Automated operating mode detection for a multi-modal system with multivariate time-series data

Patent metadata

Field	Value
Publication number	US-12561497-B2
Application number	US-202217686708-A
Country	US
Kind code	B2
Filing date	Mar 4, 2022
Priority date	Mar 4, 2022
Publication date	Feb 24, 2026
Grant date	Feb 24, 2026

Abstract

A system and method for learning a predictive function that can automatically learn different operating modes for a multi-modal system and predict the number of operating states for a multi-modal system and additionally the detailed structure for each state. Once learned, the predictive function (model) can be used to determine a mode of a new sample (an asset). Based on the determined components that maximize a log likelihood function, a mode of the new sample is detected into the model via dependency graphs. One aspect includes enforcing a lower bound for the number of sample points to form an operational mode for an asset. While a mode relates to sample points which maximizes like log-likelihood, an ability is provided to remove artifact modes due to noisy data by considering a sufficient sample data condition and maximizing log-likelihood. Domain knowledge can be incorporated into the model via dependency graphs.

First claim

What is claimed is:

1 . A method for evaluating a state of an asset, the method comprising:
receiving, at one or more hardware processors, plural multivariate time-series data values representing variables associated with a performance of the asset; running, using said one or more said hardware processors, a multi-modal prediction model trained to determine an operating mode of the asset based on said received plural multivariate time-series data values, said multi-modal prediction model comprising a mode fidelity mixture model (MFM) corresponding to a plurality of sparse Gaussian Mixture Models (GMM), each GMM of the plurality of spar se GMMs representing the values of the variables during the performance of the asset, said each GMM comprising a Gaussian mixture function having a weighted sum of multiple component Gaussian densities using at least a sparse mixture weights coefficient model parameter, a mean component model parameter, and a sparse inverse covariance matrix component model parameter, wherein a training of said MFM comprises selecting representative sample points in the training dataset to define operating modes, said selecting comprising:
tuning, using said one or more said hardware processors, one or more hyperparameters associated with the MFM to maximize a metric for selecting the sparse mixture weights coefficient, the mean component, and the sparse inverse covariance matrix component model parameters, and
utilizing said tuned hyperparameters to control a sparsity of said mixture weight, control a sparsity of said inverse covariance matrix component, and enforce a lower bound for a number of sample points in the training dataset to determine an operating mode of the operating modes for the asset;
comparing, using said one or more said hardware processors, the determined operating mode of the asset against known operating modes of the asset; detecting, using said one or more said hardware processors, whether or not the asset exhibits an anomalous behavior based on said comparing; and invoking, using said one or more said hardware processors, in response to detecting anomalous behavior, an investigation into a cause for the anomalous behavior of the asset.

2 . The method according to claim 1 , wherein said tuning said hyperparameters to maximize a metric for selecting said MFM model parameters further comprises:
estimating, using said one or more said hardware processors, a model complexity value representing a number of parameters to be estimated in a GMM.

3 . The method according to claim 2 , wherein said tuning said one or more hyperparameters comprises:
solving, at said one or more hardware processors, an optimization problem comprising a loglikelihood function adapted to determine the hyperparameter used to control said mixture weight sparsity, the hyperparameter used to control a sparsity for inverse covariance matrix, and the hyperparameter representing the lower bound on the number of sample points for establishing an operating mode of the asset using said MFM; and computing, at said one or more hardware processors, the model complexity value as a function of a number of non-zero mixture weights in one or more sparsified inverse covariance matrices associated with a subset of non-zero dominant mixture weights and a number of non-zero weights in the MFM model, said solving said optimization problem using said model complexity value.

4 . The method according to claim 1 , wherein said plural multivariate time-series data values are historical data obtained from said asset, said anomalous behavior comprising an outlier in the asset's behavior.

5 . The method according to claim 1 , wherein said plural multivariate time-series data values are
historical data obtained from said asset, said determined mode of operation being a mode of operation, said method further comprising:
assigning one or more of: said mode of operation to said asset, and an associated time period of when said asset operated in said operating mode.

6 . The method according to claim 1 , wherein said plural multivariate time-series data values are new sample values obtained from a current operation of said asset, said method further comprising:
computing, by said one or more said hardware processors running said MFM with said new sample values, an anomalous condition score; comparing, using said one or more hardware processors, said anomalous condition value against a threshold value; responsive to said anomalous condition value comparing, using said one or more hardware processors to detect an anomalous condition of said asset; and automatically generating, using said one or more hardware processors, a notification of said detected anomalous condition for communication to a user via an interface device.

7 . The method according to claim 1 , wherein said asset is used to build a product, said method further comprising:
obtaining, using said one or more said hardware processors, a behavior or measurement from multiple identical products produced by one or more assets used to build the products; assessing different groups of the identical products based on different behaviors or measurements; for each groups of identical products, identifying a respective asset used to build the products of each group; determining a root cause for the different behaviors or measurements of the products from each said respective group by:
detecting one or more modes of operation of each respective asset obtained by running, for each respective asset, using said one or more said hardware processors, said trained prediction model with respective historical multivariate time-series data values of that respective asset; and
comparing respective determined modes of operations of said respective assets; and
based on said compared respective determined modes of operation, adjusting a configuration of an asset based on a determined root cause reason for the different behaviors or measurements.

8 . A system for evaluating a state of an asset comprising:
a hardware processor and a non-transitory computer-readable memory coupled to the processor, wherein the memory comprises instructions which, when executed by the hardware processor, cause the hardware processor to: receive plural multivariate time-series data values representing variables associated with a performance of the asset; run a multi-modal prediction model trained to determine an operating mode of the asset based on said received plural multivariate time-series data values, said multi-modal prediction model comprising a mode fidelity mixture model (MFM) corresponding to a plurality of sparse Gaussian Mixture Models (GMM), each GMM of the plurality of spar se GMMs representing the values of the variables during the performance of the asset, said each GMM comprising a Gaussian mixture function having a weighted sum of multiple component Gaussian densities using at least a sparse mixture weights coefficient model parameter, a mean component model parameter, and a sparse inverse covariance matrix component model parameter, wherein a training of said MFM comprises selecting representative sample points in the training dataset to define operating modes, said selecting comprising:
tuning one or more hyperparameters associated with the MFM to maximize a metric for selecting the sparse mixture weights coefficient, the mean component, and the sparse inverse covariance matrix component model parameters, and
utilizing said tuned hyperparameters to control a sparsity of said mixture weight, control a sparsity of said inverse covariance matrix component, and enforce a lower bound for a number of sample points in the training dataset to determine an operating mode for the asset;
compare the determined

Assignees

• IBM

Inventors

• Phan Dzung Tien
• Baseman Robert Jeffrey
• Patel Dhavalkumar C
• Tipu Fateh A

Classifications

• G06F2111/08

  Probabilistic or stochastic CAD · CPC title

• G06F30/15

  Vehicle, aircraft or watercraft design · CPC title

• G06F30/27 Primary

  using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model · CPC title