EP-4738038-A1 - SYSTEMS AND METHODS TO PREDICT ASSET STATUS

EP4738038A1EP 4738038 A1EP4738038 A1EP 4738038A1EP-4738038-A1

Abstract

Disclosed here are methods and systems to determine asset status in an industrial plant. Embodiments of the method may include, during operation of an asset positioned at the industrial plant and, iteratively, at each of a plurality of selected time intervals, receiving a plurality of data streams from sensors or devices of the asset. The method may include determining, based on application of each of the plurality of pre-processed data streams received to each of a plurality of trained machine learning models, anomaly scores. The method may include determining, via a plurality of votes based on the anomaly scores, a status for the sensors and/or devices. The method may include, if the status for the sensors and/or devices for the asset indicates an anomaly for a consecutive two or more of the plurality of time intervals, generating a next action notification to indicate the anomaly and prevent unplanned downtime.

Inventors

Eskandarani, Mohammad
Alabas, Yasir
Abahussain, Noura Adel
Alqahtani, Maram Buhayshan

Assignees

SABIC Global Technologies B.V.

Dates

Publication Date: 20260506
Application Date: 20241104

Claims (15)

A method for determining asset status in an industrial plant at a selected time to prevent unplanned downtime, the method comprising: during operation of an asset positioned at the industrial plant and, iteratively, at each of a plurality of selected time intervals: receiving a plurality of data streams from one or more of (i) sensors and/or (ii) devices that correspond to the asset, each of the plurality of data streams to correspond to one of a plurality of components of the asset or one of a plurality of processes associated with the asset; pre-processing each of the plurality of data streams to form a plurality of pre-processed data streams; determining, based on application of each of the plurality of pre-processed data streams received to each of three or more trained machine learning models, (a) a plurality of anomaly scores and (b) a type of anomaly; determining, via a plurality of votes based on each of the plurality of anomaly scores and corresponding type of anomaly, a status for each of (i) the sensors and/or (ii) devices that correspond to one or more of the plurality of data streams; determining a status for the asset based on the status for each of the (i) sensors and/or (ii) devices; and if the status for the asset indicates an anomaly for a consecutive two or more of the plurality of selected time intervals, generating a next action notification to indicate the anomaly and prevent unplanned downtime.
The method of claim 1, wherein each of the plurality of votes correspond to one of the three or more trained machine learning models, and wherein each of the plurality of votes includes a weight based on an accuracy of a corresponding model.
The method of claim 1, wherein the next action notification further includes one of a predicted asset failure or predicted maintenance of the asset.
The method of claim 1, wherein pre-processing the plurality of data streams includes: determining if any one of the plurality of data streams includes missing data points or frozen data points; and in response to a determination that any one of the plurality of data streams includes the missing data points or the frozen data points: preventing application of one of the plurality of data streams that includes the missing data points or the frozen data points, and generating an error notification including indication of an error associated with one of the (i) sensors and/or (ii) devices.
The method of claim 1, wherein each of the three or more trained machine learning models comprise one of a clustering model, a Gaussian mixture model, a time series trend model, an auto encoder model, or a support vector machine.
The method of claim 1, wherein training each of three or more trained machine learning models comprises: determining an amount of missing data points and an amount of frozen data points in each data stream of a plurality of historical data streams that correspond to operation of the plurality of components of the asset and the plurality of processes related to the asset during normal operation; in response to a determination that the amount of missing data points and the frozen data points exceed a second selected threshold, removing a corresponding data stream; in response to determination that the amount of missing data points and the frozen data points do not exceed the second selected threshold, replacing the missing data points and the frozen data points from each of the plurality of historical data streams based on a linear interpolation algorithm; normalizing each of the plurality of historical data streams to form a plurality of normalized historical data streams; removing data points corresponding to an offline state and a start-up state from the plurality of normalized historical data streams to form a plurality of pre-processed historical data streams; extracting normal ranges from the plurality of pre-processed historical data streams; determining deviation from operating conditions for each two or more data streams of the plurality of pre-processed historical data streams that correspond to redundant components; removing one of the two or more data streams based on a higher determined standard deviation for the one of the two or more data streams; and training each of or more machine learning models with the pre-processed plurality of historical data streams to form each of the three or more trained machine learning models.
The method of claim 1, further comprising: determining one or more influencing features based on the anomaly score associated with each of the (i) sensors and/or (ii) devices, and wherein the one or more influencing features are included in the next action notification.
A system for determining asset status in an industrial plant at a selected time to prevent unplanned downtime, the system comprising: a communications circuitry configured to: obtain, during operation of an asset positioned at the industrial plant and at each of a plurality of selected time intervals, a plurality of data streams from one or more of (i) sensors and/or (ii) devices associated with the asset; a modeling circuitry configured to: determine (a) a plurality of anomaly scores and (b) a type of anomaly for each of the plurality of data streams based on application of each of the plurality of data streams to each of a plurality of trained machine learning models; and an asset status circuitry configured to: determine a vote associated with each of the plurality of trained machine learning models for each of the plurality of data streams based on a corresponding anomaly score of the plurality of anomaly scores and the type of anomaly, determine a status for the asset based on votes associated with each of the plurality of trained machine learning models, and if the status for the asset indicates an anomaly for a consecutive two or more of the plurality of selected time intervals, generate a next action notification to indicate the anomaly and prevent unplanned downtime.
The system of claim 8, further comprising a pre-processing circuitry configured to: determine if any one of the plurality of data streams includes missing data points or frozen data points; and in response to a determination that any one of the plurality of data streams includes the missing data points or the frozen data points, generate an error notification including indication of an error associated with one of the (i) sensors and/or (ii) devices.
The system of claim 8, wherein the vote associated with each of the plurality of trained machine learning models for each of the plurality of data streams is further based on a type of trained machine learning model.
The system of claim 8, wherein the asset status circuitry is further configured to determine, based on (a) the plurality of anomaly scores and (b) the type of anomaly for each of the plurality of data streams and for each of the plurality of anomaly scores, an influencing feature, and wherein the next action notification includes a selected amount of indicated anomalies and corresponding influencing features.
The system of claim 8, wherein the vote associated with each of the plurality of trained machine learning models for each of the plurality of data streams is further based on a weight associated with a corresponding trained machine learning model.
The system of claim 12, wherein the weight associated with the corresponding trained machine learning model is based on a type of the corresponding trained machine learning model.
A computing device for determining asset status in an industrial plant at a selected time to prevent unplanned downtime, the computing device comprising: a plurality of inputs/outputs each in signal communication with an asset at an industrial plant, the computing device including a processor and a non-transitory machine readable storage medium storing instructions configured to, when executed by the processor: obtain, during operation of an asset positioned at the industrial plant and at each of a plurality of selected time intervals, a plurality of data streams from one or more of (i) sensors and/or (ii) devices associated with the asset; determine a plurality of anomaly scores for each of the plurality of data streams based on application of each of the plurality of data streams to each of a plurality of trained machine learning models; determine a vote associated with each of the plurality of trained machine learning models for each of the plurality of data streams based on a corresponding anomaly score of the plurality of anomaly scores; determine a status for the asset based on votes associated with each of the plurality of trained machine learning models; and if the status for the asset indicates an anomaly for a consecutive two or more of the plurality of time intervals, generate a next action notification to indicate the anomaly and prevent unplanned downtime.
The computing device of claim 14, wherein each vote comprises a weighted vote, and wherein a weight associated with the weighted vote is based on a type of machine learning model corresponding to the vote, the asset corresponding to a data stream applied to a corresponding machine learning model, one or more sensors and/or one or more devices corresponding to a data stream applied to a corresponding machine learning model, and/or whether the anomaly has been previously detected by the same model for a selected amount of time..

