Unplanned downtime deteriorates a plant’s productivity and bottom line. To reduce these losses, we need a comprehensive predictive maintenance strategy. Predictive maintenance anticipates when a critical asset will fail so repairs can be made without disrupting production or maintenance activities.

This is easier said than done. Accurately predicting failures requires the analysis of extensive data sets, and with thousands of potential variations occurring simultaneously, it can be a difficult and time-consuming process. To increase its effectiveness and extend its early-warning capabilities, a facility needs a predictive maintenance strategy that’s supported by the right data technologies.

Machine Learning

Machine learning (ML) is an Artificial Intelligence (AI) data technology that enhances a data software program’s ability to predict future outcomes, such as impending asset failures, with little human interaction beyond the initial setup phase. Predictions are based on asset performance history with varying conditions, such as seasonal variations, startups and shutdowns.

During setup, the ML application evaluates data sets and process parameters from multiple sources to find patterns or warning signs of impending failures unnoticeable to the human eye. The ML algorithm performs this job so well that one study found it can predict equipment failures with up to 92% accuracy.

These early predictions carry powerful possibilities. Machine learning can:

• Help resolve issues quickly
• Help find the root cause of asset failure
• Improve asset availability
• Increase production output
• Reduce unscheduled downtime
• Reduce operating and maintenance costs

4 Types of Machine Learning

There are four types of machine learning:

• Supervised
• Unsupervised
• Semi-Supervised
• Reinforcement

Each type of machine learning utilizes some form of data, either labeled, unlabeled or a combination of both.

Labeled Data: Raw data accompanied by information that provides context about the data’s properties and characteristics. Labeled data tells ML exactly what it’s looking for and what data to utilize, increasing its efficiency and accuracy. However, it’s time-consuming to create and prone to human error. This can lead to a reduction in data quality or algorithm miscalculations.

Unlabeled Data: Raw data not accompanied by contextual information. Unlabeled data is affordable and easily accessible, and when applied to ML, it forces the system to discover its own classifications. Because it can see patterns that are overlooked or underutilized, the categories ML creates are often extremely accurate. However, unlabeled data forces ML to make inferences, which may not lead to the desired conclusion, making you reevaluate your chosen data and outlined parameters.

Supervised

Supervised is the most common type of machine learning and utilizes labeled data. By providing labeled data, the system understands what information will lead to the desired result. Because this ML is operating with historical data, the results produced are extremely unbiased.

During setup, the algorithm is given this data and told what the desired outcome is, such as predicting future failure events. The algorithm will learn to interpret the data and perform this procedure autonomously while continually looking for potential improvements to the system.

Supervised machine learning is useful for:

• Separating labeled data into two categories.
• Correctly selecting between two or more classification categories.
• Making predictions and connections between data with multiple independent values.
• Connecting prognoses from multiple ML models to create one accurate prediction.

Unsupervised

Unsupervised ML systems utilize unlabeled data. The system takes this data and sorts the information based on independently identified patterns and connections to arrive at the most logical inferences. These inferences are then used to find meaning in the data and arrive at the desired output: when an asset will fail.

Unsupervised machine learning is useful for:

• Clustering data together based on commonalities.
• Detecting anomalies within a data set.
• Recognizing data points that typically occur in conjunction with each other.
• Condensing the amount of data variables.

Semi-Supervised

Semi-supervised ML systems strike a balance between supervised and unsupervised approaches. The system is supplied with an initial set of labeled data, which is used to learn the connections and correlations that will be applied to all future data, which will be unlabeled. In this way, the ML system is given an initial direction but is allowed to advance independently, giving it the opportunity to discover more efficient methods of arriving at the desired output.

Semi-supervised machine learning is useful for:

• Identifying abnormalities with limited examples.
• Applying labels to large sets of unlabeled data.

Reinforcement

Reinforcement ML systems work with unlabeled data through a trial-and-error learning process. While it still has control over what steps are taken, the system is programmed with positive and negative reinforcements that encourage the system to arrive at a defined goal. In this way, it is encouraged to learn from its experiences and become more efficient as it repeatedly performs the process.

Reinforcement machine learning is useful for:

• Allocating limited resources to reach a defined goal.
• Teaching robotic machines to complete tasks in the physical world.

4 Steps for Establishing a Machine Learning System

Once you’ve selected the ML system that best fits your plant’s needs, it’s time to begin the setup process. While simple in nature, it’s critical that each step is completed at the highest caliber possible to ensure accurate data entry and output designation. Without these, the system could be misguided and produce inaccurate results.

1. Build a Model

Build a model of the asset you want to monitor. Include all process parameters and manufacturing equipment parameters.

2. Import Data

Import the asset’s historical operational data into the model. This is referred to as the “training data set,” and it should include one year of data to account for seasonal operational changes.

3. Identify Key Data

Identify what data to include and exclude from the training data set. This information that is kept will be used to develop an operational matrix that identifies how the asset should be operating at any given point.

4. Deploy the Machine Matrix

Deploy the machine matrix. This matrix will begin constantly monitoring the asset to predict when the asset might deviate from the preset parameters and begin to fail. If there is a large enough deviation identified, the program will create an alert. This alert is then evaluated by a team member.

There are three possible outcomes when an alert is raised and an evaluation is performed:

1. The alert is valid, and action should be taken to resolve the issue.
2. The alert needs additional research or operational data. The alerted parameter and all associated parameters will continue to be monitored for alert conditions.
3. The alert is a false positive. The model will need to be retrained with additional operational data and redeployed.

Conclusion

The key to a smooth operation is being able to accurately predict when an asset will fail so we can take corrective action and avoid production delays and unplanned downtime. Predictive maintenance strategies present an opportunity to accomplish this and create a manufacturing culture supported by data and proven maintenance methods.

Although effective, it can be challenging for humans to accomplish every step of a predictive maintenance strategy alone, especially when faced with limited manhours and the human error element. But, by partnering with a quality machine learning system, we can increase our efficiency and extend our early-warning capabilities, creating a stronger, leaner facility.