Oracle International Corporation

Amdahl Corporation

Digital Equipment Corporation

Dieter Gawlick

Zhen Hua Liu

Kenny C Gross

Guang Chao Wang

Eric S Chan

Kenneth P Baclawski

Adel Ghoneimy

Richard Paul Sonderegger

Anna Chystiakova

Amit Ganesh

Weiwei Gong

Gary C Ngai

Aravind Yalamanchi

Nimar Singh Arora

Terry Michael Olkin

Neerja Bhatt

Alan Robert Downing

Benny Souder

Edward R Wetherbee

Ekrem Soylemez

Mahesh Subramaniam

Mengying Li

Christoph Bussler

Arno S Krakauer

Richard Byron Wilmot, Ii

Jim Stamos

John A Colgrove

Wei Wang

Tirthankar Lahiri

James William Stamos

Jagannathan Srinivasan

Bipul Sinha

Sanjay Kaluskar

Srinivas Vemuri

Vasudha Krishnaswamy

Jayanta Banerjee

Matthew T Gerdes

Atrayee Mullick

Kirk A Bradley

Hasan Rizvi

Sandeep Jain

James Rohrkemper

Hariharan Balasubramanian

Edward C Cheng

Alvin H To

Samuel DeFazio

Pankaj Tandon

Vimal Chopra

Rahim Yaseem

Sashikanth Chandra

Premal Shah

Patrick E O'Neil

Menglin Liu

Anna Morav

The disclosed embodiments relate to a system that optimizes a prognostic-surveillance system to achieve a user-selectable functional objective. During operation, the system allows a user to select a functional objective to be optimized from a set of functional objectives for the prognostic-surveillance system. Next, the system optimizes the selected functional objective by performing Monte Carlo simulations, which vary operational parameters for the prognostic-surveillance system while the prognostic-surveillance system operates on synthesized signals, to determine optimal values for the operational parameters that optimize the selected functional objective.

Optimizing a prognostic-surveillance system to achieve a user-selectable functional objective

Techniques for identifying a root cause of an operational result of a deterministic machine learning model are disclosed. A system applies a deterministic machine learning model to a set of data to generate an operational result, such as a prediction of a 'fault' or “no-fault” in the system. The set of data includes signals from multiple different data sources, such as sensors. The system applies an abductive model, generated based on the deterministic machine learning model, to the operational result. The abductive model identifies a particular set of data sources that is associated with the root cause of the operational result. The system generates a human-understandable explanation for the operational result based on the identified root cause.

Root cause analysis for deterministic machine learning model

A Lock-Free Reservation mechanism is provided. When a transaction issues an update that affects a value in a 'reservable column' of a row, the database server does not immediately obtain a lock that covers the row. Instead, the database server adds a reservation to a reservation journal. At the time the transaction commits, a lock is obtained and the requested update is made. In one implementation, before adding the reservation to the reservation journal, the database server determines whether making the update would violate any constraints involving the reservable column. In one implementation, the constraint check not only takes into account the current value of the data item that is being updated and the amount of the update, but also pre-existing reservations in the reservation journal that affect the same data item.

System and method for increased transaction throughput and improved response time under high contention

Systems, methods, and other embodiments associated with auditing the results of a machine learning model are described. In one embodiment, a method accesses original time series data and machine learning estimates of the original time series data. The method generates reconstituted time series data from the machine learning estimates by reversing operations of a machine learning model trained for generating the machine learning estimates from the original time series data. The method detects tampering (or corruption) in the original time series data based on a difference between the original time series data and reconstituted time series data. And, the method generates an electronic verification report that indicates whether the tampering (or corruption) is detected in the original time series data.

System and method for ensuring that the results of machine learning models can be audited

Techniques for geometric aging data reduction for machine learning applications are disclosed. In some embodiments, an artificial-intelligence powered system receives a first time-series dataset that tracks at least one metric value over time. The system then generates a second time-series dataset that includes a reduced version of a first portion of the time-series dataset and a non-reduced version of a second portion of the time-series dataset. The second portion of the time-series dataset may include metric values that are more recent than the first portion of the time-series dataset. The system further trains a machine learning model using the second time-series dataset that includes the reduced version of the first portion of the time-series dataset and the non-reduced version of the second portion of the time-series dataset. The trained model may be applied to reduced and/or non-reduced data to detect multivariate anomalies and/or provide other analytic insights.

Geometric aging data reduction for machine learning applications

In one embodiment, a method for auditing the results of a machine learning model includes: retrieving a set of state estimates for original time series data values from a database under audit; reversing the state estimation computation for each of the state estimates to produce reconstituted time series data values for each of the state estimates; retrieving the original time series data values from the database under audit; comparing the original time series data values pairwise with the reconstituted time series data values to determine whether the original time series and reconstituted time series match; and generating a signal that the database under audit (i) has not been modified where the original time series and reconstituted time series match, and (ii) has been modified where the original time series and reconstituted time series do not match.

