The detection of gravitational waves (GWs) has opened up new avenues for studying the universe and testing fundamental physics. As LIGO, Virgo and KAGRA start another observation run (O4) this year with an improved sensitivity, non-axisymmetric neutron stars emitting quasi-monochromatic, long-standing GWs are expected to be within the detectors' frequency bands. However, their detection in the presence of noise is a challenging problem. In recent years, Convolutional Neural Networks (CNNs) have been proposed as a potential solution to this issue. This study explores the effectiveness of CNNs for detecting CWs embedded in simulated noise as well as in real noise obtained from the O3 observation run. The model is trained on a set of $10^4$ templates of Continuous Waves (CWs) signals immersed in Gaussian noise with time gaps. Moreover, we evaluated the CNN's capability to generalize in signal-to-noise-ratio (SNR) and signal frequency . The results show that the CNN model is successful in detecting signals in simulated noise for the frequency band ranging from 100 to 1000 Hz, achieving high detection accuracy and low false positive rates. Nevertheless, when evaluating the model in data with real noise, which contains non-stationary noise and instrumental artifacts, its performance deteriorates. Such limitations suggest that more advanced methods are needed to analyze gravitational wave data effectively.