Wireless sensor networks (WSNs) can make errors due to their deployment in hazardous and unpredictable environments. Identifying and diagnosing outliers/faults in WSNs is challenging because of the varied deployments and limited sensor resources. To address this, unsupervised machine learning-based one-class classifier methods are used to analyze sensor data and detect outliers/faults at the sensor node level. Data on humidity and temperature were collected using the DHT22 sensor, and a window sliding protocol was employed for data preprocessing in our experiments. The research results indicate that our proposed approach can effectively detect faults/outliers that are hard to distinguish. Additionally, it benefits from requiring only a big sample size for training, leading to high accuracy and stability in fault/outlier diagnosis in seen and unseen data. The simulation results indicate that the isolation forest method outperformed other state-of-the-art unsupervised one-class classification models with an achieved testing accuracy of 95.44% for unseen faults.