Anomaly detection in monitoring systems is a critical process that identifies irregular patterns in environmental sensor data, which can signify underlying operational failures. At present, data-driven anomaly detection techniques with supervised machine learning and deep learning models are the most commonly employed, as they rely on fully labeled datasets for training. However, the process of labeling data can be labor-intensive and slow, creating a major obstacle to the broader adoption of machine learning methods. This study proposes to address this challenge by combines a conventional technique using modality distribution assessment statistical tools to create labels from modality temperature distribution for group of data points that were previously unlabeled and these generated labels were used in the next stage to train the supervised deep learning model with convolution neural network to automatically detect anomalies sensor failures. The result highlights that the temperature distribution characteristic classifying as unimodal or bimodal to detect anomaly sensor meets at 83.9% and 92% accuracy by selected statistical tool with CNN without optimization and with CNN-OPTUNA hyperparameters optimization, respectively. Remarkably, the deep learning-based methods should not be utilized solely, and the combination of statistical quantitative analysis, along with this deep learning analysis, offers considerable advantages.