Select Page

As a co-founder of Presenso, a day does not pass without speaking to a business executive about the current state of Industrial Analytics.  Setting the bar high, Gartner claims that “manufacturing and automotive/transportation have emerged as the early adopters of IoT”.   Although it’s true that we have seen a surge of interest in Industrial Analytics for Predictive Asset Maintenance, many industrial plants have almost no meaningful capabilities.   As the industrial world embraces advances in Artificial Intelligence and Machine Learning, many plants still consider Microsoft Excel as their default statistical package.

This article covers the following topics:

  • The current state of Machine Learning and how this impacts world of Industrial Analytics for Predictive Asset Maintenance
  • An analysis of the organizational and technological constraints preventing the adoption of Industrial Analytics
  • A Maturity Model that can be used by companies to identify a relative measure of their industrial analytics capabilities

The Current State of Machine Learning and IIoT for Predictive Asset Maintenance

According to most observers, we are experiencing an industrial revolution of historic magnitude.  To understand the root cause of a revolution, consider the words of a man with personal experience, Vladimir Lenin.  According to Lenin, a “revolution is impossible without a revolutionary situation; furthermore, not every revolutionary situation leads to revolution.”

Based on Lenin’s view of a revolution, what is the “revolutionary situation” driving Industry 4.0 and Industrial Analytics for Predictive Asset Maintenance?  The biggest driver of Industrial Analytics for Predictive Asset Maintenance is cost the sharp reduction in the cost of data storage, bandwidth, transfer and computational power.

Declining Cost in Bandwidth, Storage and Computing Figure 1:  Declining Cost in Bandwidth, Storage and Computing

Furthermore, the cost of IoT sensors has declined from $1.30 in 2004 to a projected $0.30 by 2020.

A second driver of Industrial Analytics for Predictive Asset Maintenance is the democratization of the body of knowledge related to Machine Learning.  Historically, Machine Learning was a discipline limited to academia.  Today, market forces are causing the commercialization of Machine Learning as technologies are applied to new arenas.   The surge in innovation in Machine Learning is due in part to the influx of investor funding in this sector.

What is preventing the adoption of Industrial Analytics for Predictive Asset Maintenance?

Despite the seemingly supportive executive buy-in, penetration levels of IIOT for Predictive Maintenance have lagged expectations.  To some extent, industrial plants are waiting on the sidelines to determine which IIOT infrastructure will gain consensus support.

One issue that is often overlooked is the sunk costs in existing legacy systems.   Programmable Logic Controller (PLC) have been developed and customized over a period of time, and replacement could be costly or disruptive.

A further inhibiting factor is the perceived lack of organization resources required to support deployment.  The rationale is that many plants are struggling with skill shortages to support Industry 3.0 practices.   Given that Industry 4.0 is based on Big Data, the industrial sector lacks the expertise for wide-scale adoption.

Finally, many enterprises are addressing seemingly more pressing issues ranging including aging infrastructure, regulatory and commodity pricing pressure.  While executives may be paying lip-service to Industry 4.0, middle management’s incentives are aligned to short-term operational metrics.

IIOT Maturity Model of Industrial Analytics Capabilities

IIOT Maturity ModelFigure 2: Machine Learning for Predictive Analytics Maturing Model

 Each of the four levels of IIOT for Predictive Analytics is outlined below.

 IIOT Maturity ModelFigure 3:   Rule Base Machine Monitoring Architecture

 01 – Rule Base Machine Monitoring: 

 The most basic form of Industrial Analytics is the SCADA-based machine monitoring.   Manually control threshold limits are set by either the machine OEM vendor or plant reliability engineers.  Examples include temperature, humidity, vibration and pressure.  If the measurement exceeds manually-set control thresholds, then an alert is triggered resulting in further investigation.   In some cases, specific processes can be initiated automatically based on the breach of a threshold (e.g., cooling processing initiated or a machine can be automatically shut down).

SCADA Architeture

Figure 4: SCADA Based Control Thresholds

The most significant limitation of the SCADA rule based monitoring system is the fact that triggers are only alerted when control thresholds have already been breached.   In Figure 4 below, the temperature sensor data is monitored.   If the temperature reaches 40 degrees or falls below 20 degrees, then an alert is generated.   The ability to detect abnormal sensor behavior between data thresholds is not monitored and in many cases, once the breach has occurred, it is too late to remediate.

 IIOT Architecture

 

Figure 5:  – Manual Statistical Modelling Architecture

 02 – Manual Statistical Modelling for Historic Trends: 

The next level in IIOT Predictive Maintenance Maturity is Manual Statistical Modelling.  The underlying principle driving manual statistical modelling is that machine assets life can be divided into three phases:  break-in phase when machines often fail due to so-called infant mortality, the steady state during which time machine failures are not related to time and the wear out-phase where failures occur based on end-of-life factors.

Bathtub Curve IIOT

Figure 6: Hypothetical Asset Failure Rate versus Time (Bathtub Curve)

The goal of statistical modelling is to determine Mean Time Before Failure (MTBF) or Mean Time to Repair (MTTR).  Using this data, Operations and Maintenance (O&M) develop preventive maintenance programs.   These statistical models assume that although the timing of a failure for a particular machine is not calculable, failure incidence can be derived from historic data and therefore average time to failure can be determined.

