Harnessing Machine Learning to Transform Industrial Maintenance

mechanical-engineering-2993233_1920 copy

Harnessing Machine Learning to Transform Industrial Maintenance

Authored by: Matthew Erasmus – Director of Projects and Engineering at FALX


“There is one rule for the industrialist and that is: make the best quality goods possible at the lowest cost possible, paying the highest wages possible” – Henry Ford, 1922


In the fast-paced world of industrial operations, keeping key assets running smoothly is more crucial than ever. The term “sweat the asset” rings true for many manufacturing experts, with the last few drops being the lucrative prize. Traditional techniques that we see on strategy pyramids, ranging from run-to-failure, preventative or frequency, and mileage-based methods to predictive condition-based opportunities, have shown varying levels of performance depending on their level of application and focus. But can the industry reap the rewards of newer technology?

Enter Smart predictive maintenance (PdM), an enhanced strategy powered by machine learning that’s revolutionising the field by reducing both downtime and costs.

What is Predictive Maintenance?

Predictive maintenance is a proactive strategy encompassing many techniques that forecast the failure of critical assets before they occur, allowing for timely interventions and preventing expensive equipment downtime. Examples of typical predictive maintenance techniques range from vibration analysis for rotating equipment, thermography, and partial discharge analysis for electrical systems, to analysing oil for wear and contaminants to ensure the health of lubrication systems, transformers, and wind turbines. Many specialists would argue that these techniques are well-applied throughout the industry. However, the ever-pressing need for continuous improvement raises important considerations regarding effectiveness of the implementation of this strategy: Are the insights derived from these methods utilised promptly? Do they adequately account for current process conditions?

Adding the “Smart” to Predictive Maintenance

The available technology to analyse equipment health is certainly not new, but addressing the aforementioned questions can be challenging. Smart PdM can be applied to assets by eliminating the human element. Instead of relying on routines for periodic readings or samples, followed by analysis and waiting for results, the advent of IoT (Internet of Things) has opened doors to permanently installing data-gathering tools. These tools can autonomously acquire information in real-time. Combining this data with process data allows for extracting remarkable insights that can drive smarter decisions and prolong asset health, thanks to the abilities of machine learning.

Machine Learning: The Brains Behind Smart PdM

Machine learning (ML), a branch of artificial intelligence (AI), is crucial for making Smart PdM work. It can incorporate nuances, detect patterns, and learn from vast amounts of data, enabling it to determine if failure is imminent or if an unknown outlying process condition requires further investigation. Here’s how it unfolds in industrial settings:

Data Collection and Integration

In todays world of IoT and the multitude of available devices, key assets and machinery can be equipped with sensors capable of monitoring various data sources, from temperature to vibrations and acceleration. These systems also integrate process information from the machine’s SCADA system and can place compute power right at the coalface. Data can also be collected through image recognition using permanently installed camera systems and thermography, or three-dimensional point cloud data can be utilized with LiDAR technologies. Regardless of the data source, it is continuously collected, transformed, and analysed, forming the backbone of smart predictive modelling.

Anomaly Detection

The ingestion of both process and asset data expands the dimensions available for machine learning models to learn what normal operation looks like. Any deviation from this norm is automatically flagged as an anomaly for investigation.

Failure Prediction

By examining past data on previous failure modes, these models can identify conditions that typically lead to failures. With this insight, they can highlight when a particular asset is displaying signs of a similar failure occurring again, enabling timely intervention.

Process and Maintenance – Joined at Last

In many industries, the intersection between these two spheres on the Venn diagram is not as pronounced as it should be. By merging the two databases, ML models can identify important interconnected relationships. Industry experts understand that process health often mirrors machine health, and vice versa, so why not have an ML model oversee it all? Gone are the days when traditional techniques questioned whether a flagged deviation was due to different process conditions that weren’t captured. Incorporating process variables and conditions makes asset-based decision-making more intentional, precise, and inclusive.

The Benefits of a Smart Predictive Maintenance Approach

Adopting machine learning in predictive maintenance has clear advantages:

  1. Certainty: Enhance the certainty of your data-driven decisions by leveraging Smart PdM, which trains on past information to establish benchmarks for future flagging of issues.
  2. Consistency: Online and ever-present IoT devices eliminate the need for human intervention to a large extent, reducing the necessity for manual data capturing, analysis, and reporting, as these processes become autonomous.
  3. Continuity: Enhance the continuity of your assets and processes by identifying anomalies and predicting failures before they occur. This allows for informed decisions regarding intervention, which may sometimes involve opting not to remove an asset for repair based on the available information.
  4. Cost-effectiveness: Implementing Smart PdM can lead to significant cost savings by reducing unplanned downtime and preventing costly equipment failures. Additionally, predictive maintenance enables optimal resource allocation, helping to minimise operational expenses over time.

Challenges to Consider

While the benefits are significant, Smart PdM does have some hurdles that need to be considered:

  1. Capex: Setting up such systems can involve significant capital expenditure. These systems require upfront investment in necessary sensors, data handling infrastructure, and ML model building and deployment.
  2. Complexity: Alongside the capital expenditure, Smart PdM systems can be complex to establish. They require investment not only in hardware and software but also in the expertise needed to set up and maintain them effectively.
  3. Compute Power: ML systems require a fair amount of compute power, more than traditional systems, especially when systems are aimed at incorporating detail from multiple assets and the process.
  4. Change Management: Successfully implementing Smart PdM systems requires effective change management processes. This includes ensuring strong buy-in from skilled tradespersons and maintenance technicians, as well as managing potential resistance to change within the organisation.

In Conclusion

The leap into machine learning-enhanced predictive maintenance marks a significant evolution in industrial operations. Linking the process and maintenance functions through this automated approach has the potential to reveal interconnected problems that could have gone undetected before resulting in significant failures or consistent underperformance. By facilitating timely and necessary maintenance, it not only saves money but also boosts operational efficiency and safety. As the industrial world continues to evolve, the move towards Smart PdM and the incorporation of AI and ML is not just a wise choice but a necessity for which first movers will gain strategic advantage.