Localized temporal model forecasting

Patent metadata

Field	Value
Publication number	US-11295217-B2
Application number	US-201614996154-A
Country	US
Kind code	B2
Filing date	Jan 14, 2016
Priority date	Jan 14, 2016
Publication date	Apr 5, 2022
Grant date	Apr 5, 2022

Abstract

Disclosed herein are systems, computer-readable media, and methods related to modeling on multivariate time series data overlaid with event data. In particular, some examples involve selecting one or more historical time series data arrays similar to a recent time series data array and filtering the similar historical time series data arrays based on event data. Some examples can also involve training a localized temporal forecasting model using the filtered historical time series data arrays. Some examples can include building and/or training the localized temporal forecasting model at or near a time that a forecast is needed.

First claim

The invention claimed is:

1. A computing system comprising: at least one processor; a non-transitory computer-readable medium; and program instructions stored on the non-transitory computer-readable medium that are executable by the at least one processor to cause the computing system to carry out a method that comprises:
obtaining a first multivariate time series data array representing values of a set of multiple parameters related to the physical operation of a given asset from a first time represented by at least one timestamp, wherein at least some of the values comprise measurements collected by one or more sensors at the given asset;
performing a first comparison between the first multivariate time series data array and a set of historical multivariate time series data arrays, wherein each respective historical multivariate time series data array in the set of historical multivariate time series data arrays represents respective past values of the set of multiple parameters related to the physical operation of the given asset from a respective past time represented by at least one timestamp;
based on the first comparison, identifying a subset of historical multivariate time series data arrays that are similar to the first multivariate time series data array by (i) deriving a respective measure of similarity between the first multivariate time series data array and each respective historical multivariate time series data array in the set of historical multivariate time series data arrays and (ii) identifying the historical multivariate time series data arrays in the set of historical multivariate time series data arrays that have the best respective measures of similarity;
obtaining first event data associated with the first multivariate time series data array, wherein the first event data indicates one or more event occurrences that happened at or near the first time of the first multivariate time series data array;
performing a second comparison between the first event data and respective historical event data associated with each of the subset of historical multivariate time series data arrays, wherein the respective historical event data associated with each respective historical multivariate time series data array in the subset of historical multivariate time series data arrays indicates one or more past event occurrences that happened at or near the respective past time of the respective historical multivariate time series data array;
based on the second comparison, identifying historical multivariate time series data arrays in the subset of historical multivariate time series data arrays that have associated historical event data indicating one or more past event occurrences of a same type as the one or more event occurrences indicated by the first event data and then filtering the subset of historical multivariate time series data arrays down to the identified historical multivariate time series data arrays;
training a time series forecast model by using the identified historical multivariate time series data arrays as training data for a machine learning process that is executed by the computing system to produce the trained time series forecast model, and wherein the machine learning process employs a recurrent neural network and/or support vector regression;
generating a forecast of a future value of at least one parameter of the set of multiple parameters related to the physical operation of the given asset using the trained time series forecast model; and transmitting data indicating the forecast of the future value of at least one parameter to an output system.

2. The computing system of claim 1 , wherein the program instructions that are executable by the at least one processor to cause the computing system to carry out the method comprise program instructions that are executable by the at least one processor to cause the computing system to carry out the method in response to one or more of (a) a period of time expiring, (b) receiving a request to generate a forecast, (c) detecting a temporal change, or (d) detecting a change in at least one parameter of the set of multiple parameters.

3. The computing system of claim 1 , wherein the one or more event occurrences indicated by the first event data comprise one or more event occurrences that are expected to affect the physical operation of the given asset subsequent to the first time.

4. The computing system of claim 1 , wherein one or more of the first time of the first multivariate time series data array or the respective past time of each respective historical multivariate time series data array comprises a range of time.

5. The computing system of claim 1 , wherein the given asset comprises one of a transportation machine, an industrial machine, or a utility machine.

6. The computing system of claim 1 , wherein the one or more event occurrences indicated by the first event data comprise at least one of (a) an occurrence of a particular type of anomalous condition at the given asset, (b) an occurrence of a particular type of weather event, or (c) an occurrence of a particular type of man-made event.

7. The computing system of claim 1 , wherein the first event data and the historical event data comprises event codes.

8. A non-transitory computer-readable medium having program instructions stored thereon that are executable by at least one processor of a computing system to cause the computing system to carry out a method that comprises:
obtaining a first multivariate time series data array representing values of a set of multiple parameters related to the physical operation of a given asset from a first time represented by at least one timestamp, wherein at least some of the values comprise measurements collected by one or more sensors at the given asset;
performing a first comparison between the first multivariate time series data array and a set of historical multivariate time series data arrays, wherein each respective historical multivariate time series data array in the set of historical multivariate time series data arrays represents respective past values of the set of multiple parameters related to the physical operation of the given asset that were collected from a respective past time represented by at least one timestamp;
based on the first comparison, identifying a subset of historical multivariate time series data arrays that are similar to the first multivariate time series data array by (i) deriving a respective measure of similarity between the first multivariate time series data array and each respective historical multivariate time series data array hi the set of historical multivariate time series data arrays and (ii) identifying the historical multivariate time series data arrays in the set of historical multivariate time series data arrays that have the best respective measures of similarity;
obtaining first event data associated with the first multivariate time series data array, wherein the first event data indicates one or more event occurrences that happened at or near the first time of the first multivariate time series data array;
performing a second comparison between the first event data and respective historical event data associated with each of the subset of historical multivariate time series data arrays, wherein the respective historical event data associated with each respective historical multivariate time series data array in the subset of historical multivariate time series data arrays indicates one or more past event occurrences that happened at or near the respective past time of the respective historical multivariate time series data array;
based on the second comparison, identifying historical multivariate time series data arrays in the subset of hist

Assignees

• Uptake Tech Inc

Inventors

• Herzog James

Classifications

• G06N5/045

  Explanation of inference; Explainable artificial intelligence [XAI]; Interpretable artificial intelligence · CPC title

• G06N5/04 Primary

  Inference or reasoning models · CPC title

• G06N20/00 Primary

  Machine learning · CPC title