neighbors with large (l, d). Experimental results on simulated and real data show that PMM are accurate ranging from (87% to 99%) based on length of (l, d) parameters and better than state of the art motif discovery algorithms, such as qPMS10, FMotif, GADEM, and MEME-ChIP. Its performance advantage is obvious on large instances where the current ways will normally fail or take over 24 hours to accomplish. In addition, the suggested algorithm can execute large (l, d) like (30, 10). The findings indicate that the suggested method is able to obtain > 99% accuracy and F1-scores above 97% with a minimum loss rate, when using moderate parameter configurations, such as (18, 6) and (26, 6). The sampling is necessary when the problem complexity is higher to make sure that it is feasible, the classification performance tends to decrease, but the CNN can still achieve strong accuracy and F1-scores over 82% in the most difficult settings.