What Is Predictive Maintenance?

Predictive maintenance is an approach to detecting and anticipating machine anomalies and faults that allows early insight into degradation of machines and specific machine components.

There are two basic types of predictive maintenance:

• Fault Detection — Detect faulty or anomalous behavior, such as behavior changes due to damaged bearings or broken gear teeth, using sensor signals containing information such as velocity or acceleration data. Fault detection is especially useful in the early stages of degradation before overall system performance measurably degrades.

• Remaining Useful Life (RUL) — Predict remaining useful life by using condition indicators that trend as equipment degrades, along with RUL estimation models, such as survival, similarity, and degradation models, to determine when the equipment is likely to fail without maintenance.

When you can detect or predict failures, you can plan maintenance in advance, better manage inventory, reduce downtime, and increase operational efficiency.

The following diagram shows a general representation of a predictive maintenance scenario for either fault detection or RUL prediction, depending on the application.

In these scenarios, the predictive maintenance software extracts critical features, or condition indicators, from the sensor data and determines the current health state or the remaining useful life for the machine.

The processing in this scenario can occur either in real time with streaming sensor data or offline at regular intervals with stored data.

Definitions: Features, Condition Indicators, and Health Indicator

To perform predictive maintenance, you process sensor data to distill the information into a form that can help you extract key information about hardware behavior. The toolbox uses the following terms to describe this information.

• Features — General term for processed variables that provide some level of insight into system behavior

• Condition Indicators — Specific features which have been determined to provide the most useful information about system behavior. Designing effective condition indicators involves experimenting with different features and ranking them to find the best ones. The Diagnostic Feature Designer app provides an interactive tool for performing condition indicator design.

• Health Indicator — Single fused indicator that combines multiple condition indicators into a single variable that combines the indicators in a linear weighted formula. The single health indicator is easier to monitor, and is especially useful for RUL applications. The Health Indicator Designer app provides an interactive tool for combining a set of condition indicators into a health indicator.

Designing Condition Indicators

The key to predictive maintenance is a set of condition indicators that provide insight into the state of the system. A good condition indicator is any feature that is useful for distinguishing normal from faulty operation or for predicting remaining useful life. These indicators cluster similar system status together, and set different status apart.

General condition indicators are typically standard statistical metrics for raw sensor measurements that reveal anomalous behavior at a high level, such as directly observable voltage peaks or signal energy levels.

Specialized condition indicators are designed to extract information from the raw data using knowledge of the machine design and operational modes. Such specialized indicators can isolate the behavior of specific components or assemblies, such as isolating disturbances which are synchronous with component rotation or detecting anomalies in individual battery cells.

You design these indicators by working with sets of data from similar machines in various states of health, extracting features, and ranking them to determine which ones work best to differentiate between healthy and faulty machine behavior. Diagnostic Feature Designer is a tool that allows you to interactively extract and tune features, rank features using both supervised and unsupervised learning, and export the computations for the features you select as MATLAB^® code, or as a Simulink^® block, that can be converted to C/C++ for streaming data. You can also develop models from the indicators, such as by using Classification Learner, to create models using machine learning. These models process the data from the indicators they were trained on to determine whether the combined feature data represents healthy or unhealthy behavior.

Predictive Maintenance Workflow

The general workflow for importing data, designing condition indicators and developing a predictive maintenance algorithm includes the steps in the following flowchart.

1. Acquire and organize data—Import sensor measurements from multiple machines and organize the data into a single ensemble format so that they can be processed together. If you have a model of the system, you can also simulate faults in the model and generate simulated measurements. For more information, see Data Ensembles for Condition Monitoring and Predictive Maintenance.

2. Preprocess data—Perform standard preprocessing tasks such as cleaning and labeling the data. For more information, see Data Preprocessing for Condition Monitoring and Predictive Maintenance.

3. Design condition indicators—Extract multiple features and compare their effectiveness at separating data from healthy and unhealthy machines or for tracking degradation. Feature extraction often includes conversion or transformation of the imported data first into a form, such as a spectrum or a time-synchronous average, that supports the feature. Select the best features as condition indicators for the detection or prediction algorithm. For more information, see Condition Indicators for Monitoring, Fault Detection, and Prediction.

4. Train Model—As an alternative to using the feature calculations directly in the detection or prediction algorithm, use the features to train a model, such as a Classification Learner model from Statistics and Machine Learning Toolbox™, to support the detection or prediction processing. For an example of creating a model in Classification Learner, see Analyze and Select Features for Pump Diagnostics.

5. Deploy and integrate—Implement and deploy the algorithm to the machine or machines to be monitored. Deployment options include running on embedded hardware, as a stand-alone executable, or on a web application. Predictive Maintenance Toolbox™ supports MATLAB Coder™ C/C++ Code generation for real-time operation. For more information, see Deploy Predictive Maintenance Algorithms.

6. Incorporate new data into the model—Once the initial algorithm has been implemented, continue collecting data and tune the conditions indicators or model as you learn more and as the machine ages.

For a more detailed discussion of the predictive maintenance workflow, see Designing Algorithms for Condition Monitoring and Predictive Maintenance.