Top 10 tips for a successful predictive analytics implementation
Take the steps that move projects forward
Predictive Analytics Implementation Insights
- Predictive analytics implementation can lead to large cost-savings with less unplanned downtime when done correctly and efficiently.
- Starting small can be a great way to put predictive analytics implementation in motion with less fall out if it doesn’t work the first time.
- Proper guidance and training as well as incorporating expert knowledge are crucial.
Predictive analytics solutions, when implemented properly, can result in significant productivity improvements and cost-savings for manufacturers, with lower scrap, higher part quality and less unplanned downtime. However, many organizations struggle to implement these technologies in a systematic and well-thought-out manner, resulting in pilot projects that do not move forward and organizations not achieving their desired return on investment with these initiatives. With our past success stories and successful implementations, we wanted to share the top 10 tips for a successful predictive analytics implementation.
1. Obtain buy-in from higher management
Initiatives involving predictive analytics, such as predictive maintenance solutions and predictive quality solutions should be part of a long-term initiative, in that fully scaling these solutions is not a few months or short-term effort. Vision and buy-in from the top are needed to have a long-term goal on moving the solution from a pilot phase to a larger scale implementation across several plants. In addition, it is important to view these types of solutions as ones that can improve and grow over time, and to consider the total lifecycle of the deployed solution, from pilot to roll-out, solution support, updates and maintenance.
2. Select high-value use cases
Predictive analytic solutions can provide significant value to one’s organization, but it’s imperative to evaluate various use cases based on the business case. The business case for predictive maintenance would consider the cost of unplanned downtime if that machine or process goes down, and how often do these problems occur. For predictive quality applications, one would consider the amount of scrap for that process, whether quality issues result in line stoppages, and how much the scrap or line stoppages cost for each process/use case that is being considered. Rank the use cases based on business saving potential and use this list as a guide when making the final selection for pilot projects. In addition, consider the technical feasibility based on your internal capabilities and the past work and capabilities of the technology partners and vendors you are collaborating with.
3. Set up the pilot project for success
It is important that one can show the value and initial potential of the solution in a pilot project. For predictive maintenance, this would typically involve selecting a set of assets, in which one could potentially fail in a 6-month period, in which the solution could offer an early detection of the problem during the pilot phase. For predictive quality, scrap occurs daily but it would be advisable to have a way to quantify the scrap before the pilot and what is achieved during the pilot. Incorporate well defined success criteria, so management can easily review the results from the pilot and have clear metrics to judge whether expansion after the pilot should be considered.
4. Collect the right data at the right time and frequency
Although machine learning and predictive analytics are great, they are not magic. Having good quality data that is collected at the right time, at the appropriate sampling rate and with relevant tags for predictive maintenance or predictive quality are imperative for having a successful implementation. Typically, it is advisable to trigger the data collection under an expected set of operating conditions, so one can compare the sensor data in an “apples to apples” manner over time. In addition, one should list out the required signals and sampling rates (data model) for a given application and see if the signals/tags can be collected and at an equal or higher sampling rate then what is required. These data requirements can also factor into the use case selection since it would be part of the technical feasibility evaluation for each use case.
5. Consider a software solution designed for industrial applications of predictive analytics
With many predictive analytic solution providers and vendors, it is important to review their prior work, their history, and focus on ones with established work in the industrial /manufacturing sector. Ideally, ones that developed their solution for the industrial/manufacturing market would likely have analytic methods and dashboards that are more tailored for these applications. This would likely be favored over analytic providers that started out in other industries (health care, social media data, financial data), in that those solutions might require more customization and effort for implementing their software for an industrial/manufacturing application.
6. Do not be overly ambitious for the initial pilot projects
It is important to have early wins for any new implementation of technology, and for predictive analytics, one is advised to not be overly ambitious for a given pilot project. For predictive maintenance, it is likely most feasible to start out with early detection of machine health problems and consider diagnosis and failure prediction capabilities in a subsequent phase. For predictive quality, focusing on a few product families could be a reasonable approach, as opposed to monitoring the process for all product recipes. The early wins will lead to expansion of the solution for additional assets/processes and additional functionality, so the additional functionality can be items that are not covered in the initial pilot projects.
7. Incorporate expert domain knowledge when available and when appropriate
Machine learning is very powerful but using a solution only based on machine learning is typically not the best approach for a variety of reasons. Including domain knowledge for pre-processing, data quality checks, segmentation of the data on a cycle-by-cycle basis, and in certain cases for reviewing the select features/sensors that are used in the predictive analytics model can greatly enhance the accuracy of the solution. It is not typically feasible to have training data to learn all aspects of the analysis process (such as pre-processing) and combining machine learning models with domain knowledge is typically a more robust to consider for predictive analytics applications in industry.
8. Provide proper guidance and training to the end-users of the solution
The maintenance, quality and operations personnel that will be using this implemented software-based solution is new to this technology. Not only do they need to be convinced on how this solution will benefit their daily work and responsibilities, but they also need the guidance and training on how to properly use the solution. Hands-on training sessions and user feedback sessions is critical for ensuring the users feel empowered to use the solution as part of their daily routine and use the information to improve the maintenance, quality and operations of the processes and assets at the plant.
9. Promote the project success throughout the organization
The key members from the pilot project team need to put on their marketing and sales hats and promote the success of the pilot solution to the top management and throughout the organization. With a heightened level of exposure and awareness, the discussion of further expansion of the solution will be at the top of list and priorities. Building off the pilot success is key for ensuring that the solution is ultimately expanded, and the large-scale expansion of the solution is when organizations can reap the most significant cost savings from implementing the predictive analytics technology. For example, if one has 500 industrial robots at their factory and the predictive maintenance pilot project is monitoring 30 robots, the solution could potentially predict and prevent two failures in six months during the pilot. If this solution were expanded to 500 robots, perhaps 30 failures could have been avoided (if each failure avoidance saves the organization $100,000, $3 million in savings could be obtained in less than one year).
10. Start small, fail small, but do not be too hesitant to start
It is important to keep in mind that it is not advisable to wait for a perfect use case or solution. Even with following these tips, not every pilot project will lead to a wide-scale implementation and large success. It is important to not wait since competitive organizations are also likely working on initiatives on this topic and this technology can be viewed as a competitive advantage. The recommendation is to start small, fail small if the pilot project is not successful, but do not be too hesitant to start.
With these 10 tips in mind, we hope that a greater number of implementations of predictive analytics can lead to large cost-savings and significant value for various organizations.