Turning Data into Actionable Insights
Data and Insights
Insights are the new gold not the data as data is worth very little unless this data is turned into critical actionable insights. These insights can be used to support decision making and can help in refining the design and manufacturing processes. Machine learning algorithms can be used to accomplish different applied data mining tasks. These tasks can be descriptive analytics, predictive analytics, diagnostic analytics, and prescriptive analytics. Descriptive data analytics provides insight into the past and the present while predictive analytics forecasts the future. Diagnostic analytics provides root-cause analysis and prescriptive analytics advises on possible outcomes and their anticipated impacts.
Data Analytics
Descriptive data analytics provides a better understanding of the data and its nature and identifies patterns or relationships in the data. These descriptive models answer the following questions:
- What has happened?
- What happens now?
- What is the trend of a certain variable?
- What is the relationship between variables?
- How an item is performing w.r.t other item or a benchmark item?
Predictive models make prediction about future values of data and forecasts new proprieties instead of just exploring data properties like in case of descriptive analytics. These models answer the following questions:
- What would happen?
- When would it happen?
- Where would it happen?
Diagnostic analytics provides root-cause analysis to answer questions like:
- Why did it happen?
- Why would it happen?
Last but not least, prescriptive models focus on decision support. This decision support or recommendation engine answers the following questions.
- How would the predictions obtained from predictive models impact everything else?
- What are the proactive decision/actions to be made?
- How we benefit from predictions/recommendations?
- What are the best actions to make?
- What is the best time to take this action?
- What would be the impact of this action?
Predictive maintenance as an example
Predictive maintenance is a preventive maintenance approach that relies on regular monitoring and analysis of actual machine condition, operating efficiency, other indicators of the operating condition in order to detect incipient problems, minimize the number and cost of unscheduled outages and prevent catastrophic failure of critical equipment. According to McKinsey, predictive maintenance will help companies save $630 Billion by 2025. Fortunately, electro-mechanical equipment does not break without warning. Months before the faults occur, minimum vibration can be found. Weeks before the fault, apparent noise begins to develop. Days before the machine heats up and minutes before the break down, it starts to smoke.
In this context, descriptive analytics can generate insight about the current conditions of the machine. While a predictive model can perform health assessment and anticipates any incipient problem or possible failure or anomalies. Diagnostic analytics provides root-cause analysis and finally, prescriptive analytics model can process information about the current status of the machine generated by descriptive analytics module and the health assessment from predictive analytics modules and any other a priori knowledge that may be available. It then produces timely recommendations/ decisions about proactive actions or maintenance plan to mitigate any possible risk. Using these recommended decisions and anticipated impact information generated by the prescriptive analytics model, the user can proactively plan maintenance considering the highest priority maintenance needs or can create work orders and messages the assigned technicians about the needed maintenance in order to keep the productivity and reduce down time and maintenance expenditures.
Data summarization, visualization, clustering, data association and sequence discovery are common tasks in descriptive analytics. While predictive analytics uses classification, regression and time-series analysis to provide information like anomaly prediction, failure risk score, time to failure or remaining useful life estimate and/or degrading trends. Diagnostic analytics used techniques like events and causal factor analysis, change analysis, barrier analysis, fault tree analysis and natural language processing to provide explanation for the causes of possible breakdown. Prescriptive analytics employs different AI techniques from machine learning, algorithmic inference, optimization and natural language processing to provide recommendations to mitigate any possible risk in the machine.
In the current weak/narrow AI wave, classification and regression are the most commonly matured and used ML techniques. Strong/general AI wave will mature more sophisticated ML techniques to process the data relying more on cognitive data collection and preparation, clustering, data association and time-series analysis including sequence discovery, trend/cyclicity/seasonality, similarity analysis and anomaly detection in archived, steaming and live data.
