Top Obstacles to Overcome when Implementing Predictive Maintenance
We investigate how to create a systematic approach to predictive maintenance, ensuring there's enough data to create accurate systems. This post also explains how to identify a failure source and knowing how to predict it.
By Seth Deland, Product Marketing Manager, Data Analytics, MathWorks
For data scientists, predictive maintenance has several promising outcomes, including reducing machine downtime and avoiding unnecessary maintenance costs while adding revenue streams for equipment vendors with aftermarket services. However, engineers and scientists face challenges around process and data when incorporating predictive maintenance technology into their companies’ operations.
Below are four of the most common implementation obstacles that engineers and scientists should avoid when looking to bring predictive maintenance into their organizations.
Obstacle 1: Being Unaware of How to Do Predictive Maintenance
Working with any new technology requires justifiable investment, and predictive maintenance is no exception. Data scientists must realize the value of their investment and produce quantifiable results as quickly as possible. Software capabilities and tools such as MATLAB can help people new to predictive maintenance get up and running in an efficient manner. By taking advantage of such tools, engineering teams can quickly incorporate predictive maintenance algorithms into operations already in place.
Creating a systematic approach to predictive maintenance puts engineers in the best position to successfully build a real-time system using a predictive model. The five-step workflow (Figure 1) can offer guidance when just starting out:
- Accessing sensor data – Gather data from databases, spreadsheets, and web archives, and ensure the data is in the right format and organized.
- Preprocess data – Clean the data by removing outliers, aligning time series, and filtering out noise.
- Extract features – Capture higher-level condition indicators, such as frequency domain or time-frequency domain features, instead of feeding raw sensor data into the model.
- Train the model – Build models that classify equipment as healthy or faulty, that can detect anomalies, or that can estimate remaining useful life for components (Figure 2).
- Deploy the model – Generate code and deploy models as an application on hardware.
Figure 1. The basics of predictive maintenance workflow. © 1984–2018 The MathWorks, Inc.
Figure 2. Training predictive models that can estimate remaining useful life and provide confidence intervals associated with the prediction. © 1984–2018 The MathWorks, Inc.
Obstacle 2: Lacking Data to Create Proper Predictive Maintenance Systems
Because predictive maintenance relies on machine learning algorithms, enough data must exist to create an accurate model. This data typically stems from machine sensors. Model success depends on how data is logged: preferably, machines will include logging options that can be modified to record more data, or simulation tools can be used to combine simulated data with available sensor data to build and validate predictive maintenance algorithms.
Engineers should avoid a condition where their systems operate in a “feast or famine” mode where little or no data is collected until a fault occurs. To prevent this, companies can change the data logging options to record more data, perhaps on a test fleet if production data is not available. It is also possible to generate test data using simulation tools by creating models, as shown in Figure 3, that cover the mechanical, electrical, or other physical systems to be monitored and then validating against measured data.
Figure 3. Using Simulink to model a transmission. This model can be used to synthesize fault data. © 1984–2018 The MathWorks, Inc.
Obstacle 3: Lacking Failure Data to Achieve Accuracy
Failure data is a fundamental element of predictive maintenance. Yet this data may not exist if maintenance is performed so frequently that no failures occur. Simulation tools, such as Simulink, can help data engineers generate this necessary failure data.
Even without failure data, unsupervised machine learning techniques can be used to identify normal and faulty behavior. For example, data could be collected from several sensors on an aircraft engine. A dimensionality reduction technique such as principal component analysis (PCA) could then be used to reduce the sensor data into a low-dimensional representation for visualization and analysis. In this representation, healthy equipment data may be centered around a normal operating point, while unhealthy equipment may be seen as moving away from normal conditions.
Obstacle 4: Understanding Failures but Not Being Able to Predict Them
There’s a big difference between identifying a failure source and knowing how to predict it. That’s why engineers need to clearly define their goals—such as longer cycles and decreased downtime—and think about how a predictive maintenance algorithm affects them. They then should build a framework to test algorithms and estimate their performance, so they can get immediate feedback during design iterations. They can then use this framework to test simple models and apply their knowledge of the data to try more complex model types. They should keep things small, validate against data, and iterate until they are confident with their results.
Obstacles aside, data scientists and engineers can take solace in realizing that predictive maintenance is an achievable goal if they can locate the best balance of tools and guidance. The onus though is on engineers and data scientists to determine the features, methods, and models that work best for them—and keep iterating until they fully master these techniques.
Learn more about this topic covered in this blog post:
- Predictive Analytics with MATLAB (white paper): Download this white paper to learn the what and how of predictive analytics.
- Using Simulink to Generate Fault Data (example): This example shows how to use a Simulink model to generate fault and healthy data. It also uses a transmission system and models a gear tooth fault, a sensor drift fault and shaft wear fault.
- Models for Predicting Remaining Useful Life (example): These models provide examples of how to predict the remaining useful life from system data.