Ballard Power Systems, based in Burnaby, BC, Canada, is a global leader in hydrogen fuel cell technology, specializing in Proton Exchange Membrane Fuel Cells (PEMFC) for heavy-duty applications such as buses, trains, and marine vessels. As part of their growth strategy, Ballard aimed to scale production volume by 488% over three years to meet increasing demand and compete effectively against battery-electric alternatives.
In scaling production volumes, Ballard encountered these key challenges:
- Lengthy Factory Acceptance Testing (FAT): Each fuel cell stack underwent rigorous FAT lasting between 2.5 to 8 hours to ensure consistent performance and reliability. These long test cycles limited production throughput, consumed significant amounts of costly hydrogen, and strained test station capacity.
- Underutilized Data Potential: Although large volumes of high-frequency time-series data (voltage, current, temperature, etc.) were generated during testing, current tools limited the ability for Ballard to extract actionable insights. Traditional methods based on averaging techniques at the end of the test cycle didn’t allow for significant early predictions of the test outcomes or other improvements.
- High Cost of Scaling: To support their planned increase in production targets, Ballard would have needed to invest in additional automated test stations, each costing approximately $1M. Physical factory space constraints further complicated expansion plans.
Ballard partnered with Acerta to extract the latent value in their testing data and avoid costly capital expenditures. Acerta developed, validated, and deployed a custom machine learning model tailored to Ballard’s specific test sequences, integrated within Acerta’s LinePulse platform. The solution included:
- Time-Series Data Ingestion: The model ingests rich, high-frequency data captured during the FAT process, including voltage, current, stack-level performance parameters, and environmental conditions.
- Early Pass/Fail Predictions: By analyzing patterns and trends within the first 15 minutes of each test, the model accurately predicts whether a stack will pass or fail the rest of the test in 46% of cases. This allows operators to either continue the test if further verification is required, or terminate it early, saving time and resources.
- Seamless Integration: The model operates within the LinePulse platform, complementing existing modules for real-time SPC monitoring, automated root cause analysis, and predictive quality, ensuring comprehensive production oversight.
- Flexibility and Scalability: Ballard’s test engineers retain full control over model thresholds, balancing prediction coverage and accuracy to fit evolving production needs.