Why Predictive Analytics Are Important For Manufacturers

The importance of data and analytics in modern enterprise has continued to rise. In fact, IDC expects spending on AI-powered applications, such as predictive analytics, to grow from $40.1 billion in 2019 to $95.5 billion by 2022. 

The goal of predictive analytics applications is to increase efficiency by using data to understand and evaluate complex systems and processes, and anticipate what will happen next. Artificial Intelligence (AI) and machine learning technologies can quickly analyze extremely high volumes of data, allowing teams to identify insights at a much faster rate. This can benefit a variety of areas in manufacturing, including production optimization, quality, maintenance and waste reduction.

In this post, we’ll define what predictive analytics are, outline why predictive analytics are important to successful manufacturing, and discuss the drawbacks of predictive analytics that manufacturers need to know.

What are Predictive Analytics?

Predictive analytics combine the power of historical data with AI and machine learning technology to understand, monitor and optimize processes. They also detect and identify trends, predict potential problems, and provide recommendations to improve the process and maximize performance. Industrial IoT platforms that leverage predictive analytics collect and analyze real-time process data in order to predict and prevent potential problems on the factory floor. 

In manufacturing, the first step to leveraging predictive analytics software is collecting, storing and organizing the process data generated by various machines, devices and systems on the factory floor. Typically, factories need three to six months worth of historic data in order to leverage predictive analytics effectively, although this time-frame will vary depending on the quantity of data generated and the problem that is being addressed.

For example, applications like predictive performance and predictive quality generate data extremely fast because production is running on a daily basis. Equipment failures, on the other hand, happen much more infrequently so it can take months if not years to generate the quantity of data needed for specific applications.  

Once collected, the historical data can be used to derive insights and make effective predictions based-on a wide range of variables including line speed, product quality and more. This includes identifying important relationships between variables, predicting variables of interest and empowering decision-makers to take early action in order to reduce waste and increase efficiency.

As factories become ever-more connected, predictive analytics technologies will become a core component on their digital transformation journey because they can help you become more efficient and competitive, and – ultimately – profitable.

Why Are Predictive Analytics Important For Manufacturers?

It’s clear we are going to see a rapid adoption of predictive technologies in the future, but who uses predictive analytics today? In manufacturing, forward-thinking factories are using predictive analytics to decrease the time to action significantly, saving time, money and materials, and speeding up the time to market. 

Manufacturers get advance warning of problems, such as potential quality failures and/or unplanned downtime due to machine failure, and allow operators to take corrective action.  For example, machine learning can predict a quality failure will occur in 10 minutes because line speed has dropped and when line speed has dropped in the past, products do not meet quality standards. 

Factories are also turning to machine learning and predictive analytics to identify production trends, solve problems faster and manage resources more efficiently. The ability to identify potential issues early on with predictive analytics enables factories to optimize their process and avoid the costs associated material waste, high scrap rates or downtime. For example, Oden Technologies has seen manufacturers achieve anywhere from a 10 to 50% reduction in total scrap which translates into significant savings.

In the face of an upcoming skilled labor shortage, machine learning and predictive analytics technology also has the added benefit of helping manufacturers attract digital-native talent entering the workplace. At a time when many factories find it hard to recruit and retain talent, the opportunity to work with cutting-edge solutions provides a value-added benefit.

How Predictive Analytics Works

The first step to deploying a predictive analytics solution is to collect data from machines and sensors and integrated with real-time operator data, offline quality data and data from historians, MES and ERP systems. This data is then cleaned, merged, formatted and structured in the cloud.  For example, if one machine tracks temperature in Fahrenheit and another machine tracks temperature in Celsius, the temperature needs to be converted into a unified metric. 

Drawing on historical data, machine learning algorithms can identify the patterns in behavior that have previously led to problems. If real-time activity starts to follow one of those problem patterns, the system is able to predict the potential outcome and alert factory personnel. Once operators, engineers or plant managers have been alerted, they can quickly take corrective action and prevent issues from having a significant impact. 

Let’s dive into a deeper explanation of what is predictive modeling and the key components of AI predictive analytics. 

Step One: Access and Explore Available Data

Data is often talked about in rather generalized terms, but when it comes to predictive analytics, it’s important to think very specifically about the types of data-sets required for your application. Data availability is one of the top reasons that predictive applications fail. A good rule of thumb to training effective models is to train them on datasets that include a number of instances that is at least 10 times the number of key variables being monitored. For instance, if you are trying to use 5 key variables to predict quality failure, you will need at least 50 instances of those failures (in addition to normal instances) within your dataset to build an accurate prediction model. 

In addition to availability, data quality is also crucial because predictions are only useful if they’re accurate. The factory floor does not have much margin for error so it’s important that data is cleaned, corrected, and contextualized properly.  In order to identify which specific data-sets to collect, you need to be clear on the goals you want to achieve, because it is your goals that determine which data-sets will be relevant and useful.

