Industry 4.0 is the term used for the fourth industrial revolution. If we view revolution as a disruptive change that emerges from the re-ordering of established hierarchies, then what is the future of Statistical Process Control analysis in the new industrial era? To address this question, I will explore how Machine Learning for predictive analytics is changing the field of asset maintenance while usurping traditional statistical modelling.

Let’s start with a simplified explanation of how Mean Time Between Failures (MTBF) and statistical modelling have been historically applied to the field of industrial analytics for asset maintenance. MTBF can be provided by the OEM or it can be calculated at a facility level with PLC generated operational performance data.

There are over 25 different standards and methodologies for calculating MTBF such as UTE 80-810 and MIL-217 Standard. Statisticians use multiple statistical tools, ranging from Microsoft Excel to Siemens SN 29500. Although this article addresses MTBF, it is relevant for other statistical measurements including Mean Time To Failure (MTTF) and Mean Time To Repair (MTTR).

No matter what the software package or model, the commonality with statistical modelling is that mathematical equations are used to formalize relationships between dependent and independent variables. Hypotheses are generated from sample data and extrapolated to a broader asset population.

The image below is of the Bathtub Curve that represents the Useful Life of an asset. During the first phase or Break-In period, there is a relatively high failure rate due to manufacturing defects. The later phase of the asset lifetime is referred to as Wear Out. During Wear Out, there is an increase in failure rate.

MTBF is a measurement based on the theory that random failures occur during the Steady State at a constant rate. MTBF provides the average useful life period between failures. Once the MTBF period (in hours) is calculated, the probability of failure for a specific machine can be estimated. Calculating probability requires statistically significant sample sizes and in some cases, high computational power.

Predictive Maintenance (PdM) is a well-established reliability method where machinery is manually inspected by plant technicians. Evidence of machine degradation identified during these inspections is flagged and triggers further maintenance activities. MTBF is a key factor in the scheduling and frequency of PdM activities.

Is MTBF Effective? There have been several technical and methodological concerns raised about MTBF including those relating to the use of obsolete or faulty assumptions. From a reliability perspective, the following issues are noteworthy:

- MTBF is based on historic failure rates and does not use real-time data.
- MTBF is an average measurement of failure based on a sample population that is extrapolated using statistical tools. The calculation is not based on the data generated from the specific machine.
- When the probability of failure is calculated, it comes with an important caveat: that the machine is operating under the stated conditions of the equation. In the real operating environment, there are numerous variables that cannot be included in a statistical equation including environmental factors, human error and equipment abuse or overuse.
- Machine-level data is not easily accessible and other inputs for complex probability studies are either unavailable or non-exist.

**Ultimately, the limitation of MTBF is that it is rooted in the field of statistical modelling, which relies on extrapolation from a small sample and model assumptions that are often unrealistic. In the opinion of many data scientists and reliability and maintenance professionals, MTBF is inherently flawed.**

**Maintenance in the era of Machine Learning**

In recent years, there has been a move to apply the discipline of Machine Learning to industrial plant maintenance. Instead of formulas and equations, we apply algorithms to Big Data.

Here is a brief overview of how Machine Learning is used for asset maintenance. There are hundreds or thousands of sensors in an industrial plant that generate data which is extracted to an analytics platform. The learning algorithm analyzes the data and detects abnormal behavioral or patterns of abnormal behavior. Anomaly detection is performed in real time on the data to identify asset degradation or failure. Using Machine Learning, Root Cause Failure Analysis can be traced, which provides reliability and maintenance technicians with information to remediate and limit future failure incidents.

At best, statistical modelling can estimate the probability of failure based on asset class averages and model assumptions that are unstable. Machine Learning provides specific recommendations of Time To Failure for the specific asset, without relying on Mean data.

**The future of Predictive Asset Maintenance**

Time-based PdM will continue to play a significant role in maintenance and reliability operations. over time, we expect that MTBF will become less relevant as Machine Learning algorithms replace the traditional Statistical Process Control. We additionally expect the adoption of Machine Learning to ultimately reduce the need for unnecessary or redundant maintenance activities and a shift toward Predictive Maintenance (PdM).

The graph below shows a high-level depiction of the current state of maintenance. It highlights two scenarios where the cost of maintenance is excessive. In the case of under-maintenance, we see an excess of Reactive Maintenance activities which are expensive and lead to asset downtime. The alternative scenario of over-maintenance results in spending resources beyond the optimal range.

The graph below is a modification of the above graph and it includes Machine Learning. With Machine Learning, industrial plants can reduce both Reactive and Preventive Maintenance, thereby lowering the overall investment in maintenance resources.

**Summary and conclusion**

The operationalization of Big Data generated by industrial plant assets is fueling the growth of Industry 4.0. Big Data forms the basis of Industrial Analytics and over time is expected to replace traditional statistical modelling techniques. As Artificial Intelligence algorithms replace physical simulations equations, MTBF and other Mean measurements will become less relevant in the operations of the Smart Factory.

Industrial plants generate terabytes of process data. SKF Enlight AI is a SaaS Predictive Maintenance solution that uses Automated Machine Learning to identify emerging asset failure patterns within this data. It provides early warnings and sensor-level intelligence to help avert unplanned downtime and meet production goals. For more information on how SKF Enlight AI can improve performance and reliability, click here.