How to reduce operational costs by 79% with Predictive Maintenance?

Predictive Maintenance (wind farms example)

Maintenance costs and equipment downtime have a significant impact on the profitability of energy and manufacturing companies. In light of this, no wonder that one of the biggest challenges in these industries is asset usage optimization.

Let’s consider wind farms. Wind farms are usually located in isolated areas, away from the cities. Some of them are even placed in the sea! It is often too expensive to deploy a permanent maintenance department at these locations. Thus, even when minor faults occur (e.g. power outage), technicians must travel all the way to that location to fix the failure.

That’s one of the reasons why periodic maintenance (performed at set time intervals) is very costly and time-consuming.

Lack of proper estimation of the failure probability causes technical inspections to be performed more often than needed. And even with frequent inspections, costly equipment failures still occur.

How to deal with that? The answer is Predictive Maintenance, using Microsoft Azure.

In this article, we’re going to cover the topic of Predictive Maintenance and its application in the energy industry. We’ll use the example of wind farms, but this is applicable for any other business, struggling to optimize the utilization of its assets.

Key points:

  • What is Corrective, Preventive, and Predictive Maintenance?
  • How to start a Predictive Maintenance project and what are the key steps?
  • How much can you decrease the operational cost of devices using Predictive Maintenance?

What is Predictive Maintenance? (and why it’s so much better than Corrective and Preventive Maintenance)

Traditionally, there were two approaches:

  • Corrective Maintenance – maintenance is performed when equipment fails.
  • Preventive Maintenance – maintenance is performed regularly, based on the equipment lifespan.

Corrective Maintenance is the simplest solution, which minimizes the number of inspections. However, in case of failure, this approach results in long downtime and high replacement costs. Additionally, the long wait time increases the risk that the defective part will also damage others.

Preventive Maintenance (commonly used in the energy industry) significantly reduces the number of uncontrolled failures, but it comes at a high cost of frequent inspections and the replacement of parts that might still be working.

Want more updates like this? Leave your email address to get the latest insights every two weeks to your inbox. Subscribe

What if we only inspected and replaced parts when it is really needed? Apply the just-in-time strategy for maintenance? With IoT and Artificial Intelligence, it has become possible.

To allow accurate monitoring of their condition, wind turbines have a lot of IoT sensors, detecting the energy flow, wind speed, wind direction, rotation speed, temperatures, humidity, vibrations, stresses, and lubrication pressures.

This amount of sensor data enables not only environmental conditions monitoring but also powerful analytical opportunities.

By analyzing sensor data in real-time, you can predict the probability of equipment failure and act only when this probability is significant. This is what Predictive Maintenance is all about.

Corrective vs. Preventive vs. Predictive Maintenance

Corrective vs. Preventive vs. Predictive Maintenance

Predictive Maintenance provides two major benefits:

  1. Increase in revenue due to reduced downtime (downtime only lasts as long as the repair time)
  2. Decrease in costs due to:
    • limiting the number of uncontrolled failures
    • limiting the number of technical inspections
    • extending the life cycle of the equipment.

Predictive Maintenance solution architecture (using Azure services)

Now, let’s see what the Azure-based architecture for Predictive Maintenance looks like. Again, this is the case for wind farms, but it will apply to any other device equipped with IoT sensors.

Predictive Maintenance solution architecture in Azure (click to view full-size)

Below is the explanation of each step in the process.