When defining an application it’s important to start with the motivating question: why are predictive analytics important for your organization? It helps to ask yourself: 

• • What creates value and impact for your factory?
  • What can predictive analytics help you achieve?
  • What do your data and infrastructure allow you to do?
  • What should you do?

The last question is the most important, just because you can use predictive analytics to solve a problem doesn’t mean you should.  It’s crucial to start with business value and impact first, then work your way down to what you should do.  

When establishing goals for predictive analytics, many factories start with predictive maintenance. While this can certainly provide value, it takes a longer time to build up the dataset because maintenance failures typically do not occur frequently.  Predictive performance and quality applications are typically better entry points and deliver actionable insights at a much faster rate. 

Step Two: Pre-Process Data For Accuracy

Raw data is often too raw. It needs to be formatted, contextualized, and organized in order to fully maximize its potential. One key part of this process is removing any misleading data. This means removing any data that gives a false impression of how the factory is performing.

For example, if a factory wanted to identify the temperature of a specific machine during an average product run, using data that was recorded on the hottest days of the year could be problematic. It would also be misleading to include data from when the product line was inactive, because during those times the temperature would not be a true reflection of activity.

Another part of this process is organizing data from uneven time-series. Real-world data collected off machines and sensors on the factory floor is rarely reported with consistent frequency due to measurement constraints, as well as, latency from internet connections.  For example, a metric that is being collected might be initially reported within ten seconds then due to latency it’s not reported again until a minute later. In such cases instead of looking at each individual metric, the preprocessing stage can choose to average the metric recordings over a five minute span for a more consistent recording. 

At the most basic level, preprocessing takes data that’s not very easy to use and makes it usable by removing the noise and formatting it into a consistent manner. This allows machine learning technologies to build accurate predictive analytics models.  

Step Three: Build & Validate Predictive Models in the Cloud

Building and validating predictive analytics models typically happens in the cloud because it offers better economies of scale for storage and significantly more power for machine learning and predictive analytics applications – especially because it allows you to easily combine datasets from multiple sources and simultaneously explore a large number and combination of possible model parameters. 

There are two main approaches to machine learning when it comes to developing predictive models, that are typically grouped under supervised learning approaches: 

Classification – Classification models are used to predict types. For example, in predictive quality applications this means predicting whether the product is good or bad. It can also be used to predict product size; is it too large, on target or too small.

Regression – Regression models are used to predict numerical values. For example, you can predict product diameter and line speed to improve the accuracy of production across lines. This allows you to create products that are more consistent in quality and size.  

Building and validating models is an iterative process, machine learning technologies are constantly looking for new ways to improve as changes are implemented on the factory floor. Models are continually re-trained and automatically updated when there are significant improvements or changes. During this process predictive analytics models are continuously validated against unseen datasets, or in other words, datasets that they were not trained on. This is done to ensure that the models are not overfitted, or only able to predict scenarios related to their training set, and can scale to provide meaningful results on previously unseen data.

Step Four: Deploy Models and Implement Insights from Predictive Analytics

Once a predictive analytics model has been built and validated, it needs to be deployed into a production environment – either in the cloud or on the edge. From here, machine learning technologies will run predictive analytics models against live production data in order proactively identify potential problems such as quality failures or unplanned downtime. This can be done by looking for specific patterns or a set of conditions that indicate a quality or machine failure may happen in the future. Real-time alerts can then notify operators, engineers or plant managers when key parameters are reaching defined thresholds allowing them to take corrective action and minimize impact. 

Predictive analytics insights can be taken one step further and not only predict performance but prescribe recommended settings for controllable variables, such as line speed, pressure or temperature. Predictive performance applications can optimize machine throughput and output, maximizing production runs without sacrificing product quality.

What Are The Benefits Predictive Analytics?

As the industry moves towards digitalization, manufacturers face greater pressure to maintain a competitive edge and many begin asking the question why predictive analytics? 

Predictive analytics are important for applications such as predictive quality and predictive performance that enable manufacturers to identify problems at their very earliest stages, so they can be dealt with before issues start to unfold. 

With return on investment a key driver for the industry, predictive analytics is capable of delivering very quick wins, with many factories seeing measurable cost savings and opportunities for optimization after just a few months. 

Predictive Analytics Benefits: Detect Patterns to Predict Performance 

Predictive analytics sifts through vast amounts of historical data much more quickly and accurately than a human can. Machine learning technologies are able to identify repeated patterns and other relationship variables; such as when we change these settings, we increase production by 10% without sacrificing first-pass yield.  

AI and machine learning can search for patterns and explore various combinations that help your organization identify potential efficiency improvements, predict issues and reduce waste. 

