We often hear what is happening in the field of condition monitoring with regards to the Industrial Internet of Things (IIoT) and other digital transformation strategies. The promised results of utilizing machine learning (ML) and artificial intelligence (AI) as forms of condition monitoring are encouraging many organizations across industries to take advantage of data science.
In this way, we want to increase the efficiency of maintenance operations and ensure the continued health of our critical assets. Like humans, computers can learn from past experiences to make informed predictions about potential future outcomes.
But is condition monitoring really that simple?
The answer is no.
Imagine telling an organization that it needs to fail a machine at least three times to identify a particular failure mode, and then learning from the data to be able to identify patterns for that particular failure mode. Then you’ll probably be taken off-site and your technology will be ridiculed. So machine learning is problematic.
machine learning
An artificial intelligence data technology that enhances the ability of data software programs to predict future outcomes, such as impending property failure, with little human intervention after the initial setup stage.
Source: Trustworthy Factory
Some would argue that we don’t need to train a model to recognize individual failure mode levels and only get notified when a particular asset exhibits data that deviates from established norms. Machine learning can do a great job with this. But so is trend data, which has been around for decades and doesn’t require additional capital investment.
So what’s the real value in creating these machine learning models?
If we end the story here, it’s not that big of a deal. But we have vast amounts of data that help and support us. In this way, a machine learning model can be trained to understand what conditions are acceptable compared to conditions that are unacceptable.
Multi-technology and process data can also be applied to this strategy to pinpoint which data or which specific sensors are generating the outliers. This can be the focus of your analytics team.
But what value is there in doing this?
Historical data indicates that most facilities have approximately 80% of assets in good health, which means that approximately 20% of assets have identifiable defects. Utilizing this process effectively reduces her nearly 80% of the data review time required by analysts.
This frees up your schedule to focus on higher-level data and more complex problems that require a combination of equipment, process and domain knowledge to solve. Doing so can increase the percentage of good equipment and reduce the number of identifiable defects.
Most engineers and analysts don’t like flipping through streams of data to find problems. Most of the time, their real joy lies in figuring out the cause of the problem. Machine learning helps maximize analyst time, further enhances maintenance and reliability efforts, and allows the program to grow with the addition of additional assets and technologies.
did you know?
“Machine learning algorithms can predict equipment failures with 92% accuracy, improving asset reliability and product quality.”
Source: ITConvergence
As mentioned earlier, algorithms can be generated that identify anomalies down to the failure mode level, but require robust domain knowledge across several disciplines, such as prioritizing mechanical, electrical, and fixed equipment. Subject matter experts should have a basic understanding of instruments and measuring devices.
This process is not for the faint of heart. Building these accurate models requires the collaboration of software engineers, data scientists, and subject matter experts in the condition monitoring domain, but the benefits are significant.
The advantages of generating algorithms are:
- Reduced downtime by expanding condition-based maintenance coverage model.
- Reduce spending by lowering the cost per monitored asset.
- Reduce spending by eliminating time-based PM.
- Increased productivity.
- Increase employee productivity.
- Improve equipment reliability.
- Detect faults and failures early by improving local models and thresholds.
- Improved quality of life for all involved.
For example, if we consider oil analysis, the algorithm must contain information and knowledge about the asset’s individual components, parts and metadata.
Additionally, mapping the source material to a specific test slate is essential, and knowledge of proper thresholds is critical to creating a proper machine learning model for lubrication analysis.
Similarly, in vibration analysis, defining a region of interest and finding patterns in waveforms and fast Fourier transforms (FFTs) is just the starting point for the team. This base level of knowledge includes an understanding of the metadata and its inherent computations associated with specific failure modes and failure reasons.
The team should also have the knowledge and basic understanding of:
- inherent failure mode.
- What data can identify a unique failure mode?
- Which sensor is the most functional and suitable for use?
- Where failure modes are identified in the data.
These are often lacking in most, if not all, off-the-shelf products available today. Removing this basic knowledge and relying solely on simple linear regression greatly increases the number of inaccurate measurements, including both false positives and false negatives. This just makes machine learning technology look bad.
The Condition Monitoring Analyst role will develop and evolve over time, but this should be viewed as a positive transition. Your involvement in creating and maintaining these machine learning applications, and your commitment to continuously updating the models, will be invaluable to your organization.
The effort to create and maintain these databases is central to any condition monitoring program, and the accuracy of any machine learning and artificial intelligence algorithm depends on the skill, persistence and knowledge of the analyst.
