The promise

Imagine if you could predict equipment failures before they happen and systematically prevent them. That's what predictive maintenance offers. Identify warning signs of potential problems and preemptively service equipment before problems occur.

IoT is a game-changer

What used to be a manual, time-intensive procedure can now be dynamic, rapid, and automated. IoT-enabled predictive maintenance solutions take advantage of streaming data from sensors and devices to quickly assess current conditions, recognize warning signs, deliver alerts and automatically trigger appropriate maintenance processes.

Benefits of predictive maintenance solutions

Get started quickly with Microsoft Azure IoT Predictive Maintenance to anticipate maintenance needs and avoid unscheduled downtime.
Click through the six steps below for a getting-started guide on how to approach predictive maintenance projects.

Establish the business
opportunity

Gain predictive
insight

Translate insight
into action

1
Identify the
target outcome
2
Inventory
data sources
3
Capture and
combine data
4
Model, test,
and interate
5
Validate model in a live
operational setting
6
Integrate into
operations
Determine target business processes to improve and desired outcomes you ultimately want to achieve.
Identify all potential sources and types of relevant data. The outcome you are seeking will influence what data is essential and what is optional.
Connect all your data to a single place and prepare it for analysis.
Identify unexpected patterns by developing predictive models using machine learning techniques. Stank-rank models to determine which model is best at forecasting the timing of unit failures.
Apply your model to live, streaming data and observe how it works in real-world conditions. Use machine learning to improve your model and ready it for full implementation.
Operationalize the model by adjusting maintenance processes, systems, and resources to act on new insights. Make ongoing improvements by gaining insights from machine learning and advanced analytics.

What you predict must be something you can take action on—otherwise, that prediction has no value. For example, predicting that a heating and cooling unit is going to fail in the next day is not useful if there is nothing you can do to prevent it.

Start by figuring out the outcome you are looking to achieve—this determines the predictive question you need to answer, and helps you measure the success of your effort.

Common predictive questions include:

  • Timing: How much time does the equipment have left until it fails?
  • Probability: What is the probability of failure in (x) number of days or weeks?
  • Cause: What is the likely cause of a given failure?
  • Risk-level ranking: What equipment has the highest risk of failure?
  • Maintenance recommendation: Given a certain error code and other conditions, what maintenance activity is most likely to solve the problem?

Include data from a variety of sources—you may be surprised about the places where key information can come from.

Start by understanding what data is available from different data sources. This can be structured or unstructured data, and may come from internal systems or external parties.

Examples of relevant data include:

  • Operating conditions—location, temperature, equipment operator, etc.
  • Failure details—timing, weather, cause, etc.
  • Repair history

Even with partial data, you can take advantage of intermediate solutions such as anomaly detection, which involves real-time monitoring to detect unusual trends and patterns. This way you can still detect anomalies while you collect specific data required to build a robust predictive model to your problem.

Lay the groundwork for a robust predictive model by pulling in data that includes both expected behavior and failure logs.

Now you’re ready to lay the groundwork for predictive analysis. This involves:

  • Connecting data from different sources into a single, consistent system.
    Since data may live in many different places, connecting it into a single, consistent system is a key step. In some cases data may need to be moved, but in many cases it’s a matter of connecting a data source to an analysis system. Since you are likely dealing with large volumes of data, it is important to use an analysis tool that can handle big data.
  • Normalizing the data.
    Normalizing data can take time but it is also critically important, especially if you are partially relying on anecdotal information from your repair teams. Normalizing data also helps to improve the accuracy and validity of your analysis.

Make your model actionable by understanding how much advance notice the maintenance team needs in order to respond to a prediction.

Start by analyzing data to identify meaningful patterns. This involves developing a set of models using a subset of the data. As you analyze and model the data, it can be helpful to have a hypothesis you are testing. This will guide your thinking about what signals to hone in on, and will give you a baseline against which to evaluate the analytical results.

Next, stank-rank the models, using the remaining data to determine which model is best at answering your predictive question. Remember that a model must be actionable in order for it to be useful, so analysis efforts should be firmly grounded in business context. For example, if your repair team needs 48 hours’ notice for maintenance request fulfillment, an actionable model is one that predict failures more than 48 hours ahead of when they will occur.

Predictive modeling helps you identify conditions that indicate future equipment problems. With this information, you can adjust processes and systems to trigger preventive actions when those conditions occur. In other words, you can translate insights from the model into operational changes, which is where you see significant business value.

Be willing to refine your approach based on the data you gather during the real-world pilot.

Monitoring connected equipment
To run an IoT-enabled predictive maintenance pilot, your equipment needs to be connected and sending the latest operational data to the appropriate systems. That live data flow is what your model analyzes to detect problem signs and trigger alerts or preventive actions—like ordering a replacement part or scheduling a technician.

Pilot planning
Start by establishing the pilot scope, including equipment, systems and locations involved, scenarios to test, conditions under which to trigger an alert or action (for example, automatic order of a replacement part), success measures, and timing.

Applying your model and refining your results
Throughout the pilot, you will continuously gather new data that will help refine acceptable ranges and may also highlight new failure signals. Don’t be afraid to adjust your approach based on what the latest operational data and analytics tell you.

Strengthen your processes and procedures to take advantage of what you learn.

Once you’ve met pilot objectives and refined the model, you’re ready for broader implementation.

This will likely involve rolling out a number of operational changes, like a revised and/or dynamic repair schedule, or changing policies to prioritize immediate repairs when certain data exceeds a specified range. Since the operational change can be far-reaching, a phased approach is recommended so that incremental benefits can be realized.

The operational improvements that can be made when rolling out a predictive maintenance approach are extensive. For example, you can:

  • Optimize what your repair crew is doing and when—adjust repair schedules and routes to reduce breakdowns and remove extra trips.
  • Alter your purchasing approach for spare parts so you don’t need to hold excess inventory—a parts order can be triggered just in time.
  • Offer predictive maintenance as a service to capture annuity revenue and maintain ongoing relationships with your customers.

These are just a few examples of how predictive maintenance enables you to increase efficiency, reduce costs, and evolve your business.