Predictive Analytics Benefits: Improve Operations in Real-Time

As we’ve already seen, predictive analytics deliver near real-time insights by comparing data from historical production runs with live production activity. These comparisons, which convert to both predictive and prescriptive analytics, drive recommendations and alerts to improve operations in real-time.  A cloud-edge hybrid approach, such as the one Oden offers, combines the power of the cloud with the business continuity of the edge allowing factory personnel to make better decisions, faster.

Predictive Analytics Benefits: Reduce Costs

Quality failures can translate into significant losses in product, overhead labor and time. Predictive analytics can enable factories identify quality failures and take corrective action faster to minimize impact and reduce the cost associated with waste. Prescriptive analytics can further these cost savings by allowing you to replicate your most efficient runs more consistently.  Additionally, predictive analytics and condition-based monitoring can help factories reduce unplanned downtime and lost productivity by notifying plants of potential equipment problems.

Predictive Analytics Benefits: Optimize to Perfection

Manufacturers have been accustomed to ‘Lean Principles’ for decades. These best practices sought to help manufacturers achieve maximum production efficiency with minimum waste. Predictive analytics ultimately provide manufacturers with real-world data to help them optimize their processes reach perfect perfection.

What Are the Drawbacks of Predictive Analytics?

Ultimately, predictive analytics technologies are a tool for people to use. The goal is to use machine learning-based methods to effectively monitor, predict, and optimize the manufacturing process, but it is up to people to take action. In other words, using predictive analytics won’t solve management challenges or problems that arise from human error, or incomplete information.

When it comes to the drawbacks of predictive analytics it ultimately boils down to what you do with the information.

Predictive Analytics Drawbacks: Incomplete Data

The age old saying “garbage in, garbage out” applies to predictive analytics as well. The data and scenarios a machine learning model observes is what it learns from.  Models that are built and trained with incomplete information, such as data points, labels or context, can provide inaccurate predictions. 

Additionally, if models are not trained properly or processes change, models can begin to drift over time. This means that the variable, such as quality, the model is trying to predict slowly changes causes predictions to become less accurate as time changes.

Predictive Analytics Drawbacks: Collecting Data Points

Connecting your factory floor can be overwhelming, especially if none of your equipment is connected. However, not every part of every machine needs a sensor in order to execute a predictive analytics solution. Trusted Industrial IoT and applied analytics technology partners can help determine the specific data points that are needed to support your application based on the business challenges you’re looking to solve. While having enough data is extremely important, you don’t need to collect data for the sake of collecting it.

Predictive Analytics Drawbacks: Expectation Bias

While predictive analytics deliver powerful insights, people can often set unrealistic expectations or interpret the results inappropriately.It is important for manufacturers to understand the accuracy-performance properties of ML models, such as the errors (both false positives and false negatives) models can make, and how they can take actions that account for these. It is also important to provide models feedback so that they can learn from errors, and improve over time as they observe more of the process. Buying into excessive hype that ML models can work perfectly, instantaneously and without any supervision is often detrimental to the effective use of these models.  

Who Uses Predictive Analytics?

Predictive analytics can be applied to manufacturers of almost any size and in any industry.  Applications might be more relevant to certain industries than others, but since predictive analytics depend on the data available, models can be used to predict just about anything. 

Let’s look at a few key roles within the factory:

Plant managers can benefit from predictive analytics, because they help optimize production and increase contribution margins.

Engineers can solve problems faster. They can analyze data quicker than before and use analytics driven process and quality recommendations to update recipes and processes as well as debug and root cause issues.

Operators can receive alerts of potential failures, so they can take corrective action faster, and prevent any downtime associated with quality or equipment failures

To be effective, all you need is a means to collect data – such as sensors – a place to collect that data, and data-literate staff to understand what the insights mean.

Predictive Analytics In Action: Reducing Waste with Scrap Prediction 

A leading wire and cable manufacturer wanted to empower plant-floor operators to identify the cause of scrap products and take corrective action in real-time. By using Oden’s predictive analytics solution, they were able to deploy an 80-90% accurate scrap prediction model that allowed operators to proactively identify when their process resulted in poor quality or high amounts of scrap thereby reducing wasted product or material. 

What Are Some Predictive Analytics Tools?

There are a number of predictive analytics tools on the market designed to make Industrial IoT and data analytics more accessible across the factory floor. This includes platforms that allow manufacturers to leverage data visualization tools, machine learning and more. Oden’s Applied Analytics and Industrial IoT platform helps plant managers, engineers, operators and quality control managers to find the most efficient way to make a product within a robust, secure hybrid cloud-edge environment.

With Oden’s predictive analytics capabilities, you’ll receive predictive alerts that allow you to take action fast to prevent quality and other performance failures. You’ll also benefit from interactive dashboards and data exploration that provide a snapshot of real-time performance as well as allow you to investigate root cause analysis to operate more efficiently.

Get Oden’s Help

Now we’ve outlined why predictive analytics is important, it’s time to take the next step.

Are you looking to implement or improve the predictive analytics at your factory? To get more information, get in touch to request a demo and see how Oden could help.