Smart manufacturing systems transmit out streaming data from IoT devices to cloud computing; however, this could bring about several disadvantages such as high latency, immobility, and high bandwidth usage, etc. As for streaming data generated in many IoT devices, to avoid a long path from the devices to cloud computing, Fog computing has drawn in manufacturing recently much attention. This may allow IoT devices to utilize the closer resource without heavily depending on cloud computing. In this research, we set up a three-blade fan as IoT device used in manufacturing system with an accelerometer installed and analyzed the sensor data through cyber-physical models based on machine learning and streaming data analytics at Fog computing. Most of the previous studies on the similar subject are of pre-processed data open to public on the Internet, not with real-world data. Thus, studies using real-world sensor data are rarely found. A symbolic approximation algorithm is a combination of the dictionary-based algorithm of symbolic approximation algorithms and term-frequency inverse document frequency algorithm to approximate the time-series signal of sensors. We closely followed the Bayesian approach to clarify the whole procedure in a logical order. In order to monitor a fan's state in real time, we employed five different cyber-physical models, among which the symbolic approximation algorithm resulted in about 98% accuracy at a 95% confidence level with correctly classifying the current state of the fan. Furthermore, we have run statistical rigor tests on both experimental data and the simulation results through executing the post-hoc analysis. By implementing micro-intelligence with a trained cyber-physical model unto an individual IoT device through Fog computing we may alienate significant amount of load on cloud computing; thus, with saving cost on managing cloud computing facility. We would expect that this framework to be utilized for various IoT devices of smart manufacturing systems.