Recently, the reliability of buildings and their energy efficiency highly depend upon emerging robust fault identification and diagnosis (FID) methods. However, conventional FID techniques achieve better performance only when there exists adequate training labeled data which is cost-effective and data collected in real-time conditions are mostly unlabeled. To overcome this issue, the developed method introduced an innovative correlation-weighted Elman neural network model (CW-ENN) for building automation systems (AS). Here, the data is used subsequently in the developed model, which learns the features and integrates with the labeled instances to perform the training process. Moreover, the correlation-weighted scheme emphasized the ENN model to balance both labeled and unlabeled instances effectively. Finally, various faults are classified and the developed method is simulated via the Python platform, and real-time fault samples are collected from the ASHRAE Research Project database. In the experimental section, accuracy and false positive rate (FPR) are scrutinized and compared with other techniques. The overall accuracy of 97.92% and FPR of 0.03 are obtained and proved robustness on FID over building AS.