Performing preventive maintenance at regularly scheduled intervals is costly and labor-intensive. It can result in premature part replacement, machine deterioration, and unscheduled downtime.

Predictive Maintenance Analytics allows us to push the boundaries of production efficiency and reach a goal that until now has seemed unachievable – zero unplanned downtime.

To do this, an Industry 4.0 Predictive Maintenance system needs to be able to collect and store relevant production data and analyze it in order to drive actionable business outcomes. This is the true goal of Predictive Maintenance.

**How to Improve Asset Efficiency Using Predictive Maintenance Analytics**

**Historical Data**

The first step in predictive analytics is to consolidate all the relevant historical raw data. Before setting out to capture newly created machine data, it’s important to collect and organize as much historical information as possible.

The historical data is used to define and iteratively refine the predictive analytics models to get to the most accurate predictions.

Every machine in the production line should be considered as a data source. Historical machine data captured via Data Historians should be combined with contextual data of the manufacturing process – often captured in ERP and PLM systems – such as the product being manufactured, quality results, batch numbers, and raw materials.

**Current/New Data**

Collecting a complete set of asset and contextual data on an ongoing basis is central to Predictive Maintenance systems. This data is then consolidated with the historical data and fed into the predictive maintenance analytics algorithms.

Improvements in technology have enabled the development of sensors that are reliable, non-intrusive, inexpensive, and simple to integrate into existing assets.

Some common sensor types used in industrial manufacturing include:

- Pressure sensors
- Temperature sensors
- Optical sensors
- Water quality/condition sensors (pH, conductivity, turbidity etc.)
- Chemical sensors
- Gas sensors
- Smoke sensors
- Level sensors (liquid and solid materials)
- Accelerometers
- Gyroscope sensors
- Humidity sensors

**Data Transport and Analysis**

To securely transport the data from the sensors, IoT connectivity protocols such as MQTT are used, enabling the information to be sent via a gateway device to a designated central repository – whether on-premise or on-cloud.

Organizing and aggregating the data results in a formatted dataset that can be subjected to a variety of predictive maintenance analytical models.

For predictive maintenance (PdM), there are two main types of data that are relevant – Static and Temporal.

* Static Data* refers to the technical specifications and operational attributes of a particular machine. These machine features are inherently static, but if they do have the ability to change over time, they need to have timestamps associated with them.

* Temporal Data* refers to live operational telemetry, the current condition of the machine, its work order log, usage history, and failure and maintenance history.

Once the data is captured and organized in a central repository, predictive analytics algorithms are applied to recognize downtime patterns and generate actionable insights in the form of dashboards and alerts.

**Methods of Predictive Maintenance Analytics**

There are numerous predictive maintenance analytics methods, ranging from linear regression and time series models to survival analysis and geospatial predictive modeling.

These methods can be organized into 2 main categories:

**1.** **Regression **– this set of PdM analysis techniques focuses on building a mathematical equation as a model that represents interactions between the various elements in a machine’s process. Regression models are used to calculate the RUL (Remaining Useful Life) of an asset, or in other words, how much longer an asset will remain operational before its next failure.

In the diagram below, records have been made of an asset over time. Each record (5Y, 4Y, 3Y etc.) is a time unit – *n*Y – with *n* as a multiple, resulting in the RUL values taking on a continuous value.

Using a * Linear Regression* model, the relationship between the response (dependent) variable and a group of predictor (independent) variables could be analyzed and expressed as an equation that predicts the response variable (e.g. downtime) as a linear function of all the influential elements (e.g. machine parameters and related contextual data). This model can be further optimized by minimizing the size of the residual, and testing to see that it’s distributed randomly with respect to the predictions that the model proposes.

For cases where the output (response variable) is discrete instances and not continuous, a * Discrete Choice *model can be used instead of multiple instances of the Linear Regression model mentioned above.

Another predictive maintenance analytical model that falls under the Regression category is * Logistic Regression,* which suits cases where the dependent variable is binary and we’d like to transform this data into an unbound continuous variable, estimating a common multivariate model.

To construct predictions, Likelihood-ratio and Wald tests are used to measure the statistical significance of each coefficient in the model. Finally, to assess the goodness-of-fit of the classification model being used, a “Percentage Correctly Predicted” test is used.

**2.** **Machine Learning** – Predictive maintenance analysis techniques based on Machine Learning have the ability to predict dependent variables directly, without needing to focus on the relationships between the variables. This can be very useful in scenarios where defining the dependencies between variables is extremely complex.

Machine Learning techniques for manufacturing emulate characteristics of human cognition and learn by “training” with example problems, building the ability to predict future events.

**Leveraging Predictive Maintenance Analytics**

The analytical processes mentioned above give birth to Predictive Quality Models that can help us determine the optimum time to perform maintenance, and the type of service needed before an asset has dropped below its performance threshold.

Because a PdM model is based upon actual machine behavior, equipment failures of any kind can be predicted, making this technology a game-changer for manufacturers. Repairs are more focused and secondary damage is prevented since technicians know the exact type of issue to prepare for, making unplanned downtime a non-issue.