Description

FIELD OF DISCLOSURE The present disclosure generally relates to systems and methods to predict or determine asset status, via a plurality of trained machine learning models, in an industrial plant at a selected time to prevent unplanned downtime. BACKGROUND An industrial plant may include a variety of equipment or assets, such as compressors, pumps, valves, and/or other equipment utilized for various plant operations. The equipment or assets may run or be operated every day and for all or a substantial portion of each day. While some equipment or assets can be serviced without impacting plant production, other equipment or assets cannot be brought down, taken offline, or shut down without causing a significant impact to plant production and/or operations. If such components are, for example, shut down, such an action may cause a reduction of product output and an increase in cost associated with the loss of such product output. Further, each industrial plant may include different and/or in some examples similar equipment and, further still, each industrial plant may produce a variety of different products. BRIEF SUMMARY In view of the foregoing, Applicant has recognized these problems and others in the art, and has recognized a need for enhanced systems and methods to predict or determine asset status, via a plurality of trained machine learning models, in an industrial plant at a selected time to prevent unplanned downtime. The present disclosure generally relates to systems and methods that address the relevant issues as described above, among other issues. In particular, such systems and methods may enable prediction of potential asset failure and/or prediction of unplanned or unscheduled asset maintenance to prevent unplanned downtime. Such systems and methods may ensure that these potential failures are assessed and/or resolved before the associated asset is taken offline or shut down, as well as predicting when maintenance should occur, regardless of when maintenance is scheduled, thus ensuring that downtime is at least partially or even completely or substantially completely reduced or prevented. Further, the systems and methods may enable monitoring of all of the assets positioned within an industrial plant and/or within a plurality of industrial plants. The use of a plurality of trained machine learning models by such systems and methods ensure that such potential failures and unplanned or unscheduled maintenance are accurately determined or predicted, rather than being overlooked and/or incorrectly predicted based on the output of a single machine learning model. Further, the use of a voting mechanism and/or weights for each of the plurality of machine learning models by the systems and methods ensure a more accurate determination or prediction of such potential failures and/or unplanned or unscheduled maintenance of the assets. Further, the systems and methods described herein may generate a notification to indicate the potential failures and/or unplanned or unscheduled maintenance of assets and, in other embodiments, the systems and methods may automatically perform or initiate resolution of the potential failures and unplanned or unscheduled maintenance of the assets. The systems and methods may additionally enable optimization of the assets or, in other words, may enable maintenance or indicate potential refurbishment, maintenance, upgrade, and/or resolutions regarding the assets. Such systems and methods may capture, acquire, or obtain data from previous or historical industrial plant operations (for example, from a database or other type of storage), and pre-process the captured data to form a pre-processed data set. Prior to utilization with live data or data captured in real-time, the systems and methods may utilize the data to train a plurality of machine learning models. The data may include a plurality of data sets associated with one or more assets positioned at a one or more plants. Each data set may include a plurality of data points corresponding to a measurement and/or other value associated with a sensor or device of one of the plurality of assets corresponding to one of a plurality of selected times. Once the data is captured, the systems and methods may include determining if any of the plurality of data sets included in the data are missing any data points (for example, missing data points based on data inclusion or lack thereof at any of the selected times that data is captured) or includes any frozen data points (for example, a data point that remains at the same value for over a selected threshold amount of times that data is captured). If the amount of missing or frozen data points exceed a selected threshold amount, then an error notification may be generated and/or the data points may be removed from a corresponding data set to ensure a full or substantially full data set is utilized for training. After removal of data points from the data sets, pre-processing may further