Ever since the Second Industrial Revolution, manufacturing firms have continuously been working on minimizing the inefficiencies and maximizing the productivity of their system. This objective led to the creation of the Toyota Production System which follows the motto of “making [the] highest quality products at the least cost in the shortest lead time. (Ohno, 1988)” This philosophy is widely recognized and is utilized by various industries today. Currently, we are going through the Fourth Industrial Revolution (also called, Industry 4.0) where internet technologies are utilized to additionally maximize the productivity in the production processes. Process synchronization is one of the inefficiencies in cellular manufacturing. King (1980) proposed a machine-part grouping approach called Rank Order Clustering (ROC). Some of the critical challenges to this approach were, there was no consideration given to machine process and performance data when grouping machine and parts; any change in initial matrix would alter the final solution. To overcome this challenge, an enhanced grouping approach called Modified Rank Order Clustering (MROC) was proposed in this dissertation (Amruthnath & Gupta, 2016). This approach was found to be reliable in providing consistent results irrespective of the arrangement of initial matrix and also, provided considerably higher balance between clusters. Unplanned downtime is another key inefficiency manufacturing industries still struggle with today. We can apply internet technologies (such as wireless sensors) to monitor the condition of critical machines remotely on the manufacturing floor based on physical attributes, such as vibration, temperature, current, pressure, force and voltage. This methodology is often called condition-based monitoring (CBM). The machine’s condition-based monitored data can be used along with machine learning tools such as supervised and unsupervised learning to observe the degradation of the overall machine and its subcomponents. It can also perform early detection of failures using anomaly detection models, diagnose the state of the machine using classification models, predict time to failure using regression models and identify the factors that influence the degradation using variable analysis models. Today, fault diagnosis in CBM research is focused on using supervised learning tools due to its high classification accuracy. The major drawbacks of this approach identified in this research using existing literature are (1) it’s time-consuming training phase where the data for all states of the machine and its components must be captured. If any new fault is detected, the model must be re-trained with the new state (2) its time-consuming implementation and its slow and unpredictable length of time for realizing benefits. Hence, most implementations have been just a proof of concept rather than a plant-wide implementation. (3) Finally, in dynamic environment such as manufacturing where machines operate under different process parameters, supervised learning models tend not to be as robust as unsupervised learning models. In this research, a generalized method has been proposed by using unsupervised learning for implementing different levels of predictive maintenance across the manufacturing floor. In this method the model is trained once using just the healthy/normal machine state and a model- based clustering approach to detect any new states of the machine. By using this methodology, we achieve faster implementation, implement a robust fault diagnosis model in a dynamic environment, identify all the states of machine faults, eliminate the process of retraining models and identify the most significant factors contributing to each state of the machine. The proposed approach was tested in an experimental study first that resulted in a classification test accuracy of 96.08%. Subsequently, the same approach was implemented in an industrial setting with data from three different cases. A classification test accuracy of 90.91%, 97.78%, and 94.4% was achieved respectively. A test hypothesis was used to test the significance of the results with a confidence level of 95%, and in all cases, the results were found to be statistically significant. The developed method could be extended to estimating time to failure using unsupervised learning, optimize maintenance scheduling and development of a portable module.