Now that the hype surrounding Data Science has slightly diminished, we can affirm that this is not a drill but rather an exhilarating reality. Governments, large organizations, and start-ups alike have already seen and understood the value this discipline brings to the table and are fervently competing for talent and dominance in this space. As of 2019, we are finding ourselves at a new precipice in entering the Fourth Industrial Revolution: The Real-Time Enterprise! Just like the intersections of the past: First in 1784, Water and steam; Second in 1870, First conveyor belt; and the Third in 1969, Electronics and information technology, this pause will bring with it both challenges and new opportunities. Data stewards in the discipline have realized associative costs for data storage has expounded this issue while driving talent gaps. As their respective organizations continue to ingest voluminous amounts of data, they must become more tactical with the data they are creating and using to sustain or improve Return on Investment (ROI). More so, the approach of advancing insight through analyses of mountains of data must clear and consistent so that results are both reproducible and can be made autonomous. One method in doing this is by utilizing the Cross-Industry Standard Process for Data Mining (CRISP-DM). This logical method enables data stewards and stakeholders to clearly understand what, when, where, why, and how they are mining data through six easy to follow phases:
- Business understanding
- Data understanding
- Data prep
- Modeling & Application Development
- Evaluation
- Model Deployment.
Contraints, limitations, and assumptions
Constraint 1. The success of the produced mathematical model will depend on how precise the prediction of material failure is. The training set comprises historical telemetry observations from only one machine in the organization's operational plants. For instance, after analyzing several readings (independent variables) over months of consistent machine utilization hours, the model's predictions will be based on future and constant occurrences similar to the recorded and analyzed observations in the past.
Constraint 2. We do not have sufficient data to rule out time frame or seasonality significance for sensor readings in seeing how sensors perform over time and for the same period.
Limitation 1. To predict the likelihood of a future material failure for a machine after N period of running hours, we would need to know the expected hours the device will run in the future: which we do not know "accurately" today.
Limitation 2. It was also noted that in future data pulls it may be helpful to include demographic information to see how the maintainers across regions vary in skill set and maintenance practices.
Assumption 1. The classification model for this run will address threshold levels concerning mixed readings from various sensors to establish a baseline of telemetry readings pointing to machine deterioration.
Assumption 2. The classification model for this run will be able to incorporate new data from other machines in the organization and adjust accordingly, unearthing even more hidden dependencies not easily seen with five months of data.
Assumption 3. Based on the limited dataset, this analysis will yield more questions. It may require more data for future analysis keeping in mind not deteriorating the return on investment (ROI) for the project, and bearing those negative/positive findings are both outputs for this project.
The following project will demonstrate how to utilize CRISP-DM from a practical standpoint; the next analog will use it on simulated manufacturing data predicting maintenance failures for a theoretical client's manufacturing operation. Predictive maintenance is an area that has a clear use-case for data-mining and primarily due to the breakthroughs of applied machine intelligence. With the continuous advancements in the Real-time enterprise fueled through the: Internet of Things (IoT), Low-Cost Telemetry Sensors/RFID tagging, Low-Cost Digital Storage, and increasing Computing Power amongst others, the capabilities of transforming voluminous data into insight in this space does not appear to be slowing. The growth in Artificial Intelligence (AI) amongst increasing levels of Automation seen in manufacturing allows firms to be more resilient in connecting fixed-assets while improving productivity through data-driven decisions and insights not possible before. As the use of automation continues to augment and takeover manufacturing, the reduced response time required in dealing with maintenance issues will outpace the speed at which humans can intervene, requiring sophisticated and automated optimization decisions, especially concerning maintenance schedules. To cope with this transition, executives must speed up organizational learning initiatives to groom new tech talent to utilize tools to assist them in managing this transition through a structured method, or else the cost of doing business will outweigh the profits of its outputs.
Set objectives:
In this case, the client has furnished controlled sensor data on one machine, collected over five months (April 2018 through August 2018). The device has sensors that archive telemetry readings over time. Based on the data provided the analysis required in this set is to predict at which reading thresholds the machine would fail so that the company could optimize labor work schedules to support the maintenance operations proactively. The findings from this analysis will be applied globally throughout the organization's maintenance program signaling to managers when proactive maintenance is required, thus shifting labor cost to proactive sustainment to keep operations functioning.
Who might care:
• Maintenance Managers, Operations Managers, Capital Expenditure planners, and manufacturing organizations such as the Department of Defense, Exon Mobile Corporation, General Motors, Ford Motors, Apple, and Boeing, and the Department of Defense, etc. can use such a model to predict the likelihood of equipment failures to allocate resources better. They can then proactively inform their maintainers and or customers well in advance of potential disruptions in their respective operations. Understanding the probability of material failures will help sustain customer service or level of service efforts. From the customer's point of view, it would be very convenient in knowing if a supply, production, or any other disruption may occur so that they can in turn, proactively mitigate risk. On the manufacturer's hand, such a predictive model would enhance the product base and performance of the organization's operations. Moreover, there is a possibility of developing an app or other front-end communication effort in which customers and or internal users can consult with to understand the likelihood of issues well in advance.
Cost and benefits: Every maintenance hour reduced in human labor will save the company and an average of 75.69 dollars which includes fringe benefits. The company's current budget for the machine maintenance (one machine) in this analysis consists of a staff of 3 maintainers overseeing one machine with 51 telemetry sensors each week working on average 50 hours a week at an operating expense of $11,353.50 a week or about $590,382. When factoring this number across the enterprise’s other 113 machines ($67M) the cost savings in deploying machine intelligence and optimization is diametrically required.