Machine learning has the potential to transform Industrial systems via ‘Industry 4.0’ practices— which are about using data-driven insights to improve operational efficiencies of manufacturing.
In this case study we look at the experience of a client in improving their offering by applying AI/ML to an industrial-scale digital printing press.
By applying AI to a digital printing press system customers were able to achieve operational improvements in the following areas:
Proactive identification of anomalous conditions based on a holistic look at operational data
Predictive usage of supplies, such as ink and paper
Monitoring overall health of operations
The value proposition of the digital press was deemed to be high owing to the uniqueness of the product and the potential for a high ROI for the end customer.
However, while the product strategy was well thought through, the product was in a crowded category. Further consideration was needed to identify a potential new market category—eg niche down in a large market, or create a new smaller sub-segment. The client had a strong brand recognition and a sales force in adjacent markets, which would be advantageous in bringing the new offering to market. The extensive usage of digital sensors and control systems in the product made it amenable for a comprehensive data-strategy for increasing the value of the overall system. There was potential for applying industry-leading algorithms and modeling techniques to improve the operational efficiencies of the offering.
The assessment results are summarized below.
AI Business Problem and Value
Print shops are under pressure to offer faster turn around time to their clients and to improve their own profit margins by reducing wastage of supplies and paper, and optimizing the maintenance of their equipment.
The industry is facing pressure to personalize informational materials and to produce high-quality materials in small quantities and at short notice. Since modern control equipment can generate a plethora of data from which insights can be extracted to improve the operational efficiencies of the digital press to improve uptime, reduce cost, and to avoid failure and to recover quickly from failure. By leveraging the data a digital press has the potential of being vastly superior to traditional printing, which is done mostly on offset presses that are most cost effective at long print runs.
AI Use Cases
Working with the client product the following use cases were seen as important for success of the project:
Continually generating the time to failure for critical components to enable proactive maintenance before part failure to improve uptime and job turnaround time.
ML-assisted identification of anomalies in system operation for further investigation by process engineers.
Optimize spare parts inventory, and consumables such as ink, printheads, and paper by predicting usage patterns to enable just-in-time delivery and reducing the on-site spare-parts inventory.
Failed parts tracking and detection
Enable ability to look up parts version for each component as an assistance in failure triage.
Proactive firmware updates
Enable automatic firmware updates to each subsystem to ensure latest control systems innovations to be deployed quickly to the fleet.
Alerts monitoring and dashboard
Single pane of glass view of operating parameters for easy monitoring and quick identification of operational issues by line workers and process engineers.
We identified the prediction variables that would be generated by a data-driven ML system based on 3 selected use cases.
Prediction of the remaining lifetime of parts was a key prediction variable. E.g., lifetime of various drive motors, paper path rollers, print hardware.
In addition knowing when consumables such as ink, paper were expected to run out, would help the print shop optimize their inventory and to pre order consumables just in time of expected needs.
Alerts monitoring and dashboard
To aid operational efficiency it was considered desirable to produce automatic alerts of anomalous conditions, based on monitoring parameters such as tension in various parts of the paper path, temperature readings of various parts of the system and other available IIOT datasets.
In collaboration with the client’s technical team, we ran a comprehensive set of experiments on data sets collected from the prototypes of the digital press. We conducted a literature review of the Industry 4.0 best practices and identified a set of techniques that appeared promising.
After running experiments we were able to narrow down approaches that showed a lot of promise in the following areas.
Anomaly detection by looking at a holistic view of time-series data generated by various sensors and control systems.
Fault prediction by training models with labeled data from the R&D team. These predictions were validated on failures on the prototype units.
After confirming the feasibility of the ML approaches, the team built an MVP prototype to validate the benefit of the approach that was then tested by beta customers in production.
The digital sensors and control systems in the digital press are a rich source of data.
The challenge was to capture and store the relevant data sets.
Many of the IOT devices could generate millisecond granularity data, which was more useful for diagnostics rather than for failure modeling.
We designed a methodology of storing the fine-grain data in a local cache, where it could be retrieved by remote trouble shooters on-demand.
Per second data was gathered and sent to the cloud for modeling and analysis, where it was placed in archival storage after a few days.
Modeling of data was done in the cloud, but prediction was done on the edge.
Leadership team and AI readiness
We identified team gap expertise and presented an internal training and external consultancy plan.
Data pipeline, algorithms and models
Working with the system architects we laid out a comprehensive data collection, storage, and modeling strategy to make use of the data.
We conducted signal-extraction experiments to evaluate the feasibility of anomaly detection, failure prediction, and supplies optimization.
AI architecture and deployment
We recommended a hybrid approach of building predictive models in the cloud, with fleet-level data, and running predictions in the Edge with regular model refreshes pushed to the pres
Product roadmap and market category
We determined the team needed more effort to scope out new features based on ongoing customer feedback.
We identified an opportunity to build out a comprehensive category strategy and recommended the follow up steps needed.
Value proposition and competitive situation
The market was crowded and was well served by many solutions. A new niche down category or redefining a new breakout category are both possible to create a new market space to lead in.