Demystifying Machine Learning

Machine learning can help organizations improve manufacturing operations and increase efficiency, productivity, and safety by analyzing data from connected machines and sensors, machine. For example, its algorithms can predict when equipment will likely fail, so manufacturers can schedule maintenance before problems occur, thereby reducing downtime and repair costs.

How machine learning works

Machine learning teaches computers to learn from data – to do things without being specifically told how to do them. It is a type of artificial intelligence that enables computers to automatically learn or improve their performances by learning from their experiences.

machine learning stepsImagine you have a bunch of toy cars and want to teach a computer to sort them into two groups: red and blue cars. You could show the computer many pictures of red and blue cars and say, “this is a red car” or “this is a blue car” for each one.

After seeing enough examples, the computer can start to guess which group a car belongs in, even if it’s a car that it hasn’t seen before. The machine is “learning” from the examples you show to make better and better guesses over time. That’s machine learning!

Steps to translate it to industrial use case

As in the toy car example, we must have pictures of each specimen and describe them to the computer. The image, in this case, is made up of data points and the description is a label. The sensors collecting data can be fed to the machine learning algorithm in different stages of the machine operation – like when it is running optimally, needs inspection, or needs maintenance, etc.

Data taken from vibration, temperature or pressure measures, etc., can be read from different sensors, depending on the type of machine or process to monitor.

In essence, the algorithm finds a pattern for each stage of the machine’s operation. It can notify the operator about what must be done given enough data points when it starts to veer toward a different stage.

What infrastructure is needed? Can my PLC do it?

The infrastructure needed can vary depending on the algorithm’s complexity and the data volume. Small and simple tasks like anomaly detection can be used on edge devices but not on traditional automation controllers like PLCs. Complex algorithms and significant volumes of data require more extensive infrastructure to do it in a reasonable time. The factor is the processing power, and as close to real-time we can detect the machine’s state, the better the usability.

Using Guided Changeover to Reduce Maintenance Costs, Downtime

A guided changeover system can drastically reduce the errors involved with machine operation, especially when added to machines using fully automated changeovers. Processing multiple parts and recipes during a production routine requires a range of machines, and tolerances are important to quantify. Only relying on the human element is detrimental to profits, machine maintenance, and production volumes. Implementing operator assistance to guide visual guidance will reveal inefficiencies and allow for vast improvements.

Removing human error

Unverified manual adjustments may cause machine fatigue or failure. In a traditional manual changeover system, the frequency of machine maintenance is greater if proper tolerances are not observed at each changeover. Using IO-Link can remove the variable of human error with step-by-step instructions paired with precise sensors in closed-loop feedback. The machine can start up and run only when all parts are in the correct position.

Preventative maintenance and condition monitoring

Preventative maintenance is achievable with the assistance of sensors, technology, and systems. Using condition monitoring for motors, pumps and critical components can help prevent the need for maintenance and notably improve the effectiveness of maintenance with custom alerts and notifications with a highly useful database and graphing function.

A repeatable maintenance routine based on condition monitoring data and using a system to guide machine changeover will prolong machine life and potentially eliminate downtime altogether.

For more, read this real-world application story, including an automated format change to eliminate human error, reduce waste and decrease downtime.

Machine Failures and Condition Monitoring – Selecting Sensors

In previous blogs, we discussed the different types of machine failures and their implications for different maintenance approaches, the cost-benefit tradeoffs of these maintenance approaches, and the progression of machine failures and indicators that emerge at various failure phases. We now will connect the different failure indicators to the sensors which can detect them.

The Potential – Functional Failure (P-F) curve gives a rough picture of when various indicators may emerge during the progression of a failure:

Each indicator can be detected by one or more types of sensors. Selection of the “best” sensor will depend on the machine/asset being monitored, other attributes being sensed, budget/cost-benefit tradeoff, and the maintenance approach. In some cases, a single-purpose, dedicated condition-monitoring sensor may be the right choice. In other cases, a multi-function sensor (“Smart Automation and Monitoring System sensor”) which can handle both condition monitoring and standard sensing tasks may be an elegant and cost-effective solution.

The table below gives some guidance to possible single- and multi-function sensors which can address the various indicators:

* Condition monitoring sensors are specialized sensors that can often detect multiple indicators including vibration, temperature, humidity, and ambient pressure.

# Smart Automation and Monitoring System sensors add condition monitoring sensing, such as vibration and temperature, to their standard sensing functions, such as photoelectric, inductive, or capacitive sensing

There is a wide range of sensors that can provide the information needed for condition monitoring indication. The table above can provide some guidance, selecting the best fit requires an evaluation of the application, the costs/benefits, and fit with the maintenance strategy.

