We propose a dilated denoising autoencoder (DDAE) based on a multilayer one-dimensional convolutional neural network (CNN) to remove motion-induced noise—baseline wander (BW), muscle artifacts (MA), and electrode motion (EM)—from photoplethysmography (PPG) signals in the PPG-DaLiA dataset acquired using the Empatica E4 wrist-worn wearable device. The model integrates dilated convolutions, residual blocks, batch normalization, Leaky ReLU, max-pooling, and skip connections to capture long-range dependences while preserving pulse morphology. Compared with baseline methods such as the deep neural network (DNN), CNN, fully convolutional network (FCN), and convolutional denoising autoencoder (CDA), it offers three advantages: (1) pooling layers and skip connections preserve key features, (2) optimized parameter design reduces computational complexity and improves efficiency, and (3) a lightweight architecture enables efficient signal reconstruction on resource-constrained hardware. Evaluated across signal-to-noise ratio (SNR) conditions (−6 to 24 dB), the model significantly improves SNR, achieving 36.27 dB at −6 dB [root mean square error (RMSE): 0.008271; percentage root-mean-square difference (PRD): 13.97%] and 40.40 dB at 24 dB (RMSE: 0.005090; PRD: 9.00%. Deployed on a mobile device for real-time PPG processing, the model demonstrates superior performance and strong potential for medical and personal health applications.