Over the past four years, Predictronics has offered consulting services to several companies looking to make use of their data and develop industrial analytics applications.

Doing so has allowed our team to work with customers from various industries, including transportation, manufacturing, heating and cooling, oil and gas and medical. For most applications, the focus was on predictive maintenance, or predicting when a machine is going to fail and preventing it from doing so.

After so many different projects, our team has realized that industrial predictive analytics requires a specific combination of skills in domain knowledge and data science, as well as time and effort.

What Skills are Needed for Developing Industrial Analytics?

One of industrial predictive analytics’ biggest challenges is that it requires personnel with both data science skills and engineering or domain knowledge. Often times, predictive analytics teams only master either one of those skills, which in turn limits their ability to positively impact the industrial segment.

Industrial analytics and data science go hand-in-hand. That is because machine data from the field can be particularly complex, with a multitude of signals. The important signals are affected by the machine’s health condition, as well as other variables, such as the operating regime and work environment. Therefore, monitoring the value of only a few signals is impractical. Advanced machine learning algorithms are a more accurate way to model the relationships between machine signals and health.

At the same time, purely data-driven approaches often fail or aren’t as precise as they could be. At the beginning of a project, there is usually little to zero validation data (data from an unhealthy condition), which is required for supervised learning techniques to work. In addition, it is difficult to know what to look for in the data without any domain knowledge of the machine.

How Much Effort is Required during Development?

There are three steps to developing and deploying industrial predictive analytics. The first is to implement a data collection infrastructure. The second is to develop the analysis models that will convert data into information. Finally, the third is to deploy the solution so results can be displayed on an interface.

When it comes to model development, a lot of time is spent on pre-processing steps, including data parsing, outlier removal and data segmentation. Data parsing can take a lot of time, mainly because parsing scrips are typically written from scratch and data comes in many different formats. Since industrial data usually contains a lot of noise, outlier removal is needed to clean the data and remove noise that should not be included in the health model. Lastly, data segmentation adds context to the data by labelling it with whatever the machine is doing during that time period.

After the data is pre-processed, machine learning is used to convert signals or features into health values. There are many algorithms available for this, making it impossible to know which ones will perform best. The best alternative is to try multiple algorithms and see which ones stand out.

How PDX can help

An increased interest in industrial analytics has prompted many companies to release platforms designed to speed up the development and deployment time of industrial applications. However, a lack of experience in developing those applications has resulted in many incomplete platforms.

That is Predictronics’ main differentiator. PDX, our industrial predictive analytics software platform, contains domain knowledge of critical components such as bearing and gearboxes. It also includes data pre-processing tools and machine learning algorithms that are well-suited for industrial predictive analytics.

With PDX, users with little experience are capable of developing accurate predictive models. Meanwhile, those who already have significant experience can develop and deploy solutions much faster.