Research

US 12,291,957 B2

Inventors: Amey Ambade, Praprut Songchitruksa (1 of 2)

Time series data from wellsite pump equipment passes through a trained anomaly detection model, whose output feeds a second model that predicts the equipment's survival probability over time.

Two models in sequence rather than one classifier. The first flags anomalous behaviour in the raw sensor stream; the second consumes those flags and returns a survival function. The distinction matters operationally: a binary failure prediction tells you a pump is at risk, while a survival function tells you the probability it is still running at each point in the future, which is what a planner actually needs to schedule an intervention. This is the granted one, and the piece of work closest to the run-life estimation problem I spent the most years on.

Read US 12,291,957 B2 on Google Patents →

US 2025/0104436 A1

Inventors: Amey Ambade, Vigneshwaran Santhalingam, Tammy Lam, Dhananjaya Krishna, Velizar Vesselinov, Antonio Massoni Abinader, Aniket Ulhasrao Joshi (1 of 7)

Imagery from a wellsite is analyzed to detect movement, that movement is used to determine risk to a human present at the site, and an instruction is issued to reduce the risk.

The framing here is deliberately not object detection. Knowing there is a person and a suspended load in frame is not useful on its own. The claim runs movement to human risk to intervention, which is the line-of-fire problem: someone standing where a thing is about to move. Seven inventors, the largest team of the six, because the safety domain knowledge and the vision work came from different places.

Read US 2025/0104436 A1 on Google Patents →

US 2024/0401586 A1

Inventors: Amey Ambade, Piyush Umate, Supriya Gupta, Abhishek Sharma (1 of 4)

Two models split by sensor location: surface data drives a classifier for improper pump spacing, downhole data drives a deep learning model for conditions such as gas interference and fluid pounding.

Splitting by sensor location rather than training one model on everything is the whole idea. Surface and downhole measurements answer different diagnostic questions and have very different noise characteristics, so a single model ends up worse at both. The workflow runs either on an edge gateway at the wellsite or in a cloud service, and the stated effect is moving diagnosis from a timescale of days, which is how long a specialist review queue takes, to minutes.

Read US 2024/0401586 A1 on Google Patents →

US 2024/0018863 A1

Inventors: Amey Ambade, Sreekrishnan Ramachandran, Piyush Umate, Supriya Gupta (1 of 4)

Results returned by trained models running on field devices are assessed against real-time equipment data, a signal is raised when the model itself is underperforming, and training data is updated to produce a replacement model for deployment.

The one that is really about software rather than pumps. Every other item here is a model that predicts something; this is the machinery that notices a deployed model has gone stale and closes the loop back to retraining. Model drift is unremarkable as a concept, but doing it on IIoT hardware sitting at a remote wellsite, with intermittent connectivity and no one to babysit it, is a genuinely different problem from doing it in a datacenter.

Read US 2024/0018863 A1 on Google Patents →

US 2025/0270991 A1

Inventors: Amey Ambade, Saket Srivastava (1 of 2)

Pump system data is processed into card format, a machine learning model detects an operational condition from that card data, and the system then controls the operation of the pump in response.

A dynamometer card is a closed loop plot of load against position over a pump stroke, and its shape is what a specialist reads to diagnose the pump. Treating it as an image and classifying the shape is a natural fit. The part worth pointing at is the last clause: detection actuates control rather than raising an alarm for a human to action. That is a meaningfully higher bar, because a false positive now changes what the equipment does.

Read US 2025/0270991 A1 on Google Patents →

US 2025/0067164 A1

Inventors: Supriya Gupta, Lichi Deng, Amey Ambade, Miguel Angel Hernandez de la Bastida (3 of 4)

A computational device at the wellsite takes real-time time series data from pump equipment, runs it through a trained model to detect a performance issue, and issues a signal in response.

The earliest priority date of the six and the one I contributed to rather than led. Its significance is where the computation happens: the claim puts the model on a device at the wellsite operating on real-time data, not on a batch export analyzed somewhere else later. Everything else in this list builds on that being possible.

Read US 2025/0067164 A1 on Google Patents →