  1. IoT sensors installed on wind turbines send data to cloud gateway devices.
  2. Gateway devices send aggregated data to the Azure IoT Central. Optionally, they can send data to the IoT Hub. Azure IoT Central, built on top of the IoT Hub, reduces the burden and cost of developing, managing, and maintaining the solution.
  3. Azure IoT Central (or IoT Hub) continuously exports data to Events Hub for other Azure services to use.
  4. The next step is visual modeling of wind farm facilities using Azure Digital Twins. Digital Twins integrate natively with Azure Events Hub, which other applications interact with, to retrieve data from the Twins.
  5. Events Hub stores events in Azure Data Lake Storage for machine learning model training and historical data analytics.
  6. Databricks analytical platform ingests data from the Data Lake. Data scientists use this data for:
      • Data preparation
      • Data exploration
      • Model preparation
      • Model training
      • Model evaluation
  1. Trained models are deployed as an API service to Azure Kubernetes Service (AKS), using the Azure Machine Learning APIs to containerize and deploy the API.
  2. SQL Database stores historical data and prediction results for analytical reporting.
  3. Azure Stream Analytics jobs stream data from Events Hub. Optionally, we can use Databricks Structured Streaming.
  4. Stream Analytics jobs aggregate the data from Events Hub and call predictive maintenance API hosted in AKS to predict the probability of failure of every device (near real-time analytics).
  5. API Management makes the IoT data and predictions available to data users, such as web apps, mobile apps, Azure maps, and Azure Power BI. It can also make the data available to third parties. These datasets come from:
      • Stream analytics: near real-time reports/dashboards/alerts (hot patch)
      • SQL Database: analytical reports with historical data (cold patch)
  1. Web and mobile applications built with Azure App Service provide custom functionality based on data from Azure API Management.

How does the predictive algorithm work?

Below are the three examples of machine learning models worth considering when implementing Predictive Maintenance. Our experience shows that combining different models gives the best results.

Digital Twins model:

  • predicts the value of a specific parameter (e.g. temperature) based on the rest of the parameters collected by the wind turbine (humidity, wind speed, etc.).
  • observes the difference between the actual value and the predicted value
  • detects anomalies.

Chart with predictive algorithm Digital Twins

Cross Population model:

  • compares the value of the selected parameter for a given turbine to the average value for the entire farm
  • observes the difference between the actual value and the predicted value
  • detects anomalies.

chart illustrating predictive algorithm - cross population

Classification model:

  • predicts the probability of failure
  • the algorithm worth considering in the first place: XGBoost. 

classification algorithm visualized on the chart

How to start a Predictive Maintenance project and what are the key steps?

Goal definition

First, you should define the objective in business terms. We have to remember that predicting the probability of failure isn’t our real goal, but rather a component of a solution that will enable us to optimize the maintenance process.

It’s wise to consider how our solution will be used – what actions need to be taken after failure prediction. Considering the action plan will give us additional requirements (e.g., if the travel time to the wind farm is 30 minutes, we need to detect the failure at least 30 minutes earlier).

We should also think about how predictions performance will be measured. Usually, the recall performance (percentage of predicted failures) is more important than precision (percentage of correctly predicted failures).

In other words, we can accept a certain number of false positives to detect more failures. It is a good practice to perform certain business case calculations to determine the minimum performance of the algorithm (the golden mean between precision and recall) needed to reach the business objectives. This approach will give us our success criteria.

Machine learning process

Once we have a well-defined business goal and success criteria, we can follow the machine learning project process:

  • Data collection and preparation
  • Future engineering
  • Machine Learning modeling
  • Data visualization
  • Operationalization (model deployment, serving, monitoring, maintenance).

Business profits of Predictive Maintenance

So, how much can you save on it? Almost 80% of the cost.

Below are the calculations for the wind farm consisting of 100 wind turbines with the comparison of costs for Corrective, Preventive, and Predictive Maintenance.

Excel table with cost saving calculations

Cost calculations, based on our industry experience (click to view full-size)

Compared to Preventive Maintenance (often used in the energy sector, as we mentioned before), Predictive Maintenance generates 79% fewer costs due to the reduction of cost of inspections, downtime, and parts replacement.

What’s next? Can we go even further than Predictive Maintenance?

Yes, we can! With Prescriptive Maintenance. In this approach, we don’t only predict failure events, but also recommend actions to take (understanding the root cause of failure and prevention much earlier than in Predictive Maintenance). But this is much more complicated and I could cover it in another article.

If you have questions or would like to implement this approach at your organization, just let me know!

Key takeaways:

  1. In Predictive Maintenance, we predict the probability of equipment failure and take action only when this probability is significant.
  2. Azure services power the Predictive Maintenance solution, with Databricks and Azure ML at its core.
  3. To carry out a Predictive Maintenance project you should follow six major steps (goal definition, data collection and preparation, future engineering, machine learning modeling, data visualization, operationalization).
  4. Predictive algorithms should be combined to achieve the best results.
  5. Predictive Maintenance can save you almost 80% on the cost of maintenance.