Many organizations report the cost of downtime at $100,000 to $300,000 per hour or more. With the power of machine learning, big data, and condition monitoring maintenance teams can use predictive maintenance analytics to increase equipment uptime by up to 20%.
The more asset data is available the better predictions become. With advanced data science, indicators of potential failure can be detected immediately, so technicians can work on the equipment’s condition before it loses functionality.
We’ll share all the details, and show you how it’s done.
How the Data is Collected
In a predictive maintenance program, sensors are used to collect data from the selected assets. There are different types of sensors, that can be measure temperature, vibration, pressure, and more. The sensors to be used are chosen depending on the nature of the asset and installed in strategic points. Some of the sensors commonly used include:
Infrared analysis sensors. Used specifically to compare the difference in temperature between components over time. Can be used to monitor bearing temperatures in large motors, measure piping conditions, calculate process temperatures, estimate plumbing conditions, verify solar panel conditions, evaluate electrical components conditions, and more.
Vibration analysis sensors. Verifies parts vibration to detect faults related to misalignment, mechanical looseness, gear defects, lack of lubrication, resonance, rubbing, cavitation, corrosion, and more.
Ultrasonic analysis microphones. Uses a sensitive microphone to pick up high-frequency sounds and detect the need for electrical inspection, steam trap maintenance, optimal lubrification practices, leak maintenance activities, mechanical inspection, electric arc flash detection, and valve testing.
All sensors must be connected to CMMS. Through the power of the Internet of Things, the software uses real-time sensor data and predictive models to generate organized information for maintenance managers to make decisions.
This way, maintenance staff can monitor the health of the machines, act on the root cause of the issue before they become catastrophic breakdowns equipment failure.
Big Data in Predictive Maintenance Analytics
Big data refers to large and complex data sets, that grow at ever-increasing rates. Within the predictive maintenance context, it means that as sensors keep sending real-time information to the database, that grows at a constant pace.
Other than advanced methods to extract value from data, or seldom to a particular size of data set, the term big data often refers directly to the use of predictive analytics in the maintenance industry. And that’s because big data is the foundation of predictive maintenance solutions.
Two vital components are necessary to perform quality big data analytics: building a Big Data ecosystem and applying the most effective combination of predictive analytics models to the collected data.
The most substantial part of a predictive data science project is related to building a connected ecosystem of platforms that collect data from different sources.
Before any algorithms come into play, all data must be gathered, structured, aggregated, and cleaned up. This initial stage constitutes around 70% of most data science projects.
From a maintenance perspective, these datasets have been traditionally stored and analyzed independently of one another. However, with the power of the Internet of Things (IoT) and other advances in information technology, it is now possible to have a more complete picture of the asset lifecycle.
Once that initial state is done, the maintenance system can turn that data into actionable insights.
Data Requirements for Predictive Analytics
Predictive modeling is a method that uses statistics to predict outcomes, such as potential machine breakdown. Predictive models use historical data to learn patterns and predict future outcomes when they receive assets operational data.
A model’s predictive accuracy depends on three main aspects: relevancy, sufficiency, and quality of the data. It’s important to note, that “training” data is needed in order to build the history of each asset.
The new data that comes through sensors using this model should have the same features and schema as the “training” data which formed historical data.
This way, the predictions become as close to reality as they can be. Let’s look into each aspect that determines accuracy:
When generating “training” data, the sources should contain features related to the operations of the organization’s main assets. This means that the problem that managers want to avoid must be clearly stated before anything else happens. The next step is to decide which assets will be included in the PdM plan, then installing the sensors and start collecting the relevant data.
If the goal is to predict the failure of a traction system, for example, the training data has to encompass all the different components for the traction system.
We recommend that your team design prediction systems about specific components rather than larger sub-systems.
Another important aspect to consider when building the “training” data is the amount of information needed until a great predictive model is created and metrics are established.
Once you start getting the data, you might start asking yourself how many failure records or part breakage events are considered enough until the predictive analytics are considered efficient. Since every piece of equipment has different particularities, there is no one definitive answer to this query. However, there more data about specific equipment, the more accurate the predictions will be.
Keep in mind that there are failures that are rare in some types of equipment, which means that the “training” data won’t gather this event. There are different techniques used to set historical data for these cases.
Quality is perhaps the most critical element of the process of building historical data. Data is probably your most valuable intangible asset, so every data processing attribute value must be precise. Data scientists must cleanse the data before training the predictive model.
It is critical to dedicate enough time to execute data quality fundamentals into the whole project plan, maintain an audit trail as “training” data is being prepared, have an assigned individual responsible for data quality, and obtain independent quality assurance.
Condition-Based Monitoring Data Sources
There are multiple ways to collect historical data. The main predictive maintenance data sources include failure history, maintenance history, machine operating conditions, and equipment metadata. Let’s explore each one of these methods more in-depth:
As you know, the majority of pieces of equipment don’t go through unplanned downtime often. Therefore, it becomes challenging to obtain data directly from failure events.
The data needed to predict failures tend to come from a component’s normal operational pattern. When building prediction models, the algorithm uses maintenance records and parts replacement history to find failure events. With the help of some domain knowledge, anomaly detection in the training data can also be defined as failures.
Maintenance data history is a vital source to compose an asset’s historical dataset. This data includes details about spare parts used, components replaced, repair activities performed, equipment condition at the time of the repair, and more.
Every maintenance activity performed in an asset should be recorded and the algorithm considers them on degradation patterns. Failing to include this vital information in the “training” data can lead to misleading model results.
Machine operating conditions
Real-time data coming from the sensors or other sources indicate assets’ performance and operation conditions. Information about maintenance work activity performed on the machines is also an important data source. A key assumption in PdM is that a machine’s health status degrades overtime during its routine operation.
With historical data available, artificial intelligence can determine characteristics that capture this aging pattern, and any anomalies that lead to degradation.
Long-term data is required for the algorithm to learn the failure and non-failure patterns over time. Based on this information, the algorithm learns to predict how many more units of time a machine can continue to work before it needs to be replaced.
Static feature data
Different from temporal data, static data involves definite features associated with each machine’s nature — it is metadata about the equipment.
Machines’ make, model, manufactured date, the start date of service, control system location, and other technical specifications are examples of static data. Adding this type of information to the CMMS software helps the algorithm make more exact predictions.
These are a few commonly used data sources, which tend to be enough to compose effective PdM plans. However, there are other data sources that can influence failure patterns, and they should be investigated and provided by domain experts.
Leverage PdM Analytics to Streamline Decision-Making
Predictive analytics can be a game-changer for your maintenance team. But the implementation can be challenging if you don’t have the right resources to do it.
The first step is determining whether your organization needs this type of investment. If unplanned breakdowns represent large losses for your business, you probably will have a high ROI coming from a proactive asset management approach.
When problems can be predicted in advance, they become solutions — not problems. Predictive maintenance strategies allow teams to lay out a plan to action with minimal effect on production and very little cost to the operation.
Limble CMMS connects a complete Predictive Maintenance module with other crucial modules such as Work Order management, Preventive Maintenance management, Enterprise Asset Management, Spare Parts Inventory, and more.
With connected modules, information within the database becomes more powerful. That’s because the information that comes from work orders or inventory management, for example, is automatically updated on the database, and making the predictions more accurate.
If you want to hear more about Limble CMMS predictive analytics capabilities, book a demo with one of our team members.