Fog computing enables local processing for IoT sensor streams, minimizing delay and cloud dependency. Sensor-rich fog systems often encounter faults including thermal drift, mechanical vibration, and pressure surges. These faults reduce classification performance and affect decision accuracy in real-time embedded environments. This paper introduces a comparative fault-tolerant ML framework using CART, Random Forest, and Gradient Boosting. We evaluate three classifiers over environmental datasets featuring motion, acoustic, thermal, and barometric sensor signals. Each dataset includes six labeled fault classes: delay lag, thermal spike, motion glitch, static outlier, drift, and stuck fault. Data preprocessing includes noise filtering, time alignment, class balancing, and dynamic range normalization. Random Forest enhances decision boundaries using ensemble voting of uncorrelated decision estimators. Gradient Boosting incrementally improves classification by correcting previous model errors on minority fault types. CART, though simple, offers explainable classification suitable for microcontroller-based fog systems. All models are validated using 10-fold stratified cross-validation and class-specific evaluation metrics. We compare accuracy, latency, resource usage, and class-level confusion matrices under identical conditions. Results show Gradient Boosting achieves highest accuracy, while CART yields lowest energy and runtime. The framework supports flexible deployment, balancing performance with hardware constraints at the fog layer. This paper contributes to fault-resilient ML model selection for industrial fog environments and autonomous edge devices.