LESSONS WE HAVE LEARNED DEVELOPING PREDICTIVE ANALYTICS SOLUTIONS
MAY 23, 2018 - AARON SHELLY, DATA SCIENTIST
Companies that are new to predictive analytics are often very excited about the prospect of developing preventative maintenance strategies and performing real-time machine health monitoring. The only problem is that they aren’t always sure how they can start this process.
Through more than 30 projects, the Predictronics team has delivered impacts to companies in nearly every industry segment, including manufacturing, transportation and energy production. We have saved these companies, many of which are Fortune 500 companies, tens of thousands of dollars by preventing unplanned downtime and waste, as well as increasing efficiency and productivity.
Looking back on our previous work and projects, there are several lessons we have learned along the way.
Identify Realistic Project Goals Early On
Because many of the companies we have worked with do not have in-house expertise in the field of predictive analytics and the different challenges each application provides, it is important to thoroughly prepare and ask the right questions prior to setting out a new predictive analytics project.
How Connected is Your Machine?
If the answer is not at all, then choosing predictive analytics as your first goal is probably not the best option. Instead, it would be better to focus your efforts on ways to get the data out of the machine and visualize it. In some cases, the sensor signals already exist in a machine’s controller. Therefore, all that is needed is discovering or developing the protocols which allow the data acquisition devices to talk to the controller. For many legacy machines, no signals may be available and sensor instrumentation is required. Assessing the data connectivity of the assets you want to monitor is an important prerequisite to implementing predictive analytic solutions.
Are You Collecting the Right Data?
It is important to set success criterion and identify reasonable goals early on when working on predictive analytics applications, as predictive analytics solutions only solve problems that can be measured and detected. Although several different machine learning algorithms are available for developing solutions, sometimes the biggest challenges to overcome involve the data and not the modeling approach itself. When assessing the performance of a predictive solution, it is important to consider whether you’re collecting the right data, if that data is of an appropriate quality or whether more value could be gained by focusing on a different asset or stage of the manufacturing process.
Don’t Underestimate the Importance of Domain Expertise
In the field of predictive analytics, various techniques such as data mining, statistics, modeling, machine learning, and artificial intelligence are used to analyze data and make predictions about the future.
Within the field of machine learning, there are two main types of modeling approaches: supervised and unsupervised. Supervised learning models are different from unsupervised models in that they require a ground truth, or defined output labels for each data point. Therefore, the goal of supervised models is to approximate the relationship between the input data and their labeled outputs. Since unsupervised models do not have any defined output labels, their goal is to interpret the overall structure of the data using the input data alone.
For many industrial applications, the existence of labeled data in real world scenarios is few and far between. This constraint, coupled with the common imbalance in the availability of healthy data compared to faulty data, makes using supervised learning methods quite difficult. However, many unsupervised learning methods only require baseline data for training. Therefore, machine degradation can be detected by looking at the relationship of trending features away from their normal values (baseline).
The Predictronics team has used various unsupervised methods to detect anomalies across applications. Still, detecting anomalies is only the first step. Whenever possible fault events are identified, it is important to verify that these anomalies indeed correlate with actual system degradation. This is where domain expertise is paramount. By comparing model results with known expertise of the system, parameters can be adjusted to reduce false alarms and improve model accuracy. Therefore, the combination of predictive analytics and domain expertise cannot be underestimated.
As we reflect on our previous projects and work we have done, we are able to recognize the importance of assessing data connectivity when choosing to develop assets for health monitoring solutions. Evaluating the quality of data that is available is a critical step to ensure that realistic goals are set, since poor data collection and insufficient data can quickly sink a project.
It is important to not be discouraged by a lack of expertise in predictive analytics. Domain expertise concerning the industrial application is critical when evaluating the results of any predictive analytics solution. It is one thing to develop a model that can detect anomalies and suggest appropriate corrective actions, but unless those results agree with meaningful information out in the field, it is unlikely operators will act accordingly.
Whether you are just starting to pursue predictive analytic solutions or looking to improve existing predictive maintenance practices, Predictronics has the expertise and experience to help you reach your goals.
For more information and weekly updates, follow Predictronics: