Sarcasm is employed widely on various social media platforms. Due to the potential for sarcasm to alter the intended meaning of a statement, the opinion analysis technique is susceptible to inaccuracies. Detecting sarcasm is one of the most challenging problems in analyzing sentiment and mining opinions in social media. Therefore, identifying sarcasm is crucial when making informed public opinion decisions. Preliminary research indicates that sarcastic statements alone have a substantial negative impact on the accuracy of automatic sentiment analysis. Several distinct natural language processing strategies have been previously suggested. However, each technique has limits in terms of textual context and proximity, and the accuracy of classifiers is affected by noise in the dataset. This research introduces SarcAE, a unique method for combining feature-level embedding fusion using an autoencoder and fuzzy logic-based reasoning to classify sarcasm. The evaluation experiments used two benchmark datasets: the News Headlines dataset and Ironic Tweet dataset, subjected to several preprocessing techniques. Extensive experiments conducted using the proposed SarcAE approach demonstrate that the proposed method outperforms other fusion models with an accuracy of 98.53% on the News Headlines dataset and 89.83% on the Ironic Tweet dataset, respectively, surpassing baseline methods by up to 3.7%. These results indicate the effectiveness of SarcAE in capturing contextual and semantic cues needed for sarcasm detection.