From our first-hand experiences, a surprisingly large number of industrial plants still rely on manual statistical modelling despite the well-known drawbacks. Let’s put aside the limitations of using a sample population extrapolate to larger asset class instead of Big Data. Manual statistical modelling is a rudimentary short-cut for predictive asset maintenance and of limited operational value because it provides historic averages of asset failure instead of actionable forecasts for future failure incidents.

 Advanced Statistical Modelling IIOT

 

Figure 7:  – Advanced Statistical Modelling Architecture

03 – Advanced Statistical Modelling

The next layer of IIOT Predictive Asset Maintenance is Advanced Statistical Modelling.  Using offline PLC data, statistical models analyze multiple variables to predict machine failure probability.  Advanced Statistical Modelling is the bridge between Manual Statistical Modelling (trend analysis) and Machine Learning for Asset Maintenance that provides real time failure prediction based on anomalous behavior.

The drawbacks to Advanced Statistical Modelling are both practical and methodological.  From a practical perspective, it is resource intensive.   Although Manual Statistical Modelling can be performed with limited knowledge of statistics using basic software packages, Advanced Statistical Modelling requires expertise in the field of statistics and the use of advanced analytics software.  Because factory asset data is not automatically generated, there are multiple manual processes required to access, transport and cleanse the sample data that is analyzed.   These cumbersome requirements are not only time consuming.  As more data is analyzed, Advanced Statistical Modelling becomes increasingly expensive and organizational support that is often difficult to secure.

Advanced Statistical Modelling relies on analyses that have not matured at the same pace as innovations in Artificial Intelligence.  One of the biggest challenges is modelling with imbalanced data when only a fraction of the data that is analyzed is categorized as failure. While normal operational data (i.e. non-failure data) which comprises most of the data and is similar to each other, failure data may be different from one another.

Why is this important?  Metrics used to evaluate the model can be misleading. For example, in a classification model for a dataset with more than 99{e1a6989ef918f020ac33d1bc034852f3918ebb2ac141867799b34632c4d2fe93} non-failure data and less than 1{e1a6989ef918f020ac33d1bc034852f3918ebb2ac141867799b34632c4d2fe93} failure data, a near perfect accuracy could be achieved simply by assigning all instances in the data to the majority (non-failure) class. This model however is not useful as it has never learned to predict a failure.

There are workaround solutions such as random oversampling of the minority class to increase its size, or random subsampling of the majority class to decrease its size. However, random oversampling could lead to underrepresentation of “good” samples and overrepresentation of “bad” samples.  This causes overfitting and un-useful generalization due to the loss of valuable information.

 Automated Machine Learning Architecture

Figure 8:  – Automated Machine Learning for Predictive Maintenance Architecture

 04-Automated Machine Learning for Predictive Maintenance

Automated Machine Learning for Predictive Maintenance is a methodology to extract Big Data generated from plant sensors and analyze the data using advanced Artificial Intelligence algorithms.  The goal is to detect abnormal data patterns.

Machine Learning algorithms detect anomalous behavior and patterns of anomalous behavior.  These are used to detect evolving degradation or failure.  With this information, Root Cause Analysis (RCA) that and Time to Failure (TTF) estimates are generated.

Machine Learning can be categorized into two methodologies: Supervised and Unsupervised.  With Supervised Machine Learning the algorithm is “trained” using human guidance and labels of abnormal and normal machine conditions. When new data is analyzed by the algorithm, it can then classify this data as failure if it recognizes the pattern from its training.

Machine Learning Anomaly Detection

Figure 9: Machine Learning Anomaly Detection and Correlation Detection

With Unsupervised Machine Learning, the algorithm does not need to be trained using the physical blueprints or knowledge about the process itself.  Furthermore, with cloud based Unsupervised Machine Learning, plant reliability and maintenance staff are alerted to asset degradation and failure without the need for internal Big Data engineers or Data Scientist to interpret the data.   This is an important consideration for industrial plants that lack the resources to develop internal competencies in Big Data and Machine Learning.

The practical application is that there are significant resources required to “train” the model.  With Unsupervised Machine Learning the algorithm needs no knowledge of the physical layout of the machine or its mechanical processes.  In fact, the algorithm is agnostic to machine and sensor type.  This eliminates the need for the facility’s process engineers and Big Data scientists to train the algorithm.

Furthermore, with Automated Unsupervised Machine Learning, the algorithmic model is self-learning and self-maintaining and therefore can be applied to various types of machines, from various vendors.

Summary and Conclusion

With advances in Artificial Intelligence and Machine Learning, traditional SCADA and manual statistical modelling are likely to be replaced.  Advanced Statistical Modelling based on offline data is resource intensive and ultimately cannot provide real-time analytics that are actionable.  As Automated Machine Learning gains traction, we expect more industrial plants to rely on this solution for Predictive Maintenance.

Asset Maintenance Maturity Figure 10: Summary of Asset Maintenance Model Type

Are you interested in exploring practical steps to migrate towards predictive asset maintenance standards?  Click here and to schedule a demo of our Machine Learning for Asset Maintenance solution.

Like this article? Please share it using these icons....Share on Facebook
Facebook
Tweet about this on Twitter
Twitter
Share on LinkedIn
Linkedin
Deddy Lavid

Deddy Lavid

Experienced R&D Manager. Recognized expert in the field of Machine Learning and Big Data architecture. His work spans the full spectrum from researching isolated data problems to building complex production systems. At Rafael, he led a team of algorithm developers in large Software projects of national importance. Holds honors M.Sc. computer and information science.