US 12,461,797 B2
Apparatus and method for unpaired time series data translation
Uddeshya Upadhyay, Karnataka (IN); Rakesh Barve, Bengaluru (IN); Shashi Kant, Karnataka (IN); Sairam Bade, Suryapet (IN); Ashim Prasad, Bangalore (IN); and Shayan Ghosh, Karnataka (IN)
Assigned to Anumana, Inc., Cambridge, MA (US)
Filed by Anumana, Inc., Cambridge, MA (US)
Filed on Aug. 4, 2023, as Appl. No. 18/230,415.
Prior Publication US 2025/0045129 A1, Feb. 6, 2025
Int. Cl. G06F 9/54 (2006.01)
CPC G06F 9/541 (2013.01) 16 Claims
OG exemplary drawing
 
1. An apparatus for machine learning using unpaired time series time series translation, wherein the apparatus comprises:
at least a processor, and a memory communicatively connected to the at least a processor, the memory containing instructions configuring the at least a processor to:
receive, from at least a sensor, at least a time series of measured values, wherein the time series of measured values is recorded using at least an initial domain protocol, wherein the at least a time series of measured values comprises at least an additional data tag, and wherein the at least an additional data tag comprises descriptive data configured to support targeted time series data;
convert the at least a time series from the initial domain protocol to a target domain protocol using an unsupervised generative machine-learning process, wherein the target domain protocol comprises reducing a dimensionality of the initial domain protocol;
validate, using the computing device, the conversion, wherein validating the conversion comprises:
reverse-translating the time series data set back to the initial domain; and
comparing the translated time series data set to a confidence threshold using a second machine learning model, wherein the second machine learning model comprises a pretrained energy function adapted using data from both the initial domain and the target domain to guide inference within a reverse-time stochastic differential equation process;
generate training data using the converted at least a time series, wherein the training data comprises the at least an additional data tag;
accept at least a user feedback to address a deficient conversion of the time series, wherein the at least a processor applies a correction to a future conversion of the time series;
train a machine-learning model using the training data, wherein the machine-learning model is configured to identify at least a pattern within the at least an additional data tag; and
instantiate the machine-learning model on an electrocardiogram device.