IO-Link Changeover: ID Without RFID – Hub ID

When looking at flexible manufacturing, what first comes to mind are the challenges of handling product changeovers. It is more and more common for manufacturers to produce multiple products on the same production line, as well as to perform multiple operations in the same space.

Accomplishing this and making these machines more flexible requires changing machine parts to allow for different stages in the production cycle. These interchangeable parts are all throughout a plant: die changes, tooling changes, fixture changes, end-of-arm tooling, and more.

When swapping out these interchangeable parts it is crucial you can identify what tooling is in place and ensure that it is correct.

ID without RFID

When it comes to identifying assets in manufacturing today, typically the first option companies consider is Radio-Frequency Identification (RFID). Understandably so, as this is a great solution, especially when tooling does not need an electrical connection. It also allows additional information beyond just identification to be read and written on the tag on the asset.

It is more and more common in changeover applications for tooling, fixtures, dies, or end-of-arm tooling to require some sort of electrical connection for power, communication, I/O, etc. If this is the case, using RFID may be redundant, depending on the overall application. Let’s consider identifying these changeable parts without incurring additional costs such as RFID or barcode readers.

Hub ID with IO-Link

In changeover applications that use IO-Link, the most common devices used on the physical tooling are IO-Link hubs. IO-Link system architectures are very customizable, allowing great flexibility to different varieties of tooling when changeover is needed. Using a single IO-Link port on an IO-Link master block, a standard prox cable, and hub(s), there is the capability of up to: 

    • 30 Digital Inputs/Outputs or
    • 14 Digital Inputs/Outputs and Valve Manifold Control or
    • 8 Digital Inputs/Outputs and 4 Analog Voltage/Current Signals or
    • 8 Analog Input Signals (Voltage/Current, Pt Sensor, and Thermocouple)

When using a setup like this, an IO-Link 1.1 hub (or any IO-Link 1.1 device) can store unique identification data. This is done via the Serial Number Parameter and/or Application Specific Tag Parameter. They act as a 16- or 32-byte memory location for customizable alphanumeric information. This allows for tooling to have any name stored within that memory location. For example, Fixture 44, Die 12, Tool 78, EOAT 123, etc. Once there is a connection, the controller can request the identification data from the tool to ensure it is using the correct tool for the upcoming process.

By using IO-Link, there are a plethora of options for changeover tooling design, regardless of various I/O requirements. Also, you can identify your tooling without adding RFID or any other redundant hardware. Even so, in the growing world of Industry 4.0 and the Industrial Internet of Things, is this enough information to be getting from your tooling?

In addition to the diagnostics and parameter setting benefits of IO-Link, there are now hub options with condition monitoring capabilities. These allow for even more information from your tooling and fixtures like:

    • Vibration detection
    • Internal temperature monitoring
    • Voltage and current monitoring
    • Operating hours counter

Flexible manufacturing is no doubt a challenge and there are many more things to consider for die, tooling and fixture changes, and end-of-arm tooling outside of just ID. Thankfully, there are many solutions within the IO-Link toolbox.

For your next changeover, I recommend checking out Non-Contact Inductive Couplers Provide Wiring Advantages, Added Flexibility and Cost Savings Over Industrial Multi-Pin Connectors for a great solution for non-contact connectivity that can work directly with Hub ID.

Industrial Machinery Failure Types and Implications for Maintenance Approaches

Industrial machinery can fail in many different ways and for many different reasons. For critical and/or expensive equipment, it is a major challenge to find a way to detect potential failures before they happen and to take action to prevent or minimize the effects. Closely tied to this is the tradeoff between the cost of detection and the cost of failure. We discussed some of these tradeoffs in the blog “Condition Monitoring & Predictive Maintenance: Cost-Benefit Tradeoffs.”

