Tire pressure plays a crucial role in driving performance, fuel efficiency, and safety. Low tire pressure increases the risk of accidents by extending braking distances, leading to the mandatory use of tire pressure monitoring systems (TPMS) in many countries. These systems monitor wheel speed or pressure using sensors, but the signals they produce are often contaminated with noise, including sensor and environmental noise. This study focuses on denoising vibration signals from nitrogen-filled pneumatic tires, which are known to maintain more stable pressure levels. Vibrations recorded by an accelerometer on the vehicle’s wheel hub were used to create vibration plots. To improve signal clarity, a non-local fully convolutional neural network (NL-FCNN) with a wide kernel principle was proposed. This network incorporates a non-local block (NLB), inspired by the non-local means (NLM) technique, enabling it to capture long-range dependencies in the signal. Several pre-trained networks, including AlexNet, VGG-16, DenseNet121, MobileNet, GoogleNet (InceptionV3), and ResNet50, were evaluated for performance, with ResNet50 achieving the best results. The model attained a classification accuracy of 94.05%, precision of 88.36%, and recall of 80.46%. This research presents a robust approach to improving the accuracy of fault diagnosis in TPMS through advanced denoising techniques.