Spontaneous battery failures continue to expose a gap between conventional testing and real-world risk. In a webinar, Advanced Battery Testing: Catching Defects in Cells and Modules with AI & ML, Emerson’s Michi Kubo and EECOMOBILITY Founder & CEO Dr. Saeid Habibi outlined how advanced testing approaches combine modular instrumentation with artificial intelligence to identify defects that traditional methods miss.

Why It Matters

Battery manufacturing operates at scale, where even rare defects can have an outsized impact. A single faulty cell can compromise an entire pack, leading to safety risks, recalls, and warranty costs. Traditional end-of-line testing methods often lack the resolution needed to detect these anomalies in high-volume production. The approaches discussed here focus on improving defect detection without slowing throughput, while also providing traceability back to individual cells.

Key Takeaways

• Conventional measurement methods, such as open-circuit voltage and resistance checks, cannot detect all defect types
• High-speed, automated test systems can combine multiple measurements within seconds
• Machine learning techniques can identify subtle anomalies using structured test data
• Feature extraction from spectral and temporal responses enables deeper defect visibility
• Data reduction and clustering methods help isolate fault signatures from large datasets
• Traceability through data “fingerprints” supports targeted recalls and root-cause analysis

From Measurement to Classification: A Three-Step Approach

Milan opened the session by framing battery testing as a three-step process: measurement, feature extraction, and classification. First, current is applied to the cell, and the voltage response is measured. Second, data is processed to calculate metrics such as impedance and statistical variation. Third, machine learning models classify results to identify anomalies.

Michi emphasized the role of modular instrumentation in the first two steps. By using hardware directly connected to a high-speed system bus, test engineers can reduce cycle time and increase channel density, which is critical when measuring large volumes of cells. Software tools enable integration, visualization, and feature extraction from these measurements, providing the structured inputs required for advanced analysis.

The key shift comes in the third step. Instead of relying solely on threshold-based checks, machine-learning-based classification allows engineers to detect patterns indicating defects, even when those patterns are not visible in conventional metrics.

Limits of Conventional Battery Testing

Dr. Saeid Habibi highlighted a central challenge: many critical defects do not appear in standard measurements. Techniques such as open-circuit voltage (OCV), alternating current internal resistance (ACIR), and electrochemical impedance spectroscopy (EIS) provide useful information, but they can miss fault conditions that lead to safety events.

One issue is variability. In high-volume production, normal variation creates a “band” in measurement data. This variation can obscure the signal from a defect, even when using frequency-based methods such as EIS. Some fault types do not appear in these measurements at all, or require long test times that are impractical in production environments.

As Saeid explained, this gap is especially important in applications such as electric vehicles, where a single defective cell can lead to thermal events. Detecting these rare defects requires a different approach that goes beyond conventional metrics.

High-Speed Testing with Deeper Data Capture

The approach presented combines conventional testing with additional high-frequency current and voltage measurements. These measurements are collected within a few seconds per cell, enabling deployment at scale in gigafactory environments.

Rather than replacing traditional tests, this method layers additional analysis on top of them. Within the same test window, engineers can capture open-circuit voltage, resistance, and impedance data while also generating more detailed spectral and temporal information.

From this data, hundreds to thousands of features can be extracted. For example, the system described generates numerous indicators from each test, each representing a different aspect of the battery’s response. These features provide the raw material for machine learning models to identify subtle deviations from normal behavior.

The result is a testing process that operates within production constraints while capturing significantly more information about each cell or module.

Machine Learning for Anomaly Detection

Handling large volumes of test data requires structured analysis. The process begins with dimensionality reduction, where the system identifies which features contain useful information about potential defects. This reduces the dataset to a manageable set of indicators without losing critical insights.

Next, unsupervised learning techniques group data into clusters. These clusters represent patterns of normal and abnormal behavior. Because defect cases are rare, clustering helps isolate these outliers without requiring prior labeling.

Once identified, clusters can be labeled through teardown analysis, where cells are physically examined to confirm defect types. This enables semi-supervised learning, allowing the system to associate specific patterns with known fault conditions and improve diagnostic capability over time.

This structured approach differs from general-purpose artificial intelligence models. As Saeid noted, the data is tightly controlled and purpose-built, which improves reliability and interpretability in a production environment.

Enabling Traceability and Process Improvement

Beyond detection, the approach supports traceability through what Saeid described as a “fingerprint” of each cell. Instead of storing all raw data, the system retains compressed metadata that uniquely represents the cell’s characteristics.

This enables manufacturers to track defects through modules, packs, and end products. In the event of a recall, this traceability allows targeted action rather than broad, fleet-wide responses.

The data also provides insight into production stability. Variability in feature distributions can reveal inconsistencies in manufacturing processes, particularly during ramp-up phases. By monitoring these patterns, engineers can identify and correct process issues earlier, improving overall yield and quality.

In this way, advanced testing supports not only defect detection but also continuous improvement in battery manufacturing operations.

Extending Testing Beyond the End of the Line

While current implementations focus on end-of-line testing, the same methods can be applied earlier in the manufacturing process. Saeid noted potential applications in formation and aging stages, where early detection of defects could reduce downstream costs.

The approach also scales from individual cells to modules and beyond. At the module level, additional measurements allow identification of defects within specific stacks, providing even greater diagnostic resolution.

This flexibility reflects a broader shift toward integrating testing throughout the battery lifecycle, rather than relying on a single checkpoint before shipment.

Conclusion

As battery production scales, the limits of conventional testing become more visible. Combining high-speed measurement with structured machine learning offers a practical path to identifying defects that would otherwise go undetected.

To learn more about how these methods are being applied in production environments, watch the full session.