Predictive maintenance is the analysis of the condition and performance of critical machines in a plant to reduce instances of machine failure. In the past, highly experienced plant operators may have predicted outcomes through experience, sound of operation, or other asset behaviors. Today, this is done with a variety of software tools and analysis types.
Understanding predictive maintenance is essential if you have major assets or machines nearing end of life, or if unplanned downtime would have a high impact on your production schedules. Choosing it as a solution for your plant can lead to savings by way of more operational uptime and faster diagnoses of issues.
To get the most value out of predictive maintenance for your plant, you should understand the following process:
1. Data acquisition. For predictive maintenance, the data collected are usually time series process data such as historian data. Common tags analyzed include current, temperature, pressure, flow, vibration, etc.
2. Data cleansing. To avoid using your data ineffectively, you must find the outliers and missing values and use corrective techniques to preserve the data. This includes removing outliers, filtering out meaningless data, and correcting offset time parameters. Cleaning your data reduces problems down the road and adds value to the insights an analytics team can provide.
3. Identifying conditional indicators. This involves distinguishing between normal asset operation and various fault types. Examples include healthy motor operational parameters, seal leakages, worn bearings, blocked inlets, or a combination of faults. Methods to identify fault features include time-based and frequency analyses.
4. Training the model. After healthy operation and fault states are identified, the model is trained. This is an important step for understanding the accuracy of the fault indicators. By running multiple tests and ensuring consistency, you can accurately choose your data model type. Machine learning algorithms come in many varieties. The five families of algorithms used to build advanced models are: classification, regression, clustering, density estimation, and dimensionality reduction. The regression family is most commonly used with continuous data.
5. Deployment and integration. These steps can occur in three places: On-premises, where your local networks and data systems are stored; at the edge, where data is so high speed you need it as close to the data source as possible to reduce latency; or in the cloud, where your company already has cloud-based systems to serve a network of remote engineers who need to see the data.
6. Retraining the model. Retraining is based on live process data, as new features or faults appear over time. As new faults occur, analysts can identify the issue and seek resolution with the client and client teams.
These steps encompass the process of predictive maintenance. You will achieve more operational uptime when you can observe the trend of your assets and can see when a machine is going down. As a result, unplanned downtime is turned into planned, effective maintenance. Additionally, when you avoid suddenly shutting down machines, you extend their life by reducing wear and tear.
Fault type identification provides value by getting your machines up and working faster. When you can see the behavior, you can fix the break rather than relying on guesswork. Predictive maintenance also helps you estimate the time of failure, taking away the guessing game and giving you a solid idea of when a machine will fail so you can prepare for maintenance and shutdowns on your own time.
Ultimately, investing in predictive maintenance can pay off for many types of industrial plants with various machine types. To learn more about how predictive maintenance can benefit your plant and prevent downtime, choose an analytics team with proven experience in industrial analytics.