An essential component of buildings that provides occupant-centered and energy-efficient interior amenities is the heating, ventilation, and air conditioning (HVAC) system. Since inaccurate sensory readings have the potential to disrupt system functioning, sensor fault diagnosis, is crucial for HVAC systems. The majority of sensor fault diagnosis approaches for HVAC systems require either extensive data with labels or extensive manual expertise, both of which are scarce in various buildings. The deep learning approach is generally suitable to diagnose HVAC sensor problems in order to save labor costs, but it still faces substantial obstacles such as uneven distribution of information and negligible fault characteristics. This research proposes a novel hybrid deep learning technique to improve the performance sensor fault diagnosis in HVAC systems. Initially, Multi-linear Subspace Learning Principal Component Analysis (MSL-PCA) is utilized as a preprocessing technique in the proposed research. Secondly, in order to replace missing sensor readings, the data imputation module utilizes the generative adversarial network (GAN) model. Variational autoencoder is employed as the fault detection module with Deep Convolutional Neural Networks implemented as the Fault classification module. Compared to traditional preprocessing methods, Multi-linear Subspace Learning Principal Component Analysis (MSL-PCA) improves feature extraction and raises the caliber of input data for deep learning models. By combining dimensionality reduction, data imputation, and deep learning approaches for dependable performance, the hybrid architecture addresses the uneven distribution of data, a prevalent problem in HVAC systems. The integration of the methods used in the proposed system is used to overcome the drawbacks of conventional techniques, such as sparse information and insignificant fault features, and bridges the gap between robust fault detection and data reliability. Experimental validation is carried out using ASHRAE project 1312-RP and the performance of the proposed system is compared with conventional machine learning and deep learning algorithms. The proposed system outperforms other algorithms by producing an accuracy of 98.72% for effective sensor fault classification in HVAC systems.