Key Features

  • Survival, similarity, and time-series models for remaining useful life (RUL) estimation
  • Time, frequency, and time-frequency domain feature extraction methods for designing condition indicators
  • Organizing sensor data imported from local files, Amazon S3™, Windows Azure® Blob Storage, and Hadoop® Distributed File System
  • Organizing simulated machine data from Simulink® models
  • Examples for developing predictive maintenance algorithms for motors, gearboxes, batteries, and other machines

Remaining Useful Life (RUL) Estimation

Estimating the remaining useful life (RUL) of a machine can help you predict its time-to-failure, optimize maintenance schedules, and manage spare inventory efficiently. The type of RUL estimation algorithm used depends on the condition indicators extracted from the data, as well as how much data is available.

Predictive Maintenance Toolbox™ provides multiple algorithms for estimating RUL. If you only have data corresponding to when your machine failed, you can use survival analysis to determine RUL. Similarity-based models can be used when you have run-to-failure data for your machines that captures how sensor measurements change from healthy to failed states. When you have data corresponding to condition indicator values over time, as well as information about critical threshold values that indicate failure for those condition indicators, you can fit linear and exponential time-series models to your data to forecast when these thresholds will be crossed.

Training predictive models that can estimate remaining useful life and provide confidence intervals associated with the prediction.

Condition Indicator Design

Condition indicators are features extracted from your data using time, frequency, and time-frequency domain methods. The values of these indicators typically change in a predictable manner as the health of the machine degrades over time with use.

You can detect the onset of faults in your machine and take remedial action by continuously monitoring these indicators for significant changes. Condition indicators are also used to estimate remaining useful life and to develop diagnostics using machine learning techniques such as classification and regression.

Predictive Maintenance Toolbox provides capabilities for designing condition indicators using both signal-based and model-based approaches. You can calculate time-frequency moments that enable you to capture time-varying dynamics, which are frequently seen when analyzing vibration data. To detect sudden changes in data collected from machines displaying nonlinear behavior or characteristics, you can compute features based on phase-space reconstructions that track changes in your system’s state over time.

The toolbox also provides a wide range of capabilities for feature extraction using statistical methods, spectral analysis, time-series modeling, modal analysis, and other methods for designing condition indicators.

Designing condition indicators using signal-based and model-based methods to monitor the health of your machinery.

Data Organization and Labeling

Predictive maintenance and condition monitoring algorithm development involves analyzing multiple datasets collected from your machines under different operating and fault conditions. This data can come from multiple sensors, and can be time-series data or discrete-event data.

Predictive Maintenance Toolbox provides data ensembles that let you manage and organize these multiple datasets. The ensemble can access files stored locally, in distributed file systems such as HDFS, and on cloud storage platforms such as Azure® Blob Storage. After linking your data files to an ensemble, you can label different components of your data as dependent or independent variables that can be used for analyzing and describing the condition of the machine. You can also transform this data to create derived data that can be added back to the ensemble.

Managing multiple files and datasets by using a data ensemble. 

Sensor data is not always available for the multiple failure modes possible in a machine. In its place, you can use simulation data that is representative of failures by creating a model of your machine and simulating faulty operating conditions. Simulink® and Simscape™ enable you to build a model of your machine that can describe its behavior in terms of its physical components and dynamics. You can represent different failure modes of the machine by modifying parameter values, injecting faults, and changing model dynamics.

Predictive Maintenance Toolbox provides an ensemble for managing this simulation data. You can run multiple simulations of your model, and use the ensemble to manage the generated signal data in MAT files. You can also organize and label these signals according to the multiple scenarios to which they correspond. The ensemble also enables you to add any derived data you may create back to the appropriate MAT files.

Managing data generated from Simulink models using simulation ensembles.

Reference Examples for Algorithm Development

Predictive Maintenance Toolbox provides a set of reference examples that help you get started with condition monitoring and predictive maintenance. The focus of these examples is on algorithm development for machines such as batteries, pumps, and gearboxes, which are frequently the focus of predictive maintenance programs.

The examples step you through the workflow of preprocessing your data, extracting features from it, building predictive models to estimate RUL or classify the root cause of failures, and validating your results. They demonstrate the usage of various machine learning, signal processing, and dynamic modeling techniques in each stage of the workflow and can serve as templates when you develop algorithms for your own machines.