Overcoming the obstacles to predictive maintenance

Predictive maintenance makes a lot of promises including reduced downtime and eliminating unnecessary maintenance. However it is important to keep engineering and business challenges from getting in the way. Common obstacles and arguments to implementation include:

• We do not have enough data to create a predictive maintenance system: Many predictive maintenance approaches rely on machine learning algorithms, so there needs to be enough data to create an accurate model. For predictive maintenance, this data usually originates from sensors on machinery. If the sensors are new or the way readings are logged limits the information, you will need to think about the best way to access enough data to build your models. 

Take a close look at your list of data sources – You might find that your department does not collect enough data to power a predictive maintenance system. Consider whether other departments collect data as well. Depending on where you are in the supply chain, it is  also worth looking at agreements with your suppliers or customers. Cooperating to prolong the health and efficiency of equipment components may put you in a win-win situation that fosters data access between business entities. 

Some systems operate in a feast or famine mode where data isn’t collected until a fault occurs. Others only log event codes and time stamps: engineers are notified that an event occurred, but not the sensor values at the time of the failure. Although this data may be useful for diagnostics, it is likely insufficient for developing models that can predict failures. Consider changing the data logging options to record more data, perhaps on a test fleet if production data is not available. Depending on the load on existing embedded devices, it may be possible to reconfigure them to collect and transmit sensor data, or external data loggers may be necessary to get started.

Use simulation tools to synthesise data – Generate test data using simulation tools and combine that data with what sensor data is available to build and validate predictive maintenance algorithms. This is done by creating models that cover the mechanical, electrical, or other physical system to be monitored. Synthesise sample data and validate this against measured data to ensure the model is well-calibrated. This can be done at the component level first, then later at the system level for complex systems.

When considering data for a predictive maintenance system, begin to analyse it early to understand which features are important and which may be redundant. It can be costly to keep data that is not going to be used. 

• We lack the failure data needed for accurate results: Failure data is a crucial part of teaching algorithms to recognise the warning signs to trigger just-in-time maintenance. Failure data may not exist if maintenance is performed so often that no failures have occurred, or the system is safety critical and cannot be left to fail. To stop this from becoming a fatal deficiency, you can simulate failure data and learn how to recognise warning signs.

An engineer with deep system knowledge of how the physical components work will be able to generate sample failure data with the right tools. Tools such as failure mode effects analysis (FMEA) provide useful starting points for determining which failures to simulate. An engineer with sufficient domain knowledge can incorporate these behaviours into the model in a variety of scenarios, which simulate failures by adjusting temperatures, flow rates, or vibrations or adding a sudden fault. These scenarios can then be simulated, and the resulting failure data is labeled and stored for further analysis.

While failure data might not be present, operations data might show trends about how a machine degrades over time. Statistical techniques such as principal component analysis (PCA) can provide valuable insight into how equipment operates over time, transforming raw sensor data into something which can be visualised and analysed more easily. 

• We understand the failures, but we cannot predict them: Understanding the cause of a failure is important, but there is a significant difference between identifying what went wrong and knowing how to predict it. Root cause analysis is an integral part of domain knowledge that, paired with predictive maintenance algorithms, creates an effective predictive maintenance program. If the algorithm part of the equation is a new and intimidating undertaking, you can take these steps to reduce the learning curve.

Define goals – It is important to define your goals upfront. You should then think about how the predictive maintenance algorithm will affect these goals. Building a framework that can test an algorithm and estimate its performance relative to your goals will enable faster design iterations. 

Start small – If you already know the causes behind failures, then the domain knowledge is there. Choose a project using a deeply understood system to practice on. Make sure you understand the features and factors that affect the performance of the system, and build a predictive maintenance algorithm. As the simplest starting point, consider if thresholding a feature is a significant maintenance indicator (typically done via control charts). Once you and your team are comfortable with building the algorithms for a simple problem, you can apply that knowledge to more complex systems.

Gain confidence – When predictive maintenance algorithms begin to show promising your domain knowledge to tune models to predict different outcomes based on the cost/severity of those outcomes. To further validate models, add generated failure data similar to known historical conditions and test the system. This will build confidence that the process is working.

• We do not know how to do predictive maintenance: Every new technology requires investment that must be justified. If machine learning has only recently been introduced, it is only natural to see what might be considered an advanced application of it as a risk. However, you can take steps to minimise that risk and get up and running with a working predictive maintenance model as quickly as possible. 

Instead of trying to introduce a new technology and technique, take advantage of new capabilities in the software already in place and focus on the new techniques. Some tools already have specific predictive maintenance capabilities, enabling engineers to continue working in an environment they know. 

Data can be gathered from multiple sources, such as databases, spreadsheets, or web archives. Make sure the data is in the right format including date and time stamps. Pain points are often around how to organise the data for analysis. If you don’t have enough data, you can generate this from a physical model of the machine to supplement normal usage, varying parameter values, different system dynamics, or signal faults.

If data has come from different sources, it will also need to be combined. If you remove anomalies, think about whether you want to replace them with approximate values or work with a smaller data set.
 

Instead of feeding sensor data directly into machine learning models, it is common to extract features from the sensor data. These features capture higher-level information in the sensor data, for example moving averages or frequency content. The use of familiar tools to perform feature extraction techniques simplifies this step. An iterative approach—in which features are added, new models are trained, and their performance is compared—can work well to determine the effectiveness of different features on the results.

To train the model you must classify data as healthy/faulty, set thresholds and estimate remaining useful life for components. You will need to create a list of failure scenarios to predict, choose classification methods, and simulate models. Apps provide graphical interfaces for applying machine learning that make it easy to get started and compare the results of training many different types of models.

Models may be deployed to embedded devices by converting them to a low-level language such as C, or they may be integrated with other applications in an IT environment. The pain here is often around lack of familiarity with code generation and IT integration. There are tools that can automatically package models to run in a production environment. 

Predictive maintenance is an achievable goal with the right tools, guidance, and motivation. Find the features, models, and methods that work for your business and iterate until you get it right—and remember you do not have to do it alone.