When assessing how equipment might fail, several industry studies* have identified six primary failure types which may be considered:

    • Type A: Lower probability of failure in early- and mid-life of the asset, with a dramatic increase in probability of failure in late-life. This is typical for mechanical devices, such as engines, fans, compressors, and machines.
    • Type B: Higher initial probability of failure when the asset is new, with a much lower/steady failure probability over the rest of the asset’s life. This is often the profile for electronic devices such as computers, PLCs, etc.
    • Type C: Lower initial probability of failure when the asset is new, with an increase to a steady failure probability in mid- and late-life. These are often devices with high stress work conditions, such as pressure relief valves.
    • Type D: Consistent probability of failure throughout the asset life, similar failure probability in early-, mid- and late-life. This can cover many types of industrial machines, often with stable, proven design and components.
    • Type E: Higher probability of failure in early- and late-life, a lower and constant probability of failure in mid-life (often called a “bathtub curve”). This can be devices that “settle in” after a wear-in period and then experience higher failures at the end of life, such as bearings.
    • Type F: Lower probability of failure when new, with a gradual increase over time and without the steep increase in failure probability at the end of life of Type A. This is often typical where age-based wear is steady and gradual in equipment such as turbine engines and structural components (pressure vessels, beams, etc.).

Age-related and non-age-related failures

These six failure types fall into two categories: age-related and non-age-related failures. The studies show that 15-30% of failures are age-related (Types A, E & F) and 70-85% of failures are non-age-related (Types B, C & D). The age-related failures have a clear correlation between the age of the asset and the likelihood of failure. In these cases, preventative maintenance at regular time-based intervals may be appropriate and cost-effective. The non-age-based failures are more “random,” due to improper design/installation, operator error, quality issues, machine overuse, etc. In these cases, preventative maintenance will likely not prevent failure and may waste time and money on unnecessary maintenance.

Table is based on data from studies conducted by United Airlines (1978), Broberg (1973), U.S. Navy (1993 MSDP) and U.S. Navy (2001 SUBMEPP) and ARC Consulting

The fact that approximately 80% of failures are non-age-related has major implications for manufacturers trying to decide on a maintenance approach. The traditional preventative-maintenance approach is not likely to address these failures and may even cause failures when improperly done. It is therefore important to consider a more proactive approach, such as condition-based monitoring or predictive maintenance, especially for assets that are critical to the process and/or expensive.

Preventative maintenance and regular inspection may be a good approach for assets more likely to experience age-based failures in Types A, E, and F. These include fans, bearings, and structural components – and in many cases, the cost of condition monitoring or predictive maintenance may not be worth the cost. But for critical components or equipment, such as bearings on an expensive milling machine or transfer line, it may be worthwhile to apply condition monitoring or predictive maintenance.

And when the assets are more likely to experience non-age-related failures (Types B, C, and D), the proactive approaches are better. Many industrial machines and industrial control/motion components fall into this category, and condition monitoring or predictive maintenance can significantly reduce preventative maintenance costs and unplanned failures while improving machine uptime and Overall Equipment Effectiveness.

You can use this information to improve your maintenance operations. Start by considering your maintenance approach(es), especially for your most critical assets:

    • Are they more likely to experience age-related failures or non-age-related failures?
    • Should you change your maintenance approach to be more proactive?
    • What components and indicators should you measure?

We’ll discuss ideas on how to assess your equipment for condition monitoring/predictive maintenance and what you might measure in separate blogs.

* Studies conducted by United Airlines (1978), Broberg (1973), U.S. Navy (1993 MSDP) and U.S. Navy (2001 SUBMEPP)

Predictive Maintenance vs. Predictive Analytics, What’s the Difference?

With more and more customers getting onboard with IIoT applications in their plants, a new era of efficiency is lurking around the corner. Automation for maintenance is on the rise thanks to a shortage of qualified maintenance techs coinciding with a desire for more efficient maintenance, reduced downtime, and the inroads IT is making on the plant floor. Predictive Maintenance and Predictive Analytics are part of almost every conversation in manufacturing these days, and often the words are used interchangeably.

This blog is intended to make the clear distinction between these phrases and put into perspective the benefits that maintenance automation brings to the table for plant management and decision-makers, to ensure they can bring to their plants focused innovation and boost efficiencies throughout them.

Before we jump into the meat of the topic, let’s quickly review the earlier stages of the maintenance continuum.

Reactive and Preventative approaches

The Reactive and Preventative approaches are most commonly used in the maintenance continuum. With a Reactive approach, we basically run the machine or line until a failure occurs. This is the most efficient approach with the least downtime while the machine or line runs. Unfortunately, when the machine or line comes to a screeching stop, it presents us with the most costly of downtimes in terms of time wasted and the cost of machine repairs.

The Preventative approach calls for scheduled maintenance on the machine or line to avoid impending machine failures and reduce unplanned downtimes. Unfortunately, the Preventative maintenance strategy does not catch approximately 80% of machine failures. Of course, the Preventative approach is not a complete waste of time and money; regular tune-ups help the operations run smoother compared to the Reactive strategy.

