Implementing predictive maintenance without machine-learning skills - Embedded.com

Implementing predictive maintenance without machine-learning skills

A growing perception among engineers these days is that predictive maintenance is now an almost exclusive domain of artificial intelligence (AI) techniques and that they first need to learn machine learning (ML) and neural network skillsets for implementing such applications. According to Aditya Baru, senior product marketing manager at MathWorks, engineers can still deploy predictive maintenance without learning new AI and ML skillsets.

In a recent talk with EDN, Baru outlined four basics steps for implementing predictive maintenance and added that specialized tools are available for each step.


Figure 1. A basic predictive maintenance workflow comprises four basics steps. Source: MathWorks

1. Data processing

For engineers who are not data scientists or the ones who don’t have background in ML, looking at large amounts of data generated by sensors and industrial units like wind turbines, generators, pumps and motors isn’t easy. The data that engineers are dealing with is primarily raw data; it’s messy and unclean.

A jet engine or an oil pump in an exploration operation can easily create a terabyte of data every single day; now imagine looking for faulty conditions in a terabyte of data. So, what can engineers do? “Engineers can look at the data coming in large quantities, figure out if anything is changing in the raw data, identify any system degradation, and determine why the system is behaving with an abnormality,” Baru said.

For instance, in an oil exploration pump, one thing with the raw data that engineers can look at is spectral analysis for a pump that keeps spinning. So, they can identify the frequencies at which faults appear. “While engineers understand the machine already, what they have to do now is identify what works best.”


Figure 2. Engineers can detect leaks and clogs in pumps by tracking changes in motor friction. Source: MathWork

That brings us to the second basic step, condition indicators, a data reduction method.

2. Condition indicators

If an engineer has 100 samples of time-series data, he should manage to reduce it to a single number, and that single number must capture all the relevant information in those 100 samples. “The idea is that you take a huge dataset and reduce it to a smaller number of features.”

Baru mentioned a recent project in which MathWorks worked with Daimler Mercedes on an anomaly detection application that analyzes a large amount of time-series data and figures out if the manufacturing line has some anomaly. Here, MathWorks tools reduce the large amounts of data to a smaller set of features—things like patterns and time delays—to reduce the data handling by a factor of 250.


Figure 3. Engineers can extract features from raw sensor data and create condition indicators using time- and frequency-based techniques. Source: MathWork

Now that engineers are looking at a smaller number of condition indicators, they can build a predictive model based on these condition indicators.

3. Predictive model

With a much smaller dataset, which represents the entire large dataset and captures unique information, engineers can employ suitable tools to create predictive learning models without necessarily learning AI and ML skillsets.

A variety of models—such as time series models, statistical models, and probability-based models—are equally applicable to building predictive models. “There are a lot of traditional engineering techniques for building predictive models,” Baru said.


Figure 4. Predictive Maintenance Toolbox enables engineers to estimate the remaining useful life (RUL) and provide confidence intervals associated with the prediction. Source: MathWorks

Engineers can also repurpose a tool for a slightly different application. Baru mentioned Safran, an aerospace company that uses signal conditioning techniques to predict when a system might fail. The work is done in MATLAB, a programming environment for algorithm development, data analysis, visualization, and numeric computing.

4. Algorithm deployment

The fourth step is probably the most important: deploying the algorithm for a predictive model in a production environment. Engineers can deploy the algorithms in several ways. That includes a predictive model embedded locally in a machine, a small computer running locally as an on-premise server, or data streamed to the cloud services when connectivity is viable.

Predictive maintenance implemented in this four-step workflow allows engineers to deploy a maintenance service that can guarantee that a machine will stay operational 90% of the time. And tools are available for efficiently managing all these four basic steps.

>> This article was originally published on our sister site, EDN.


Related Contents:

For more Embedded, subscribe to Embedded’s weekly email newsletter.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.