How Rajesh Mattaparthi Is Using AI To Detect Hidden Faults In Standby Power Generators

Power infrastructure failures do not announce themselves. In hyperscale data centers, standby generators can sit idle for extended periods, accumulating insulation degradation that goes undetected until a critical moment. For hospitals, emergency dispatch centers, and water treatment facilities, that moment can have serious consequences. Rajesh Mattaparthi, a Principal Data Engineer with over 20 years of experience in industrial AI and machine learning, has been working to change that by applying transformer-based deep learning to the problem of partial discharge detection in large-scale standby power generators.

His published research, Transformer-Based Fault Diagnosis for Large-Scale Standby Power Generators: Partial Discharge Pattern Recognition at Hyperscale Data Center Installations, published in the Journal of Computational Analysis and Applications, presents a detailed methodology for using transformer architectures trained on multi-sensor data to identify early-stage electrical faults. The paper draws on real-world case studies from two hyperscale data center installations with combined standby generator capacity exceeding 10 megawatts per site.

The Problem with Conventional Monitoring

Standby generators present a specific monitoring challenge. Unlike primary generation equipment that runs continuously, standby units may remain offline for months at a time between maintenance tests. Electrical insulation within these machines can degrade quietly during dormant periods, producing partial discharge activity that classical diagnostic methods struggle to catch. Traditional rule-based monitoring systems and manual inspection routines tend to flag issues only after damage is already in progress.

Mattaparthi’s research frames this as a data architecture problem as much as a signal-processing one. The paper identifies six distinct partial discharge categories relevant to hyperscale standby installations, ranging from corona discharge and internal insulation faults to thermal-electrical stress patterns and noise-induced signatures. Each carries a different operational risk profile and requires a different detection approach. Handling that variety at scale demands something more flexible than conventional threshold-based alarms.

Transformer Architecture and Multi-Sensor Fusion

The core technical contribution of Mattaparthi’s work is a transformer-based deep learning framework that ingests time-series signals from multiple sensor types simultaneously. Ultra-high-frequency sensors capture electromagnetic discharge emissions. Acoustic sensors record discharge sound patterns via audio spectrograms. Temperature sensors provide ambient condition data used for compensation. These streams are fused into a unified feature representation before being passed through the transformer’s attention mechanism, which learns to weight and relate signals across time and sensor type.

The attention mechanism is particularly suited to this task because partial discharge activity is temporal and contextual. A single anomalous reading may be noise. A pattern of readings correlated with generator load levels, cooldown phases, or ambient temperature departures is a different matter entirely. By processing data within a 10-minute time window and converting sensor snapshots into “waterfall” datasets, the model captures both discrete and time-variable discharge patterns that point-in-time approaches miss.

The research also implements a one-class transformer model designed to detect anomalous patterns without requiring labeled examples of every possible fault condition. This matters in practice because rare fault types, by definition, do not appear frequently in training data. The one-class model learns a representation of normal generator behavior and scores deviations from it, enabling detection of previously unseen discharge conditions that a supervised classifier might never encounter during training.

Findings from Field Deployments

Detection accuracy improved steadily across training epochs, reaching approximately 96 percent at epoch 10. Precision, recall, and F1 scores all performed at high levels, and the ROC curve demonstrated strong separability between normal and fault-positive conditions. The temperature compensation analysis revealed a clear relationship between ambient heat and discharge activity, with the PD activity index rising substantially as ambient temperature climbed from 20 to 45 degrees Celsius. This finding underlines the importance of environmental context in condition monitoring, particularly for installations in warmer climates.

"The goal is not just to detect faults but to detect them early enough and with enough interpretability that maintenance teams can act before failure occurs," Mattaparthi notes. "A fault score that cannot be explained offers limited operational value. This is why attention head analysis and head-wise feature integration are central to the framework, not optional additions."

Interpretability as an Operational Requirement

One of the distinguishing aspects of Mattaparthi’s approach is its attention to model interpretability. The paper shows how log-likelihood anomaly scores can be decomposed across transformer decoder heads, with specific attention heads assigned to identifying partial discharge signal embeddings in the low-frequency input range. This allows maintenance teams to trace which aspects of the sensor input contributed most to a fault classification, rather than receiving an opaque probability score.

The decision intelligence layer described in the research connects model outputs to practical operator workflows. Attention maps feed into fault localization. Latent feature representations support signal strength analysis. Likelihood scores inform anomaly visualization dashboards. Each component is designed to translate complex model behavior into action items that specialists can evaluate and act on in the field.

Speaker at GatherVerseXREvolve Summit 2026

Beyond his published research, Mattaparthi recently participated as a speaker at the GatherVerseXREvolve Summit 2026, held June 23–24, 2026. He joined a roundtable panel titled "Building XR and AI Without Leaving People Behind," which examined how immersive technologies and artificial intelligence can be developed and deployed in ways that remain accessible and equitable. The panel brought together practitioners from across North America and offered a forum for discussing the societal dimensions of AI adoption alongside its technical applications.

Mattaparthi’s participation at the summit reflects a thread running through his broader work: the recognition that AI deployed in industrial and critical infrastructure settings carries real-world stakes that go beyond model performance metrics. Systems protecting hospital power supplies or 911 dispatch centers must be not only accurate but also auditable, maintainable, and legible to the teams responsible for them.

Broader Research Contributions

Mattaparthi’s work on standby generator fault detection sits within a larger body of research spanning diesel engine prognostics, combustion anomaly detection, digital twin calibration, and agentic AI architectures for industrial systems. He holds two UK design patents, two Indian patents, and has published more than ten peer-reviewed articles covering topics from NOx sensor degradation and ECM adaptive algorithms to GenAI-augmented diagnostic reasoning for field technicians. Three additional papers have been accepted to IEEE, with further submissions to Springer currently under review.

He has also presented at ICMLBDCC 2025 and ICIDEAIA 2025, with a paper at ICAMS 2026 forthcoming. Two books on industrial AI and edge computing architectures have been submitted to Barnes & Noble and Deep Science. His academic profiles are active on Google Scholar, Web of Science, ORCID, SSRN, and Academia.edu.

"Standby generators are exactly the kind of asset where conventional monitoring leaves a gap," Mattaparthi observes. "They are idle most of the time, which means degradation accumulates quietly. Transformer-based models give us a way to close that gap using the data the sensors are already producing. The challenge now is building the infrastructure to deploy these approaches consistently across large fleets at different sites and under different operating conditions."

As AI adoption in critical infrastructure continues to accelerate, work like Mattaparthi’s offers a grounded example of what applied industrial machine learning looks like at scale: less about headline capability claims, more about building systems that hold up under real operating conditions, serve the people who depend on them, and can be understood by the experts responsible for keeping them running.