What Is Predictive Maintenance?
Predictive maintenance means catching battery problems before they turn into expensive problems. The concept gets thrown around a lot in sales pitches, but the actual work involves staring at voltage curves and arguing with your BMS vendor about why their alerts keep firing at 3 AM.
How Cells Actually Fail
Lithium-ion cells do not die suddenly. They give warnings. The trick is knowing what to look for.
A cell losing capacity will start showing lower voltage at the same state of charge compared to its neighbors. Maybe 20 millivolts at first. Easy to miss if you are only checking pack-level data. Three months later that gap is 80 millivolts and now you have a thermal management headache because one cell is working harder than the rest.
Internal resistance creep is the other big one. We pulled a module last spring from a grid storage system that had been running for four years. Cells looked fine on paper. Capacity was still at 84%. But the DC internal resistance had climbed from 1.2 milliohms to 2.8 milliohms. That module was dumping heat into its neighbors every time the system responded to a frequency regulation call.

What Data Actually Matters
Every battery management system collects voltage, current, and temperature. That part is easy. The useful stuff comes from what you do with that data over time.
Capacity fade rate tells you more than absolute capacity. A cell that dropped from 100% to 94% in year one and then to 91% in year two is behaving normally. A cell that held at 97% for eighteen months and then fell to 88% in three months has something wrong internally. Maybe a manufacturing defect finally caught up. Maybe it got cooked during a heat wave when the HVAC system was down for maintenance.
Coulombic efficiency matters for cells doing frequent cycling. Grid storage systems might see 300 cycles a year. An EV fast charging hub might push cells through 500 or 600. If a cell is returning 99.1% of what you put in instead of 99.7%, something is consuming lithium internally. Usually SEI growth. Sometimes lithium plating if the pack got charged in cold weather without proper preheating.
The Temperature Problem
I have seen more battery failures blamed on bad cells when the real cause was bad thermal design. A cabinet in direct afternoon sun. A cooling fan that failed and nobody noticed for two weeks. Intake vents blocked by warehouse dust.
Cell temperature during charging matters more than ambient temperature. A pack charging at 0.5C might run 8 or 10 degrees Celsius above ambient. That same pack at 1C could hit 20 degrees above. Now multiply that across hundreds of cycles.
We had a customer in Nevada complaining about premature capacity loss in their second year of operation. Turned out their facility faced west. Late afternoon summer sun heated the south end of their battery room to 95°F while the north end sat at 78°F. The cells at the hot end aged almost twice as fast. Predictive maintenance caught the divergence, but the damage was already done by the time anyone looked at the data seriously.
Impedance Testing
Running electrochemical impedance spectroscopy on a live system is not practical for most operators. The equipment costs real money and you need someone who knows how to interpret the results.
What most sites do instead is track DC resistance during normal operation. Measure voltage sag under load. Compare it to baseline measurements from commissioning. A cell showing 40% higher resistance than its neighbors is telling you something.
Some BMS platforms calculate resistance automatically from charge and discharge events. The accuracy varies. I have seen systems that nail it within 5% and systems that give you garbage because they are sampling at the wrong moments in the cycle.

When to Replace Modules
The 80% state of health threshold gets repeated everywhere like it came down from a mountain on stone tablets. Real replacement decisions are messier.
A grid storage system getting paid for capacity might swap modules at 82% because the revenue loss from reduced output exceeds the module cost. A backup power installation at a data center might run cells down to 70% because they only discharge a few times a year during outages.
Warranty terms drive a lot of decisions. Most cell manufacturers warrant to 70% or 80% capacity at some cycle count. If your data shows a module will drop below warranty threshold six months early, that conversation with your supplier gets a lot more interesting.
Software Platforms
The big players sell analytics packages that promise machine learning and AI-driven insights. Some of it is useful. A lot of it is trend lines with fancy graphics.
What actually helps is having historical data organized in a way that lets you compare similar systems. If you operate fifteen identical battery containers and one of them is aging faster than the others, you want to know that. If cell lot 2847 from your supplier is showing problems across multiple sites, you really want to know that.
Data storage requirements are not trivial. Cell-level logging at one-second intervals generates about 2 gigabytes per megawatt-hour per year. Most operators sample less frequently to keep storage costs down, but then you miss transient events.

Practical Limits
Predictive maintenance cannot fix bad cells. It cannot undo thermal damage. It cannot make up for a facility that was poorly designed.
What it can do is give you lead time. A module that would have failed during a summer peak demand event fails instead during a planned maintenance window in April. That difference might save you a penalty payment or an insurance claim.
The sites that get real value from predictive maintenance are the ones that actually act on the data. I have walked through facilities with beautiful dashboards showing all kinds of warnings that nobody had looked at in months. That is not predictive maintenance. That is expensive data hoarding.

