No Unexpected Outages: AI Predicts Charging Station Failures
A U.S.-based electric vehicle charging network operator is building a sustainable competitive advantage through AI-driven predictive maintenance.
A U.S. EV charging network operator, in business for over a decade, wanted to turn reactive maintenance into a competitive edge. In a market growing this fast, reliability is what keeps drivers loyal — yet on aging hardware, success rates were sliding below 70%. Mondrian came on as lead AI partner to speed up development of a proprietary predictive maintenance platform that flags charger failures days before they happen. Within weeks of agile iteration, the model was producing reliable forecasts across a range of failure types. The payoff: a more dependable network, field teams dispatched on schedule instead of scrambling, and a software product the client can scale across its expanding footprint.
A U.S. operator bets on AI to fast-track a proprietary predictive maintenance platform
EV mobility infrastructure
Thousands of stations deployed nationwide
United States
This client is an industry original — it sold the first smart EV charger in the U.S. in 2008, back when most people doubted the market would ever take off. More than a decade later, it runs thousands of commercial stations across the country, serving everyone from corporate fleets and municipalities to airports and retail chains. As EV adoption took off, the client placed a strategic bet: reliability would be what separates winners from the rest. Building a proprietary predictive maintenance platform wasn't about keeping up with the market — it was about setting the bar.
Reactive rather than proactive interventions
Maintenance crews responded on an as-needed basis: a malfunction was discovered by a dissatisfied driver, reported, and then scheduled for an on-site visit. Each unplanned service call resulted in a direct cost and left a terminal out of service for hours, sometimes even days.
With an average annual maintenance cost of $400 per charging station, a network of thousands of stations already represents millions of dollars in operating costs. Reactive maintenance amplifies this cost: an emergency service call adds between $500 and $1,000 in labor and logistics costs for a failure that could have been detected remotely. And the situation worsens over time: as charging stations age, their success rate drops from 85% to less than 70% after three years of operation.
Three challenges made the situation complex without specialized expertise. First, failure data is scarce and extremely skewed: less than 0.5% of recorded events correspond to an actual failure. Second, the precursor signals are subtle and buried in a massive volume of time-series data. Third, the network is constantly expanding geographically, which required a solution capable of adapting—without a complete overhaul—every time a new stations came online.
Predicting terminal failures several days in advance
Mondrian has developed a machine learning model capable of predicting terminal failures several days in advance by continuously analyzing time-series data from the network. The solution identifies precursor patterns that are invisible to the human eye: voltage fluctuations, communication anomalies, and gradual component degradation. It generates actionable alerts for maintenance teams, who can schedule their visits before a failure occurs. The architecture was designed to integrate new stations without requiring a redesign, as the network grows. The entire system is encapsulated in a proprietary software product that the customer can deploy and scale independently.
Mondrian worked in an agile manner directly with the client’s infrastructure and data experts, iterating rapidly to refine the model’s performance. The initial iterations targeted the most costly types of breakdowns—those that result in the most emergency callouts and the highest driver dissatisfaction. The client’s on-the-ground expertise, accumulated over more than a decade of operations, was central to every modeling decision.
| Component | Role in the solution |
|---|---|
| Machine Learning Models for Time Series | Detect early warning signs of failure in charger data streams |
| Specialized Resampling Techniques | Addressing extreme data skew (less than 0.5% defects) |
| Extensible Modular Architecture | Enable the seamless integration of new stations into the network without requiring a technical overhaul |
Shorter response times for fault detection and thousands in avoided costs
| Indicator | Result | Business impact |
|---|---|---|
| Fault Detection Time | Responsive → Several days in advance | Scheduled maintenance, no surprises for the driver |
| Cost Avoided Per Intervention | $500 to $1,000 saved for each emergency visit avoided | Significant cumulative savings across a network of several thousand charging stations |
| Time-to-market MVP | A few weeks of agile iterations | Quick entry into the market with a competitive and scalable product |
Do you have a similar challenge in your industry?
A 30-minute call is enough to find out where AI can create value for you.
Planifier un appel