We describe a system that performs prognostic-surveillance operations based on an inferential model that dynamically adapts to evolving operational characteristics of a monitored asset. During a surveillance mode, the system receives a set of time-series signals gathered from sensors in the monitored asset. Next, the system uses an inferential model to generate estimated values for the set of time-series signals, and then performs a pairwise differencing operation between actual values and the estimated values for the set of time-series signals to produce residuals. Next, the system performs a sequential probability ratio test (SPRT) on the residuals to produce SPRT alarms. When a tripping frequency of the SPRT alarms exceeds a threshold value, which is indicative of an incipient anomaly in the monitored asset, the system triggers an alert. While the prognostic-surveillance system is operating in the surveillance mode, the system incrementally updates the inferential model based on the time-series signals.

Prognostic-surveillance technique that dynamically adapts to evolving characteristics of a monitored asset

The disclosed embodiments relate to a system that automatically adapts a prognostic-surveillance system to account for aging phenomena in a monitored system. During operation, the prognostic-surveillance system is operated in a surveillance mode, wherein a trained inferential model is used to analyze time-series signals from the monitored system to detect incipient anomalies. During the surveillance mode, the system periodically calculates a reward/cost metric associated with updating the trained inferential model. When the reward/cost metric exceeds a threshold, the system swaps the trained inferential model with an updated inferential model, which is trained to account for aging phenomena in the monitored system.

Automatically adapting a prognostic-surveillance system to account for age-related changes in monitored assets

First, the system obtains time-series sensor data. Next, the system identifies missing values in the time-series sensor data, and fills in the missing values through interpolation. The system then divides the time-series sensor data into a training set and an estimation set. Next, the system trains an inferential model on the training set, and uses the inferential model to replace interpolated values in the estimation set with inferential estimates. If there exist interpolated values in the training set, the system switches the training and estimation sets. The system trains a new inferential model on the new training set, and uses the new inferential model to replace interpolated values in the new estimation set with inferential estimates. The system then switches back the training and estimation sets. Finally, the system combines the training and estimation sets to produce preprocessed time-series sensor data, wherein missing values are filled in with imputed values.

Missing value imputation technique to facilitate prognostic analysis of time-series sensor data

Embodiments of the invention provide systems and methods for managing and processing large amounts of complex and high-velocity data by capturing and extracting high-value data from low value data using big data and related technologies. Illustrative database systems described herein may collect and process data while extracting or generating high-value data. The high-value data may be handled by databases providing functions such as multi-temporality, provenance, flashback, and registered queries. In some examples, computing models and system may be implemented to combine knowledge and process management aspects with the near real-time data processing frameworks in a data-driven situation aware computing system.

Knowledge-intensive data processing system

Techniques related to an in-memory key-value store for a multi-model database are disclosed. In an embodiment, a relational database may be maintained on persistent storage. The relational database may be managed by a database server and may include a database table. The database table may be stored in a persistent format. Key-value records may be generated within volatile memory accessible to the database server by converting data in the database table to a key-value format. The key-value format may be different from and independent of the persistent format. A database statement referencing the database table may be executed based on determining whether to access one or more key-value records in the volatile memory or to access the data in the database table. In response to determining to access the one or more key-value records, the database server may access the one or more key-value records in the volatile memory.

In-memory key-value store for a multi-model database

The disclosed embodiments relate to a system that preprocesses sensor data to facilitate prognostic-surveillance operations. During operation, the system obtains training data from sensors in a monitored system during operation of the monitored system, wherein the training data comprises time-series data sampled from signals produced by the sensors. The system also obtains functional requirements for the prognostic-surveillance operations. Next, the system performs the prognostic-surveillance operations on the training data and determines whether the prognostic-surveillance operations meet the functional requirements when tested on non-training data. If the prognostic-surveillance operations do not meet the functional requirements, the system iteratively applies one or more preprocessing operations to the training data in order of increasing computational cost until the functional requirements are met.

Intelligent preprocessing of multi-dimensional time-series data

The disclosed embodiments relate to a system for analyzing telemetry data. During operation, the system obtains telemetry data gathered from sensors during operation of a monitored system. Next, the system applies a univariate model to the telemetry data to identify an operational phase for the monitored system, wherein the univariate model analyzes an individual signal in the telemetry data without reference to other signals in the telemetry data. The system then selects a phase-specific multivariate model based on the identified operational phase, wherein the phase-specific multivariate model was previously trained based on telemetry data gathered while the system was operating in the identified operational phase. Finally, the system uses the phase-specific multivariate model to monitor the telemetry data to detect incipient anomalies associated with the operation of the monitored system.

Hybrid univariate/multivariate prognostic-surveillance technique

The disclosed embodiments relate to a system that caches time-series data in a time-series database system. During operation, the system receives the time-series data, wherein the time-series data comprises a series of observations obtained from sensor readings for each signal in a set of signals. Next, the system performs a multivariate memory vectorization (MMV) operation on the time-series data, which selects a subset of observations in the time-series data that represents an underlying structure of the time-series data for individual and multivariate signals that comprise the time-series data. The system then performs a geometric compression aging (GAC) operation on the selected subset of time-series data. While subsequently processing a query involving the time-series data, the system: caches the selected subset of the time-series data in an in-memory database cache in the time-series database system; and accesses the selected subset of the time-series data from the in-memory database cache.

Multivariate memory vectorization technique to facilitate intelligent caching in time-series databases

Growing community of inventors

Palo Alto, CA, United States of America

Dieter Gawlick