Predictive Maintenance vs. Predictive Analytics

As more companies implement IIoT solutions, data has become exponentially more important to the way we automate machines and processes within a production plant, including maintenance processes. The idea behind Predictive Maintenance (PdM), aka condition-based maintenance, is that by frequently monitoring critical components of the machine, such as motors, pumps, or bearings, we can predict the impending failures of those components over time. Hence, we can prevent the failures by scheduling planned downtime to service machines or components in question. We take action based on predictive conditions or observations. The duration between the monitored condition and the action taken is much shorter here than in the Predictive Analytics approach.

Predictive Analytics, the next higher level on the maintenance continuum, refers to collecting the condition-based data over time, marrying it with expert knowledge of the system, and finally applying machine learning or artificial intelligence to predict the event or failure in the future. This can help avoid the failure altogether. Of course, it depends on the data sets we track, for how long, and how good our expert knowledge systems are.

So, the difference between Predictive Maintenance and Predictive Analytics, among other things, is the time between condition and action. In short, Predictive Maintenance is a stepping-stone to Predictive Analytics. Once in place, the system monitors and learns from the patterns to provide input on improving the system’s longevity and uptime. Predictive Maintenance or Preventative Maintenance does not add value in that respect.

While Preventative Maintenance and Predictive Maintenance promises shorter unplanned downtimes, Predictive Analytics promises avoidance of unplanned downtime and the reduction of planned downtime.

The first step to improving your plant floor OEE is with monitoring the conditions of the critical assets in the factory and collecting data regarding the failures.

Other related Automation Insights blogs:

How IO-Link Sensors With Condition Monitoring Features Work With PLCs

As manufacturers continually look for ways to maximize productivity and eliminate waste, automation sensors are taking on a new role in the plant. Once, sensors were used only to provide detection or measurement data so the PLC could process it and run the machine. Today, sensors with IO-Link measure environmental conditions like temperature, humidity, ambient pressure, vibration, inclination, operating hours, and signal strength. By setting alarm thresholds, it’s possible to program the PLC to use the resulting condition monitoring data to keep machines running smoothly.

Real-time data for real-time response

A sensor with condition monitoring features allows a PLC to use real-time data with the same speed it uses a sensor’s primary process data. This typically requires setting an alarm threshold at the sensor and a response to those alarms at the PLC.

When a vibration threshold is set up on the sensor and vibration occurs, for example, the PLC can alert the machine operator to quickly check the area, or even stop the machine, to look for a product jam, incorrect part, or whatever may be causing the vibration. By reacting to the alarm immediately, workers can reduce product waste and scrap.

Inclination feedback can provide diagnostics in troubleshooting. Suppose a sensor gets bumped and no longer detects its target, for example. The inclination alarm set in the sensor will indicate after a certain degree of movement that the sensor will no longer detect the part. The inclination readout can also help realign the sensor to the correct position.

Detection of other environmental factors, including humidity and higher-than-normal internal temperatures, can also be set, providing feedback on issues such as the unwanted presence of water or the machine running hotter than normal. Knowing these things in real-time can stop the PLC from running, preventing the breakdown of other critical machine components, such as motors and gearboxes.

These alarm bits can come from the sensors individually or combined together inside the sensor. Simple logic, like OR and AND statements, can be set on the sensor in the case of vibration OR inclination OR temperature alarm OR humidity, output a discrete signal to pin 2 of the sensors. Then pin 2 can be fed back through the same sensor cable as a discrete alarm signal to the PLC. A single bit showing when an alarm occurs can alert the operator to look into the alarm condition before running the machine. Otherwise, a simple ladder rung can be added in the PLC to look at a single discrete alarm bit and put the machine into a safe mode if conditions require it.

In a way, the sensor monitors itself for environmental conditions and alerts the PLC when necessary. The PLC does not need to create extra logic to monitor the different variables.

Other critical data points, such as operating hours, boot cycle counters, and current and voltage consumption, can help establish a preventative and predictive maintenance schedule. These data sets are available internally on the sensors and can be read out to help develop maintenance schedules and cut down on surprise downtimes.

Beyond the immediate benefits of the data, it can be analyzed and trended over time to see the best use cases of each. Just as a PLC shouldn’t be monitoring each alarm condition individually, this data must not be gathered in the PLC, as there is typically only a limited amount of memory, and the job of the PLC is to control the machines.

This is where the IT world of high-level supervision of machines and processes comes into play. Part two of my blog will explore how to integrate this sensor data into the IT level for use alongside the PLC.