Back to results
US 11,378,946 B2Granted
Predictive maintenance method for component of production tool and computer program product thererof
AssigneeNational Cheng Kung University
InventorsChin-Yi Lin, Yu-Ming Hsieh, Fan-Tien Cheng, Hsien-Cheng Huang
Life of this patent
Estimating the term…
Abstract
Embodiments of the present invention provide a predictive maintenance method for a component of a production tool, in which a time series prediction (TSP) algorithm and an information criterion algorithm are adapted to build a TSP model, thereby forecasting the complicated future trend of accidental shutdown of the component of the production tool. In addition, an alarm scheme is provided for performing maintenance immediately when the component is very likely to enter a dead state, and a death related indicator (DCI) is provided for quantitatively showing the possibility of the component entering the dead state.
Claims18
A predictive maintenance method, comprising:
- obtaining a plurality of sets of process data used or generated by a component of a production tool when a plurality of workpieces are processed in sequence, wherein each of the sets of process data comprises values of a plurality of parameters, the values of each of parameters in each of the sets of process data is a set of time series data values within a period of processing time when the component of the production tool is processing one of the workpieces, and the sets of process data are one-to-one corresponding to the sets of workpieces;
- according to if an abnormal event occurs when the component is processing each of the workpieces, obtaining a plurality of event indicative values that are one-to-one corresponding to the sets of process data;
- converting the set of time series data values of each of the parameters in each of the sets of process data to a value of a parameter indicator by using a plurality of algorithms respectively, wherein the parameter indicators are one-to-one corresponding to the parameters;
- conducting a correlation analysis between each of parameter indicators and the event indicative values, thereby obtaining a plurality of correlation coefficients one-to-one corresponding to the parameter indicators;
- selecting one of the parameter indicators that is corresponding to a maximum one of the correlation coefficients as an aging feature (yT);
- performing a first determination operation to determine if the component is in a sick state according to the value of the aging feature (yT) corresponding to each of the workpieces, wherein one of the workpieces is set as a sample selection point once the component is in the sick state when processing the one of the workpieces;
- using N values of the aging feature as a set of model-building sample data, wherein the N values of the aging feature (yT) are corresponding to N sets of the sets of process data belonging to successive N of the workpieces that are processed directly before the sample selection point, wherein N is a positive integer;
- performing a model-building operation to build an aging-feature prediction model by using the set of model-building sample data in accordance a time series prediction algorithm, thereby obtaining a plurality of predicted values of the aging feature (yT) that are arranged in a processing order;
- obtaining a plurality of time points at which the component processes the workpieces respectively in accordance with a plurality of process times used by the component for processing the wokpieces respectively;
- obtaining a dead spec that is a value of the aging feature (yT) used or generated by the component when the component cannot work;
- sequentially determining if the predicted values are substantially equal to the dead spec until an earliest one of the predicted values is found, wherein one of the time points corresponding to the earliest one is a death time point at which the component cannot work;
- and
- computing differences between the death time point and the respective time points at which the workpieces are processed, thereby obtaining a plurality of predicted remaining useful life values (RULt), wherein t stands for the tth workpieces and t is an integer,
- wherein the model-building operation comprises:
- building the aging-feature prediction model by using an autoregressive integrated moving average (ARIMA) model as the time series prediction algorithm, wherein the ARIMA model comprising a moving average (MA) model and an autoregressive (AR) model;
- selecting the biggest lag of the MA model by using an autocorrelation function (ACF), and selecting the biggest lag of the AR model by using a Partial autocorrelation function (PACF);
- performing a white noise test on the values of the aging feature in the set of model-building sample data, wherein when one of the values of the aging feature in the set of model-building sample data is a white noise, a value of the aging feature corresponding to another set of process data belonging to a (N+1)th workpiece processed before the selection sample point is added to the set of model-building data;
- creating a plurality of ARIMA model combinations by using the biggest lag of the MA model and the biggest lag of the AR model;
- calculating an information quantity of each of the ARIMA model combinations by using an information criterion algorithm;
- and
- selecting one of the ARIMA model combinations that has the biggest information quantity as an optimal model.
18 more elements
A non-transitory tangible computer readable recording medium storing instructions which when executed by a processor configured to perform a predictive maintenance method, the predictive maintenance method comprising:
- obtaining a plurality of sets of process data used or generated by a component of a production tool when a plurality of workpieces are processed in sequence, wherein each of the sets of process data comprises values of a plurality of parameters, the values of each of parameters in each of the sets of process data is a set of time series data values within a period of processing time when the component of the production tool is processing one of the workpieces, and the sets of process data are one-to-one corresponding to the sets of workpieces;
- according to if an abnormal event occurs when the component is processing each of the workpieces, obtaining a plurality of event indicative values that are one-to-one corresponding to the sets of process data;
- converting the set of time series data values of each of the parameters in each of the sets of process data to a value of a parameter indicator by using a plurality of algorithms respectively, wherein the parameter indicators are one-to-one corresponding to the parameters;
- conducting a correlation analysis between each of parameter indicators and the event indicative values, thereby obtaining a plurality of correlation coefficients one-to-one corresponding to the parameter indicators;
- selecting one of the parameter indicators that is corresponding to a maximum one of the correlation coefficients as an aging feature (yT);
- performing a first determination operation to determine if the component is in a sick state according to the value of the aging feature (yT) corresponding to each of the workpieces, wherein one of the workpieces is set as a sample selection point once the component is in the sick state when processing the one of the workpieces;
- using N values of the aging feature as a set of model-building sample data, wherein the N values of the aging feature (yT) are corresponding to N sets of the sets of process data belonging to successive N of the workpieces that are processed directly before the sample selection point, wherein N is a positive integer;
- performing a model-building operation to build an aging-feature prediction model by using the set of model-building sample data in accordance a time series prediction algorithm, thereby obtaining a plurality of predicted values of the aging feature (yT) that are arranged in a processing order;
- obtaining a plurality of time points at which the component processes the workpieces respectively in accordance with a plurality of process times used by the component for processing the wokpieces respectively;
- obtaining a dead spec that is a value of the aging feature (yT) used or generated by the component when the component cannot work;
- sequentially determining if the predicted values are substantially equal to the dead spec until an earliest one of the predicted values is found, wherein one of the time points corresponding to the earliest one is a death time point at which the component cannot work;
- and
- computing differences between the death time point and the respective time points at which the workpieces are processed, thereby obtaining a plurality of predicted remaining useful life values (RULt), wherein t stands for the tth workpieces and t is an integer,
- wherein the model-building operation comprises:
- building the aging-feature prediction model by using an autoregressive integrated moving average (ARIMA) model as the time series prediction algorithm, wherein the ARIMA model comprising a moving average (MA) model and an autoregressive (AR) model;
- selecting the biggest lag of the MA model by using an autocorrelation function (ACF), and selecting the biggest lag of the AR model by using a partial autocorrelation function (PACF);
- performing a white noise test on the values of the aging feature in the set of model-building sample data, wherein when one of the values of the aging feature in the set of model-building sample data is a white noise, a value of the aging feature corresponding to another set of process data belonging to a (N+1)th workpiece processed before the selection sample point is added to the set of model-building data;
- creating a plurality of ARIMA model combinations by using the biggest lag of the MA model and the biggest lag of the AR model;
- calculating an information quantity of each of the ARIMA model combinations by using an information criterion algorithm;
- and
- selecting one of the ARIMA model combinations that has the biggest information quantity as an optimal model.
18 more elements
Dependencies are inferred from the claim text.