Using Machine Learning to Improve Maintenance Costs in Manufacturing

Download "Using Machine Learning to Improve Maintenance Costs in Manufacturing" to Learn

• How NineTwoThree helped create a predictive maintenance system for a hydraulic system.
• The benefits of predictive vs. reactive and preventative maintenance.
• How NineTwoThree optimized data for better model accuracy.
• Why AI-mature companies outsource to experts for model refinement.
• The success of the project, with models achieving 99%+ accuracy.

Boosting AI Maturity with Outsourced Machine Learning Expertise

While many companies look to completely outsource machine learning to an expert venture studio, organizations that have achieved strong results with their own internal machine learning teams can still benefit from outsourcing.

This was the case when NineTwoThree worked with a client with a complex hydraulic system to create a predictive maintenance system. The organization already had the monitoring and data collection infrastructure in place and they turned to NineTwoThree to reach a higher level of AI maturity and create a preventative machine learning-based system with a high level of accuracy.

Benefits of Predictive Maintenance for Manufacturing

When a single component malfunction can disrupt an entire hydraulic system, detecting faults before they become larger issues is critical. The most common maintenance strategies are:

• Reactive Maintenance: Equipment is fixed when it breaks. This results in more downtime - which can be extremely costly - in an effort to reduce upkeep costs.
• Preventative Maintenance: Equipment is regularly maintained to increase its life and prevent unexpected downtime. This approach increases maintenance costs because equipment may be worked on or replaced earlier than necessary, and doesn’t take into account components breaking down unexpectedly.
• Predictive Maintenance: Uses machine learning-based models to predict when machines will require maintenance and schedule tasks before components break down.

The benefits of predictive maintenance are clear from a cost, performance, and safety perspective. In terms of improved ROI, predictive maintenance increases the lifespan of machines, ensures equipment is running more efficiently and reduces waste. Avoiding unexpected component failures keeps those parts from negatively impacting other parts of the machine and from threatening operator safety.

Our client was aware of the risks associated with unexpected downtime and the benefits of predictive maintenance. They had a complex hydraulic system with a primary working and secondary cooling-filtration circuit, with both circuits connected via the oil tank. They reached out to us to create a machine learning model which could predict the degradation of specific components, including the cooler, valve, pump, and hydraulic accumulator.

Why ML Mature Organizations Still Outsource

The data science team working on the hydraulic system had already installed condition monitoring sensors and gathered a dataset with labeled degradation levels for each separate component. They reached out to NineTwoThree to develop the machine learning model to make use of this data.

Many organizations that are able to reach a level of AI maturity and data readiness on their own will then reach out to a venture studio like NineTwoThree to take them to the next level of expertise. In our experience, some of the reasons an organization will outsource machine learning include:

• Reaching the Next Level of AI Maturity: Organizations may find they lack the AI and machine learning expertise to continue to grow toward their goals of becoming an AI-driven organization. When attempting to make operational or systemic changes to the organization - such as implementing a predictive maintenance system - it pays to employ more experienced AI and data science experts.
• Achieving Greater Model Accuracy: Thanks to AutoML, many organizations can experiment with machine learning models and see good results. However, there are limitations in the degree of accuracy achievable. Employing ML experts like NineTwoThree help an organization reach a higher level of model accuracy than achievable alone, and reach it more quickly. For example, this client had the goal of reaching 95% accuracy with their predictive model.
• Models Need to Be Retrained: There are a few reasons data science models need to be retrained, which include new sensors added to the system, new settings of the system, and changes in the elements of the system. This is a good opportunity to bring in greater expertise and make improvements.

Challenges of a Predictive Maintenance Project

Predictive maintenance projects can become extremely complex when dealing with a large amount of data, creating accurate predictions for degradation, and developing machine learning models.

Data collection and classification usually make up a large portion of any machine learning project. Luckily, our hydraulic system client provided a raw dataset based on their many physical and virtual sensors. The physical sensors measured parameters during the load cycle of the hydraulic system. The virtual sensors used heterogeneous physical sensors to combine types of data to compute a measurement and generate information to get the observed synergistic effects.

Data Analysis

With all the sensors and load cycles, we ended up with a massive total of 43,680 features. After an exploratory analysis, we also found that many of the sensors were highly correlated, which could lead to a multicollinearity problem. 

Dealing with a large number of features can create additional complexities when building predictive maintenance models. One challenge is overfitting. When dealing with real-life production data, an overfitted model is unlikely to prove as effective.

Using NineTwoThree’s expertise in building predictive models, we were able to reduce the number of features from 43,680 to a more manageable and easier-to-analyze 136.

Modeling

Once all the sensor data is organized, we worked to develop machine learning models to predict degradation levels for each component of the hydraulic system. Understanding which models to use, and which hypotheses to test, can save a lot of time and training and help achieve greater model accuracy quicker.

For this project, we use four machine learning models to predict the degradation of each of the main components - cooler condition, valve condition, internal pump leakage, and hydraulic accumulator.

Benefits of Working with a Machine Learning Partner

Our initial goal for this project was to achieve accuracy above 95%. The four machine learning models we developed achieved accuracies of 99.0% or better.

NineTwoThree’s decades of machine learning experience helped reach high levels of accuracy and to avoid costly and time-consuming challenges. We were able to quickly identify problems and workshop solutions to our clients' data challenges and efficiently develop ML models with the best chance of achieving success.

Now, when any of the models identify the inefficient performance of a component, the client’s maintenance experts receive an immediate notification of the problem. This alert provides data on exactly which component of the hydraulic system requires their attention and what the issue may be. This allows for intelligent, preventative maintenance to limit downtime, reduce maintenance costs, and extend the life of their equipment.

Whether a company already has a high degree of data maturity, or they are just becoming aware of the benefits of ML for manufacturing, working with experienced ML developers can make a huge difference. 

If you’re looking to get started with predictive maintenance, reach out to NineTwoThree. We have a strong background in manufacturing and offer machine learning consulting and software solutions to make the most of predictive maintenance. Our team even holds IoT sensor patents from our past work in manufacturing.

‍